diff options
Diffstat (limited to 'tests/validation')
293 files changed, 27711 insertions, 5676 deletions
diff --git a/tests/validation/CL/ActivationLayer.cpp b/tests/validation/CL/ActivationLayer.cpp index fa95594157..133b39d154 100644 --- a/tests/validation/CL/ActivationLayer.cpp +++ b/tests/validation/CL/ActivationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -70,6 +70,7 @@ AbsoluteTolerance<float> tolerance(ActivationLayerInfo::ActivationFunction activ case ActivationLayerInfo::ActivationFunction::SOFT_RELU: case ActivationLayerInfo::ActivationFunction::ELU: case ActivationLayerInfo::ActivationFunction::SQRT: + case ActivationLayerInfo::ActivationFunction::GELU: return AbsoluteTolerance<float>(data_type == DataType::F16 ? 0.01f : 0.00001f); case ActivationLayerInfo::ActivationFunction::TANH: return AbsoluteTolerance<float>(data_type == DataType::F16 ? 0.001f : 0.00001f); diff --git a/tests/validation/CL/ArgMinMax.cpp b/tests/validation/CL/ArgMinMax.cpp index 1d849ed0c7..8566972f81 100644 --- a/tests/validation/CL/ArgMinMax.cpp +++ b/tests/validation/CL/ArgMinMax.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,15 +22,11 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" #include "arm_compute/runtime/CL/functions/CLArgMinMaxLayer.h" -#include "arm_compute/runtime/CL/functions/CLReductionOperation.h" #include "tests/CL/CLAccessor.h" #include "tests/datasets/ShapeDatasets.h" -#include "tests/datasets/SplitDataset.h" -#include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/validation/Validation.h" #include "tests/validation/fixtures/ArgMinMaxFixture.h" @@ -46,6 +42,8 @@ namespace const auto ArgMinMaxSmallDataset = framework::dataset::make("Shape", { TensorShape{ 1U, 7U, 1U, 3U }, + TensorShape{ 3U, 1U, 3U, 2U }, + TensorShape{ 2U, 1U, 3U, 2U }, TensorShape{ 149U, 5U, 1U, 2U }, TensorShape{ 166U, 5U, 1U, 2U }, TensorShape{ 322U, 5U, 1U, 2U }, @@ -53,6 +51,22 @@ const auto ArgMinMaxSmallDataset = framework::dataset::make("Shape", TensorShape{ 2560, 2U, 2U, 2U }, }); +const auto ArgMinMaxSmallDatasetAxis0 = framework::dataset::make("Shape", +{ + TensorShape{ 1U, 5U }, + TensorShape{ 2U, 3U }, + TensorShape{ 1U }, + TensorShape{ 3U }, + TensorShape{ 2U }, + TensorShape{ 5U }, + TensorShape{ 17U }, + TensorShape{ 15U, 2U }, +}); + +const auto OpsDataset = framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }); +const auto AxisDataset = framework::dataset::make("Axis", { 0, 1, 2, 3 }); +const auto QInfoDataset = framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }); + const auto ArgMinMaxLargeDataset = framework::dataset::make("Shape", { TensorShape{ 517U, 123U, 13U, 2U } }); } // namespace @@ -85,47 +99,78 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( // clang-format on // *INDENT-ON* -template <typename T> -using CLArgMinMaxValidationFixture = ArgMinMaxValidationFixture<CLTensor, CLAccessor, CLArgMinMaxLayer, T>; +template <typename T1, typename T2> +using CLArgMinMaxValidationFixture = ArgMinMaxValidationFixture<CLTensor, CLAccessor, CLArgMinMaxLayer, T1, T2>; + +using CLArgMinMaxValidationFixture_S32_S32 = CLArgMinMaxValidationFixture<int32_t, int32_t>; +using CLArgMinMaxValidationFixture_F16_S32 = CLArgMinMaxValidationFixture<half, int32_t>; +using CLArgMinMaxValidationFixture_F32_S32 = CLArgMinMaxValidationFixture<float, int32_t>; +using CLArgMinMaxValidationFixture_F32_S64 = CLArgMinMaxValidationFixture<float, int64_t>; TEST_SUITE(S32) +FIXTURE_DATA_TEST_CASE(RunSmallAxis0, + CLArgMinMaxValidationFixture_S32_S32, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(ArgMinMaxSmallDatasetAxis0, + framework::dataset::make("DataTypeIn", DataType::S32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + framework::dataset::make("Axis", { 0 })), + OpsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunSmall, - CLArgMinMaxValidationFixture<int32_t>, + CLArgMinMaxValidationFixture_S32_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(ArgMinMaxSmallDataset, framework::dataset::make("DataType", DataType::S32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxSmallDataset, + framework::dataset::make("DataTypeIn", DataType::S32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(CLAccessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, - CLArgMinMaxValidationFixture<int32_t>, + CLArgMinMaxValidationFixture_S32_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(ArgMinMaxLargeDataset, framework::dataset::make("DataType", DataType::S32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxLargeDataset, + framework::dataset::make("DataTypeIn", DataType::S32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(CLAccessor(_target), _reference); } + TEST_SUITE_END() // S32 TEST_SUITE(Float) TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, - CLArgMinMaxValidationFixture<half>, + CLArgMinMaxValidationFixture_F16_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(ArgMinMaxSmallDataset, framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxSmallDataset, + framework::dataset::make("DataTypeIn", DataType::F16)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(CLAccessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, - CLArgMinMaxValidationFixture<half>, + CLArgMinMaxValidationFixture_F16_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(ArgMinMaxLargeDataset, framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxLargeDataset, + framework::dataset::make("DataTypeIn", DataType::F16)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(CLAccessor(_target), _reference); @@ -134,49 +179,77 @@ TEST_SUITE_END() // FP16 TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, - CLArgMinMaxValidationFixture<float>, + CLArgMinMaxValidationFixture_F32_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(ArgMinMaxSmallDataset, framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxSmallDataset, + framework::dataset::make("DataTypeIn", DataType::F32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunSmall_F32_S64, + CLArgMinMaxValidationFixture_F32_S64, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(ArgMinMaxSmallDataset, + framework::dataset::make("DataTypeIn", DataType::F32)), + framework::dataset::make("DataTypeOut", DataType::S64)), + AxisDataset), + OpsDataset)) { // Validate output validate(CLAccessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, - CLArgMinMaxValidationFixture<float>, + CLArgMinMaxValidationFixture_F32_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(ArgMinMaxLargeDataset, framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxLargeDataset, + framework::dataset::make("DataTypeIn", DataType::F32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(CLAccessor(_target), _reference); } + TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float -template <typename T> -using CLArgMinMaxQuantizedValidationFixture = ArgMinMaxValidationQuantizedFixture<CLTensor, CLAccessor, CLArgMinMaxLayer, T>; +template <typename T1, typename T2> +using CLArgMinMaxQuantizedValidationFixture = ArgMinMaxValidationQuantizedFixture<CLTensor, CLAccessor, CLArgMinMaxLayer, T1, T2>; + +using CLArgMinMaxQuantizedValidationFixture_U8_S32 = CLArgMinMaxQuantizedValidationFixture<uint8_t, int32_t>; +using CLArgMinMaxQuantizedValidationFixture_S8_S32 = CLArgMinMaxQuantizedValidationFixture<int8_t, int32_t>; TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, - CLArgMinMaxQuantizedValidationFixture<uint8_t>, + CLArgMinMaxQuantizedValidationFixture_U8_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(ArgMinMaxSmallDataset, framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxSmallDataset, + framework::dataset::make("DataTypeIn", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(CLAccessor(_target), _reference); } - FIXTURE_DATA_TEST_CASE(RunLarge, - CLArgMinMaxQuantizedValidationFixture<uint8_t>, + CLArgMinMaxQuantizedValidationFixture_U8_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(ArgMinMaxLargeDataset, framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxLargeDataset, + framework::dataset::make("DataTypeIn", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(CLAccessor(_target), _reference); @@ -185,28 +258,32 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, - CLArgMinMaxQuantizedValidationFixture<int8_t>, + CLArgMinMaxQuantizedValidationFixture_S8_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(ArgMinMaxSmallDataset, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxSmallDataset, + framework::dataset::make("DataTypeIn", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(CLAccessor(_target), _reference); } - FIXTURE_DATA_TEST_CASE(RunLarge, - CLArgMinMaxQuantizedValidationFixture<int8_t>, + CLArgMinMaxQuantizedValidationFixture_S8_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(ArgMinMaxLargeDataset, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxLargeDataset, + framework::dataset::make("DataTypeIn", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE_END() // QASYMM8_SIGNED - TEST_SUITE_END() // Quantized TEST_SUITE_END() // ArgMinMax TEST_SUITE_END() // CL diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp index c74f6a3b23..1ed3a105dc 100644 --- a/tests/validation/CL/ArithmeticAddition.cpp +++ b/tests/validation/CL/ArithmeticAddition.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -41,26 +41,12 @@ namespace test { namespace validation { +/** Synced with tests/validation/dynamic_fusion/gpu/cl/Add.cpp from the dynamic fusion interface. + * Please check there for any differences in the coverage + */ namespace { /** Input data sets **/ -const auto ArithmeticAdditionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)), framework::dataset::make("DataType", - DataType::U8)); -const auto ArithmeticAdditionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataType", - DataType::QASYMM8)); -const auto ArithmeticAdditionQASYMM8SignedDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)); -const auto ArithmeticAdditionQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)), - framework::dataset::make("DataType", - DataType::QSYMM16)); -const auto ArithmeticAdditionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), - framework::dataset::make("DataType", DataType::S16)); -const auto ArithmeticAdditionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("DataType", DataType::F16)); -const auto ArithmeticAdditionFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataType", DataType::F32)); const auto EmptyActivationFunctionsDataset = framework::dataset::make("ActivationInfo", { ActivationLayerInfo() }); const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", @@ -68,6 +54,8 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(CL) @@ -76,22 +64,19 @@ TEST_SUITE(ArithmeticAddition) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes }), - framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false})), + framework::dataset::make("Expected", { true, false, false})), input1_info, input2_info, output_info, expected) { ARM_COMPUTE_EXPECT(bool(CLArithmeticAddition::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), ConvertPolicy::WRAP)) == expected, framework::LogLevel::ERRORS); @@ -129,8 +114,10 @@ using CLArithmeticAdditionFixture = ArithmeticAdditionValidationFixture<CLTensor TEST_SUITE(Integer) TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticAdditionU8Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -138,15 +125,19 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixture<uint8_t>, framework TEST_SUITE_END() // U8 TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticAdditionS16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticAdditionS16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -159,48 +150,65 @@ using CLArithmeticAdditionQuantizedFixture = ArithmeticAdditionValidationQuantiz TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticAdditionQASYMM8Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } template <typename T> using CLArithmeticAdditionBroadcastQuantizedFixture = ArithmeticAdditionValidationQuantizedBroadcastFixture<CLTensor, CLAccessor, CLArithmeticAddition, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapesBroadcast(), - ArithmeticAdditionQASYMM8Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), - framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), - framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(datasets::SmallShapesBroadcast(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, CLArithmeticAdditionBroadcastQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(datasets::TinyShapesBroadcastInplace(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 255.f, 10) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(1.f / 255.f, 10) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 10) })), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticAdditionQASYMM8SignedDataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 10) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE(QSYMM16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticAdditionQSYMM16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QSYMM16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -213,16 +221,21 @@ using CLArithmeticAdditionFloatFixture = ArithmeticAdditionValidationFloatFixtur TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionFP16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::F16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticAdditionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ArithmeticAdditionFP16Dataset), +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticAdditionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::TinyShapes(), + framework::dataset::make("DataType", + DataType::F16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -230,24 +243,32 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticAdditionFloatFixture<half> TEST_SUITE_END() // FP16 TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", + DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticAdditionFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ArithmeticAdditionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticAdditionFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::TinyShapes(), + framework::dataset::make("DataType", + DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticAdditionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -256,27 +277,30 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFloatFixture<float>, framew template <typename T> using CLArithmeticAdditionBroadcastFloatFixture = ArithmeticAdditionBroadcastValidationFloatFixture<CLTensor, CLAccessor, CLArithmeticAddition, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapesBroadcast(), - ArithmeticAdditionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapesBroadcast(), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLArithmeticAdditionBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), - ArithmeticAdditionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLArithmeticAdditionBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::TinyShapesBroadcast(), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticAdditionBroadcastFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapesBroadcast(), - ArithmeticAdditionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticAdditionBroadcastFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapesBroadcast(), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); diff --git a/tests/validation/CL/ArithmeticDivision.cpp b/tests/validation/CL/ArithmeticDivision.cpp index 9dcdfb83e1..94bacba7e5 100644 --- a/tests/validation/CL/ArithmeticDivision.cpp +++ b/tests/validation/CL/ArithmeticDivision.cpp @@ -59,6 +59,8 @@ const auto ActivationFunctionsDataset = framework::dataset::make("Activatio ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(CL) @@ -95,14 +97,16 @@ using CLArithmeticDivisionIntegerFixture = ArithmeticDivisionValidationIntegerFi TEST_SUITE(Integer) TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmallInteger, CLArithmeticDivisionIntegerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticDivisionS32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmallInteger, CLArithmeticDivisionIntegerFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticDivisionS32Dataset), + EmptyActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunIntegerWithActivation, CLArithmeticDivisionIntegerFixture, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ArithmeticDivisionS32Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunIntegerWithActivation, CLArithmeticDivisionIntegerFixture, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ArithmeticDivisionS32Dataset), + ActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -116,14 +120,16 @@ using CLArithmeticDivisionFloatFixture = ArithmeticDivisionValidationFloatFixtur TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticDivisionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ArithmeticDivisionFP16Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticDivisionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ArithmeticDivisionFP16Dataset), + EmptyActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticDivisionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ArithmeticDivisionFP16Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticDivisionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ArithmeticDivisionFP16Dataset), + ActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); @@ -131,21 +137,24 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticDivisionFloatFixture<half> TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticDivisionFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticDivisionFP32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticDivisionFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticDivisionFP32Dataset), + EmptyActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticDivisionFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ArithmeticDivisionFP32Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticDivisionFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ArithmeticDivisionFP32Dataset), + ActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticDivisionFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticDivisionFP32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticDivisionFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticDivisionFP32Dataset), + EmptyActivationFunctionsDataset), + InPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); @@ -154,24 +163,27 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticDivisionFloatFixture<float>, framew template <typename T> using CLArithmeticDivisionBroadcastFloatFixture = ArithmeticDivisionBroadcastValidationFloatFixture<CLTensor, CLAccessor, CLArithmeticDivision, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticDivisionBroadcastFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticDivisionBroadcastFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapesBroadcast(), ArithmeticDivisionFP32Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLArithmeticDivisionBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLArithmeticDivisionBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), ArithmeticDivisionFP32Dataset), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticDivisionBroadcastFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticDivisionBroadcastFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapesBroadcast(), ArithmeticDivisionFP32Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); diff --git a/tests/validation/CL/ArithmeticSubtraction.cpp b/tests/validation/CL/ArithmeticSubtraction.cpp index 2709fcaedb..5825ce2e5d 100644 --- a/tests/validation/CL/ArithmeticSubtraction.cpp +++ b/tests/validation/CL/ArithmeticSubtraction.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -41,27 +41,12 @@ namespace test { namespace validation { +/** Synced with tests/validation/dynamic_fusion/gpu/cl/Sub.cpp from the dynamic fusion interface. + * Please check there for any differences in the coverage + */ namespace { /** Input data sets **/ -const auto ArithmeticSubtractionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)), - framework::dataset::make("DataType", - DataType::U8)); -const auto ArithmeticSubtractionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataType", - DataType::QASYMM8)); -const auto ArithmeticSubtractionQASYMM8SignedDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)); -const auto ArithmeticSubtractionQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)), - framework::dataset::make("DataType", - DataType::QSYMM16)); -const auto ArithmeticSubtractionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), - framework::dataset::make("DataType", DataType::S16)); -const auto ArithmeticSubtractionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("DataType", DataType::F16)); -const auto ArithmeticSubtractionFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataType", DataType::F32)); const auto EmptyActivationFunctionsDataset = framework::dataset::make("ActivationInfo", { ActivationLayerInfo() }); const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", @@ -79,22 +64,19 @@ TEST_SUITE(ArithmeticSubtraction) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes }), - framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false})), + framework::dataset::make("Expected", { true, false, false})), input1_info, input2_info, output_info, expected) { ARM_COMPUTE_EXPECT(bool(CLArithmeticSubtraction::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), ConvertPolicy::WRAP)) == expected, framework::LogLevel::ERRORS); @@ -159,7 +141,8 @@ using CLArithmeticSubtractionFixture = ArithmeticSubtractionValidationFixture<CL TEST_SUITE(Integer) TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionU8Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -169,7 +152,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<uint8_t>, framew TEST_SUITE_END() // U8 TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -177,7 +161,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixture<int16_t>, framew validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS16Dataset), +FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -193,11 +178,22 @@ using CLArithmeticSubtractionQuantizedFixture = ArithmeticSubtractionValidationQ TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticSubtractionQASYMM8Dataset), + framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunTinyInPlace, CLArithmeticSubtractionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::TinyShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 255.f, 20) })), InPlaceDataSet)) { // Validate output @@ -206,12 +202,12 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionQuantizedFixture<uint8_t TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticSubtractionQASYMM8SignedDataset), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 10) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -219,7 +215,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionQuantizedFixture<int8_t> TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE(QSYMM16) FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticSubtractionQSYMM16Dataset), + framework::dataset::make("DataType", DataType::QSYMM16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), @@ -237,7 +233,8 @@ using CLArithmeticSubtractionFloatFixture = ArithmeticSubtractionValidationFloat TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionFP16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::F16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), EmptyActivationFunctionsDataset), OutOfPlaceDataSet)) @@ -246,7 +243,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFloatFixture<half>, fram validate(CLAccessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticSubtractionFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::TinyShapes(), - ArithmeticSubtractionFP16Dataset), + framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), ActivationFunctionsDataset), InPlaceDataSet)) @@ -258,7 +255,7 @@ TEST_SUITE_END() // FP16 TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), EmptyActivationFunctionsDataset), OutOfPlaceDataSet)) @@ -267,7 +264,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFloatFixture<float>, fra validate(CLAccessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticSubtractionFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::TinyShapes(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), ActivationFunctionsDataset), InPlaceDataSet)) @@ -277,7 +274,7 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLArithmeticSubtractionFloatFixture<fl } FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapes(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), EmptyActivationFunctionsDataset), OutOfPlaceDataSet)) @@ -290,7 +287,7 @@ template <typename T> using CLArithmeticSubtractionBroadcastFloatFixture = ArithmeticSubtractionBroadcastValidationFloatFixture<CLTensor, CLAccessor, CLArithmeticSubtraction, T>; FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticSubtractionBroadcastFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapesBroadcast(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), EmptyActivationFunctionsDataset), OutOfPlaceDataSet)) @@ -298,8 +295,18 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticSubtractionBroadcastFloatF // Validate output validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInplace, CLArithmeticSubtractionBroadcastFloatFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(datasets::TinyShapesBroadcastInplace(), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + EmptyActivationFunctionsDataset), + InPlaceDataSet)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLArithmeticSubtractionBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::TinyShapesBroadcast(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), ActivationFunctionsDataset), OutOfPlaceDataSet)) @@ -309,7 +316,7 @@ FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLArithmeticSubtractionBroadc } FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticSubtractionBroadcastFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapesBroadcast(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), EmptyActivationFunctionsDataset), OutOfPlaceDataSet)) diff --git a/tests/validation/CL/BatchNormalizationLayer.cpp b/tests/validation/CL/BatchNormalizationLayer.cpp index 8b3bdbc3ea..3b87b9d1b5 100644 --- a/tests/validation/CL/BatchNormalizationLayer.cpp +++ b/tests/validation/CL/BatchNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,7 @@ namespace { RelativeTolerance<float> rel_tolerance_f32(0.05f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ -constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr AbsoluteTolerance<float> tolerance_f16(0.02f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ const auto act_infos = framework::dataset::make("ActivationInfo", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), diff --git a/tests/validation/CL/BatchToSpaceLayer.cpp b/tests/validation/CL/BatchToSpaceLayer.cpp index e90ac921c5..ca12b76e8a 100644 --- a/tests/validation/CL/BatchToSpaceLayer.cpp +++ b/tests/validation/CL/BatchToSpaceLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,56 +50,38 @@ using CLBatchToSpaceLayerFixture = BatchToSpaceLayerValidationFixture<CLTensor, // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blockx > blocky - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blocky > blockx - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), // Mismatching data types - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), // Wrong data type block shape - TensorInfo(TensorShape(32U, 13U, 2U, 2U, 4U), 1, DataType::F32), // Wrong tensor shape - }), - framework::dataset::make("BlockShapeInfo",{ TensorInfo(TensorShape(2U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(2U, 4U), 1, DataType::S32), - TensorInfo(TensorShape(4U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(2U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(2U, 2U), 1, DataType::F16), - TensorInfo(TensorShape(2U, 2U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(64U, 16U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 32U, 2U, 1U), 1, DataType::F32), - - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F16), - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - })), - framework::dataset::make("Expected", { true, true,true, false, false, false})), - input_info, block_shape_info, output_info, expected) -{ - bool has_error = bool(CLBatchToSpaceLayer::validate(&input_info.clone()->set_is_resizable(false), &block_shape_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))); - ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS); -} -DATA_TEST_CASE(ValidateStatic, framework::DatasetMode::ALL, zip(zip(zip(zip( +DATA_TEST_CASE(ValidateStatic, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blockx > blocky - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blocky > blockx - TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Mismatching data types - TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Negative block shapes - TensorInfo(TensorShape(32U, 16U, 2U, 4U, 4U), 1, DataType::F32), // Wrong tensor shape + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Supported: blockx != blocky && blockx > blocky + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Supported: blockx != blocky && blocky > blockx + TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Invalid: Mismatching data types + TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Invalid: Negative block shapes + TensorInfo(TensorShape(32U, 16U, 2U, 4U, 4U), 1, DataType::F32),// Unsupported tensor rank + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Invalid output tensor shape (invalid batch dimension) + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Invalid output tensor shape (invalid spatial dimension) + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Supported: correct tensor shape with cropping + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Invalid tensor shape with cropping }), - framework::dataset::make("BlockShapeX", { 2, 4, 2, 2, 2, 2 })), - framework::dataset::make("BlockShapeY", { 2, 2, 4, 2, -2, 2 })), + framework::dataset::make("BlockShapeX", { 2, 4, 2, 2, 2, 2, 2, 2, 2, 2 })), + framework::dataset::make("BlockShapeY", { 2, 2, 4, 2, -2, 2, 2, 2, 2, 2 })), + framework::dataset::make("CropInfo", { + CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{3, 2, 1, 3}, CropInfo{3, 2, 1, 3} + })), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(64U, 16U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 32U, 2U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(64U, 16U, 2U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 32U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F16), TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F32), TensorInfo(TensorShape(32U, 8U, 2U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(33U, 32U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(27, 12U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 16U, 2U, 4U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true,true, false, false, false})), - input_info, block_shape_x, block_shape_y, output_info, expected) + framework::dataset::make("Expected", { true, true, true, false, false, false, false, false, true, false})), + input_info, block_shape_x, block_shape_y, crop_info, output_info, expected) { - bool has_error = bool(CLBatchToSpaceLayer::validate(&input_info.clone()->set_is_resizable(false), block_shape_x, block_shape_y, &output_info.clone()->set_is_resizable(false))); + bool has_error = bool(CLBatchToSpaceLayer::validate(&input_info.clone()->set_is_resizable(false), block_shape_x, block_shape_y, &output_info.clone()->set_is_resizable(false), crop_info)); ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS); } // clang-format on @@ -114,6 +96,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchToSpaceLayerFixture<float>, framework::D // Validate output validate(CLAccessor(_target), _reference); } + +FIXTURE_DATA_TEST_CASE(RunSmallWithCropping, CLBatchToSpaceLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::SmallBatchToSpaceLayerWithCroppingDataset(), framework::dataset::make("DataType", + DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, CLBatchToSpaceLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeBatchToSpaceLayerDataset(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) @@ -131,6 +123,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchToSpaceLayerFixture<half>, framework::Da // Validate output validate(CLAccessor(_target), _reference); } + +FIXTURE_DATA_TEST_CASE(RunSmallWithCropping, CLBatchToSpaceLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::SmallBatchToSpaceLayerWithCroppingDataset(), framework::dataset::make("DataType", + DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, CLBatchToSpaceLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeBatchToSpaceLayerDataset(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) diff --git a/tests/validation/CL/Cast.cpp b/tests/validation/CL/Cast.cpp index 2ca8b58040..2f943e84d8 100644 --- a/tests/validation/CL/Cast.cpp +++ b/tests/validation/CL/Cast.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,6 +48,10 @@ constexpr AbsoluteTolerance<float> one_tolerance(1); constexpr AbsoluteTolerance<float> zero_tolerance(0); /** Input data sets **/ +// QASYMM8 +const auto CastQASYMM8toF32Dataset = combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::F32)); +const auto CastQSYMM8toF32Dataset = combine(framework::dataset::make("DataType", DataType::QSYMM8), framework::dataset::make("DataType", DataType::F32)); + // U8 const auto CastU8toS8Dataset = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::S8)); const auto CastU8toU16Dataset = combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U16)); @@ -119,6 +123,26 @@ const auto CastF32toS16Dataset = combine(framework::dataset::make("DataType", Da const auto CastF32toU32Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::U32)); const auto CastF32toS32Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::S32)); const auto CastF32toF16Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F16)); + +// U64 +const auto CastU64toU8Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::U8)); +const auto CastU64toS8Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::S8)); +const auto CastU64toU16Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::U16)); +const auto CastU64toS16Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::S16)); +const auto CastU64toU32Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::U32)); +const auto CastU64toS32Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::S32)); +const auto CastU64toF16Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::F16)); +const auto CastU64toF32Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::F32)); + +// S64 +const auto CastS64toU8Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::U8)); +const auto CastS64toS8Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::S8)); +const auto CastS64toU16Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::U16)); +const auto CastS64toS16Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::S16)); +const auto CastS64toU32Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::U32)); +const auto CastS64toS32Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::S32)); +const auto CastS64toF16Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::F16)); +const auto CastS64toF32Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::F32)); } // namespace TEST_SUITE(CL) @@ -149,6 +173,12 @@ using CLCastToF32Fixture = CastValidationFixture<CLTensor, CLAccessor, CLCast, T } \ TEST_SUITE_END() +// QASYMM8 +CAST_SUITE(QASYMM8_to_F32, DataType::QASYMM8, DataType::F32, CLCastToF32Fixture<uint8_t>, CastQASYMM8toF32Dataset, zero_tolerance) +// QSYMM8 +CAST_SUITE(QSYMM8_to_F32, DataType::QSYMM8, DataType::F32, CLCastToF32Fixture<int8_t>, CastQSYMM8toF32Dataset, zero_tolerance) + + // U8 CAST_SUITE(U8_to_S8, DataType::U8, DataType::S8, CLCastToS8Fixture<uint8_t>, CastU8toS8Dataset, zero_tolerance) CAST_SUITE(U8_to_U16, DataType::U8, DataType::U16, CLCastToU16Fixture<uint8_t>, CastU8toU16Dataset, zero_tolerance) @@ -221,6 +251,26 @@ CAST_SUITE(F32_to_U32, DataType::F32, DataType::U32, CLCastToU32Fixture<float>, CAST_SUITE(F32_to_S32, DataType::F32, DataType::S32, CLCastToS32Fixture<float>, CastF32toS32Dataset, one_tolerance) CAST_SUITE(F32_to_F16, DataType::F32, DataType::F16, CLCastToF16Fixture<float>, CastF32toF16Dataset, zero_tolerance) +// S64 +CAST_SUITE(S64_to_U8, DataType::S64, DataType::U8, CLCastToU8Fixture<int64_t>, CastS64toU8Dataset, one_tolerance) +CAST_SUITE(S64_to_S8, DataType::S64, DataType::S8, CLCastToS8Fixture<int64_t>, CastS64toS8Dataset, one_tolerance) +CAST_SUITE(S64_to_U16, DataType::S64, DataType::U16, CLCastToU16Fixture<int64_t>, CastS64toU16Dataset, one_tolerance) +CAST_SUITE(S64_to_S16, DataType::S64, DataType::S16, CLCastToS16Fixture<int64_t>, CastS64toS16Dataset, one_tolerance) +CAST_SUITE(S64_to_U32, DataType::S64, DataType::U32, CLCastToU32Fixture<int64_t>, CastS64toU32Dataset, one_tolerance) +CAST_SUITE(S64_to_S32, DataType::S64, DataType::S32, CLCastToS32Fixture<int64_t>, CastS64toS32Dataset, one_tolerance) +CAST_SUITE(S64_to_F16, DataType::S64, DataType::F16, CLCastToF16Fixture<int64_t>, CastS64toF16Dataset, zero_tolerance) +CAST_SUITE(S64_to_F32, DataType::S64, DataType::F32, CLCastToF32Fixture<int64_t>, CastS64toF32Dataset, zero_tolerance) + +// U64 +CAST_SUITE(U64_to_U8, DataType::U64, DataType::U8, CLCastToU8Fixture<uint64_t>, CastU64toU8Dataset, one_tolerance) +CAST_SUITE(U64_to_S8, DataType::U64, DataType::S8, CLCastToS8Fixture<uint64_t>, CastU64toS8Dataset, one_tolerance) +CAST_SUITE(U64_to_U16, DataType::U64, DataType::U16, CLCastToU16Fixture<uint64_t>, CastU64toU16Dataset, one_tolerance) +CAST_SUITE(U64_to_S16, DataType::U64, DataType::S16, CLCastToS16Fixture<uint64_t>, CastU64toS16Dataset, one_tolerance) +CAST_SUITE(U64_to_U32, DataType::U64, DataType::U32, CLCastToU32Fixture<uint64_t>, CastU64toU32Dataset, one_tolerance) +CAST_SUITE(U64_to_S32, DataType::U64, DataType::S32, CLCastToS32Fixture<uint64_t>, CastU64toS32Dataset, one_tolerance) +CAST_SUITE(U64_to_F16, DataType::U64, DataType::F16, CLCastToF16Fixture<uint64_t>, CastU64toF16Dataset, zero_tolerance) +CAST_SUITE(U64_to_F32, DataType::U64, DataType::F32, CLCastToF32Fixture<uint64_t>, CastU64toF32Dataset, zero_tolerance) + TEST_SUITE_END() // Cast TEST_SUITE_END() // CL } // namespace validation diff --git a/tests/validation/CL/Col2Im.cpp b/tests/validation/CL/Col2Im.cpp index b651bf8918..4b004e2472 100644 --- a/tests/validation/CL/Col2Im.cpp +++ b/tests/validation/CL/Col2Im.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,7 +22,7 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "src/core/CL/kernels/CLCol2ImKernel.h" +#include "src/gpu/cl/kernels/ClCol2ImKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/framework/Asserts.h" @@ -40,7 +40,7 @@ namespace validation TEST_SUITE(CL) TEST_SUITE(Col2Im) -using CLCol2Im = CLSynthetizeFunction<CLCol2ImKernel>; +using ClCol2Im = ClSynthetizeOperatorWithBorder<opencl::kernels::ClCol2ImKernel>; /** Negative tests * @@ -59,7 +59,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto input = TensorInfo(TensorShape(10U, 12U, 1U, 2U), 1, DataType::SIZET); const auto output = TensorInfo(TensorShape(3U, 4U, 10U, 1U, 2U), 1, DataType::F32); const auto conv_size = Size2D(3, 4); - const auto status = CLCol2ImKernel::validate(&input, &output, conv_size); + const auto status = opencl::kernels::ClCol2ImKernel::validate(&input, &output, conv_size); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -68,7 +68,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto input = TensorInfo(TensorShape(10U, 12U, 1U, 2U), 1, DataType::F32); const auto output = TensorInfo(TensorShape(3U, 4U, 10U, 1U, 2U), 1, DataType::F32, DataLayout::NHWC); const auto conv_size = Size2D(3, 4); - const auto status = CLCol2ImKernel::validate(&input, &output, conv_size); + const auto status = opencl::kernels::ClCol2ImKernel::validate(&input, &output, conv_size); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -77,13 +77,13 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto input = TensorInfo(TensorShape(10U, 12U, 1U, 2U), 1, DataType::F32); const auto output = TensorInfo(TensorShape(3U, 4U, 10U, 2U, 2U), 1, DataType::F32); const auto conv_size = Size2D(3, 4); - const auto status = CLCol2ImKernel::validate(&input, &output, conv_size); + const auto status = opencl::kernels::ClCol2ImKernel::validate(&input, &output, conv_size); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } } template <typename T> -using CLCol2ImFixture = Col2ImValidationFixture<CLTensor, CLAccessor, CLCol2Im, T, true>; +using ClCol2ImFixture = Col2ImOpValidationFixture<CLTensor, CLAccessor, ClCol2Im, T, true>; /** Test kernel for single-precision floating point * @@ -99,7 +99,7 @@ using CLCol2ImFixture = Col2ImValidationFixture<CLTensor, CLAccessor, CLCol2Im, * Kernel tested col2im */ FIXTURE_DATA_TEST_CASE(FP32, - CLCol2ImFixture<float>, + ClCol2ImFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine( framework::dataset::make("InputShape", { TensorShape(8U, 16U, 3U, 1U), TensorShape(17U, 16U, 3U, 1U), TensorShape(7U, 16U, 3U, 1U) }), @@ -125,7 +125,7 @@ FIXTURE_DATA_TEST_CASE(FP32, * Kernel tested col2im */ FIXTURE_DATA_TEST_CASE(F16, - CLCol2ImFixture<half>, + ClCol2ImFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(17U, 16U, 3U, 1U)), @@ -151,7 +151,7 @@ FIXTURE_DATA_TEST_CASE(F16, * Kernel tested col2im */ FIXTURE_DATA_TEST_CASE(QASYMM8, - CLCol2ImFixture<uint8_t>, + ClCol2ImFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(17U, 16U, 3U, 1U)), diff --git a/tests/validation/CL/Comparisons.cpp b/tests/validation/CL/Comparisons.cpp index d015528b0e..dd3dbd8d59 100644 --- a/tests/validation/CL/Comparisons.cpp +++ b/tests/validation/CL/Comparisons.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -59,7 +59,7 @@ TEST_SUITE(Comparison) DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid output type TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching input types - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Window shrink + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), }), @@ -75,7 +75,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), })), - framework::dataset::make("Expected", { false, false, false, false, true})), + framework::dataset::make("Expected", { false, false, true, false, true})), input1_info, input2_info, output_info, expected) { Status s = CLComparison::validate(&input1_info.clone()->set_is_resizable(false), diff --git a/tests/validation/CL/Convolution3D.cpp b/tests/validation/CL/Convolution3D.cpp new file mode 100644 index 0000000000..a2848560c3 --- /dev/null +++ b/tests/validation/CL/Convolution3D.cpp @@ -0,0 +1,300 @@ +/* + * Copyright (c) 2021, 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/functions/CLConv3D.h" +#include "arm_compute/runtime/FunctionDescriptors.h" +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/DirectConvolution3DFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +const RelativeTolerance<half> rel_tolerance_fp16(half(0.2)); /**< Relative tolerance for FP16 tests */ +constexpr float abs_tolerance_fp16(0.05f); /**< Absolute tolerance for FP16 tests */ +constexpr RelativeTolerance<float> rel_tolerance_fp32(0.05f); /**< Relative tolerance for FP32 tests */ +constexpr float abs_tolerance_fp32(0.0001f); /**< Absolute tolerance for FP32 tests*/ +constexpr AbsoluteTolerance<uint8_t> abs_tolerance_qasymm8(1); /**< Absolute tolerance for quantized tests */ +constexpr float tolerance_num = 0.07f; /**< Tolerance number */ +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DirectConvolution3D) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 5U, 3U), // Unsupported data layout + TensorShape(27U, 13U, 5U, 3U), // Unsupported activation enabled + TensorShape(27U, 13U, 5U, 3U), // Mismatching data type + TensorShape(27U, 13U, 5U, 3U), // Unsupported data type + TensorShape(27U, 13U, 5U, 3U), // Mismatching input feature maps + TensorShape(27U, 13U, 5U, 3U), // Mismatching output feature maps + TensorShape(27U, 13U, 5U, 3U), // Mismatching bias shape + TensorShape(27U, 13U, 5U, 3U), // Unsupported number of weights dimensions + TensorShape(27U, 13U, 5U, 3U), // Unsupported number of biases dimensions + TensorShape(27U, 13U, 5U, 3U), // Mismatching output shape + TensorShape(27U, 13U, 5U, 3U) + }), + framework::dataset::make("WeightsShape", { TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 32U, 3U, 3U, 3U), + TensorShape(8U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U, 2U), + TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U), + TensorShape(4U, 27U, 3U, 3U, 3U) + })), + framework::dataset::make("BiasesShape", { TensorShape(4U), + TensorShape(4U), + TensorShape(4U), + TensorShape(4U), + TensorShape(4U), + TensorShape(4U), + TensorShape(8U), + TensorShape(4U), + TensorShape(4U), + TensorShape(4U), + TensorShape(4U) + })), + framework::dataset::make("OutputShape", { TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U, 2U), + TensorShape(4U, 11U, 5U, 3U), + TensorShape(4U, 13U, 5U, 3U) + })), + framework::dataset::make("Conv3dInfo", { Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false), + Conv3dInfo(Size3D(1U, 1U, 1U), Padding3D(1U, 1U, 1U), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false) + })), + framework::dataset::make("SrcDataType", { DataType::F32, + DataType::F32, + DataType::F32, + DataType::U32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32 + })), + framework::dataset::make("WeightsDataType", { DataType::F32, + DataType::F32, + DataType::F16, + DataType::U32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32, + DataType::F32 + })), + framework::dataset::make("DataLayout", { DataLayout::NCDHW, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC, + DataLayout::NDHWC + })), + framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, false, true })), + input_shape, weights_shape, biases_shape, output_shape, conv3d_info, src_data_type, weights_data_type, data_layout, expected) +{ + TensorInfo input_info = TensorInfo(input_shape, 1, src_data_type); + TensorInfo weights_info = TensorInfo(weights_shape, 1, weights_data_type); + TensorInfo biases_info = TensorInfo(biases_shape, 1, src_data_type); + TensorInfo output_info = TensorInfo(output_shape, 1, src_data_type); + + input_info.set_data_layout(data_layout); + weights_info.set_data_layout(data_layout); + biases_info.set_data_layout(data_layout); + output_info.set_data_layout(data_layout); + + bool is_valid = bool(CLConv3D::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv3d_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} + +template <typename T> +using CLDirectConvolution3DFixture = DirectConvolution3DValidationFixture<CLTensor, CLAccessor, CLConv3D, T>; +template <typename T> +using CLDirectConvolution3DQuantizedFixture = DirectConvolution3DValidationQuantizedFixture<CLTensor, CLAccessor, CLConv3D, T>; + +TEST_SUITE(NDHWC) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolution3DFixture<half>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(7U, 5U, 3U, 13U, 3U), + TensorShape(15U, 7U, 11U, 7U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U) + }), + framework::dataset::make("StrideX", { 1, 3, 2, 1 })), + framework::dataset::make("StrideY", { 2, 1, 3, 1 })), + framework::dataset::make("StrideZ", { 3, 2, 1, 1 })), + framework::dataset::make("PadX", { 0, 2, 1, 0 })), + framework::dataset::make("PadY", { 1, 0, 2, 0 })), + framework::dataset::make("PadZ", { 2, 1, 0, 0 })), + framework::dataset::make("KernelWidth", { 3, 7, 5, 1 })), + framework::dataset::make("KernelHeight", { 5, 3, 7, 1 })), + framework::dataset::make("KernelDepth", { 7, 5, 3, 1 })), + framework::dataset::make("NumKernels", { 5, 3, 1, 11 })), + framework::dataset::make("HasBias", { true, true, true, false })), + framework::dataset::make("Activation", ActivationLayerInfo())), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", DataLayout::NDHWC))) +{ + validate(CLAccessor(_target), _reference, rel_tolerance_fp16, tolerance_num, abs_tolerance_fp16); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolution3DFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(7U, 5U, 3U, 13U, 3U), + TensorShape(15U, 7U, 11U, 7U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U) + }), + framework::dataset::make("StrideX", { 1, 3, 2, 1 })), + framework::dataset::make("StrideY", { 2, 1, 3, 1 })), + framework::dataset::make("StrideZ", { 3, 2, 1, 1 })), + framework::dataset::make("PadX", { 0, 2, 1, 0 })), + framework::dataset::make("PadY", { 1, 0, 2, 0 })), + framework::dataset::make("PadZ", { 2, 1, 0, 0 })), + framework::dataset::make("KernelWidth", { 3, 7, 5, 1 })), + framework::dataset::make("KernelHeight", { 5, 3, 7, 1 })), + framework::dataset::make("KernelDepth", { 7, 5, 3, 1 })), + framework::dataset::make("NumKernels", { 5, 3, 1, 11 })), + framework::dataset::make("HasBias", { true, true, true, false })), + framework::dataset::make("Activation", ActivationLayerInfo())), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NDHWC))) +{ + validate(CLAccessor(_target), _reference, rel_tolerance_fp32, 0.0, abs_tolerance_fp32); +} + +// clang-format on +// *INDENT-ON* +TEST_SUITE_END() // FP32 + +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolution3DQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(7U, 5U, 3U, 13U, 3U), + TensorShape(15U, 7U, 11U, 7U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U) + }), + framework::dataset::make("StrideX", { 1, 3, 2, 1 })), + framework::dataset::make("StrideY", { 2, 1, 3, 1 })), + framework::dataset::make("StrideZ", { 3, 2, 1, 1 })), + framework::dataset::make("PadX", { 0, 2, 1, 0 })), + framework::dataset::make("PadY", { 1, 0, 2, 0 })), + framework::dataset::make("PadZ", { 2, 1, 0, 0 })), + framework::dataset::make("KernelWidth", { 3, 7, 5, 1 })), + framework::dataset::make("KernelHeight", { 5, 3, 7, 1 })), + framework::dataset::make("KernelDepth", { 7, 5, 3, 1 })), + framework::dataset::make("NumKernels", { 5, 3, 1, 11 })), + framework::dataset::make("HasBias", { true, true, true, false })), + framework::dataset::make("Activation", ActivationLayerInfo())), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("DataLayout", DataLayout::NDHWC)), + framework::dataset::make("SrcQuantizationInfo", QuantizationInfo(0.1f, 10))), + framework::dataset::make("WeightsQuantizationInfo", QuantizationInfo(0.3f, 20))), + framework::dataset::make("DstQuantizationInfo", QuantizationInfo(0.2f, 5)))) +{ + validate(CLAccessor(_target), _reference, abs_tolerance_qasymm8); +} + +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolution3DQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(7U, 5U, 3U, 13U, 3U), + TensorShape(15U, 7U, 11U, 7U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U) + }), + framework::dataset::make("StrideX", { 1, 3, 2, 1 })), + framework::dataset::make("StrideY", { 2, 1, 3, 1 })), + framework::dataset::make("StrideZ", { 3, 2, 1, 1 })), + framework::dataset::make("PadX", { 0, 2, 1, 0 })), + framework::dataset::make("PadY", { 1, 0, 2, 0 })), + framework::dataset::make("PadZ", { 2, 1, 0, 0 })), + framework::dataset::make("KernelWidth", { 3, 7, 5, 1 })), + framework::dataset::make("KernelHeight", { 5, 3, 7, 1 })), + framework::dataset::make("KernelDepth", { 7, 5, 3, 1 })), + framework::dataset::make("NumKernels", { 5, 3, 1, 11 })), + framework::dataset::make("HasBias", { true, true, true, false })), + framework::dataset::make("Activation", ActivationLayerInfo())), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataLayout", DataLayout::NDHWC)), + framework::dataset::make("SrcQuantizationInfo", QuantizationInfo(0.1f, 10))), + framework::dataset::make("WeightsQuantizationInfo", QuantizationInfo(0.3f, 20))), + framework::dataset::make("DstQuantizationInfo", QuantizationInfo(0.2f, 5)))) +{ + validate(CLAccessor(_target), _reference, abs_tolerance_qasymm8); +} + +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE_END() // NDHWC +TEST_SUITE_END() // DirectConvolution3D +TEST_SUITE_END() // CL + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp index 31eed7646c..8820a6a31e 100644 --- a/tests/validation/CL/ConvolutionLayer.cpp +++ b/tests/validation/CL/ConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,6 +22,7 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h" @@ -37,12 +38,16 @@ #include "tests/validation/Validation.h" #include "tests/validation/fixtures/ConvolutionLayerFixture.h" +/** Synced with tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp + * Please check there for any differences in the coverage + */ namespace arm_compute { namespace test { namespace validation { +using framework::dataset::make; namespace { class SmallConvolutionLayerDatasetCases final : public datasets::ConvolutionLayerDataset @@ -61,32 +66,32 @@ constexpr AbsoluteTolerance<float> tolerance_qasymm8(1); /**< T constexpr float tolerance_num = 0.07f; /**< Tolerance number */ /** CNN data types */ -const auto CNNDataTypes = framework::dataset::make("DataType", +const auto CNNDataTypes = make("DataType", { DataType::F16, - DataType::F32, - DataType::QASYMM8, - DataType::QASYMM8_SIGNED, + DataType::F32, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED, }); /** Grouped CNN data types */ -const auto GroupedCNNDataTypes = framework::dataset::make("DataType", +const auto GroupedCNNDataTypes = make("DataType", { DataType::F16, - DataType::F32 + DataType::F32 }); -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +const auto ActivationFunctionsDataset = make("ActivationInfo", { ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) }); -const auto ActivationFunctionsSmallDataset = framework::dataset::make("ActivationInfo", +const auto ActivationFunctionsSmallDataset = make("ActivationInfo", { ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) }); } // namespace @@ -96,7 +101,7 @@ TEST_SUITE(ConvolutionLayer) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32), // Select GEMM + make("InputInfo", { TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32), // Select GEMM TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32), // Select GEMM TensorInfo(TensorShape(23U, 27U, 5U, 4U), 1, DataType::F32), // Select GEMM TensorInfo(TensorShape(23U, 27U, 31U, 4U), 1, DataType::F32), // Select WINOGRAD @@ -106,7 +111,7 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32), // Select GEMM TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::QASYMM8_SIGNED), // Select GEMM }), - framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32), + make("WeightsInfo", { TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32), TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 31U, 21U), 1, DataType::F32), @@ -116,7 +121,7 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32), TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::QASYMM8_SIGNED), })), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32), + make("OutputInfo", { TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32), TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32), TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32), TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32), @@ -126,7 +131,7 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z TensorInfo(TensorShape(17U, 31U, 19U), 1, DataType::F32), TensorInfo(TensorShape(17U, 31U, 19U), 1, DataType::QASYMM8_SIGNED), })), - framework::dataset::make("ConvInfo", { PadStrideInfo(1, 2, 1, 1), + make("ConvInfo", { PadStrideInfo(1, 2, 1, 1), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), @@ -136,7 +141,7 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z PadStrideInfo(1, 1, 2, 2), PadStrideInfo(1, 1, 2, 2), })), - framework::dataset::make("GpuTarget", { GPUTarget::BIFROST, + make("GpuTarget", { GPUTarget::BIFROST, GPUTarget::MIDGARD, GPUTarget::G71, GPUTarget::G71, @@ -146,7 +151,7 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z GPUTarget::BIFROST, GPUTarget::BIFROST, })), - framework::dataset::make("Dilation", { Size2D(1U, 1U), + make("Dilation", { Size2D(1U, 1U), Size2D(1U, 1U), Size2D(1U, 1U), Size2D(1U, 1U), @@ -156,8 +161,8 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z Size2D(2U, 1U), Size2D(2U, 1U), })), - framework::dataset::make("EnableFastMath", { false, false, false, false, false, false, true, true, true })), - framework::dataset::make("Expected",{ ConvolutionMethod::GEMM, + make("EnableFastMath", { false, false, false, false, false, false, true, true, true })), + make("Expected",{ ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::WINOGRAD, @@ -188,15 +193,14 @@ template <typename T> using CLGEMMConvolutionLayerFixture = ConvolutionValidationFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; template <typename T> using CLGEMMConvolutionLayerMixedDataLayoutFixture = ConvolutionValidationFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T, true>; +template <typename T> +using CLConvolutionValidationWithPaddingFixture = ConvolutionValidationWithPaddingFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(Float) TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ReshapeWeights", { true })), make("DataType", DataType::F16)), make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsSmallDataset)) { // Validate output @@ -207,31 +211,40 @@ TEST_SUITE_END() // FP16 TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ReshapeWeights", { true })), make("DataType", DataType::F32)), make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLGEMMConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), - framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), - framework::dataset::make("Bias", TensorShape(2U))), - framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1, 1))), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType",DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsSmallDataset)) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLGEMMConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("Input", TensorShape(23U, 27U, 5U)), + make("Weights", TensorShape(3U, 3U, 5U, 2U))), + make("Bias", TensorShape(2U))), + make("Output", TensorShape(11U, 25U, 2U))), + make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), + make("Dilation", Size2D(1, 1))), + make("ReshapeWeights", { true })), + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } +FIXTURE_DATA_TEST_CASE(RunSmallWithPadding, CLConvolutionValidationWithPaddingFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerPrePaddingDataset(), + make("ReshapeWeights", { true })), + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() })), +make("PrePadLayer", { PaddingList({ { 1, 1 }, { 1, 1 } }) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float @@ -242,64 +255,108 @@ using CLGEMMConvolutionLayerQuantizedMixedDataLayoutFixture = ConvolutionValidat template <typename T> using CLGEMMConvolutionLayerQuantizedPerChannelFixture = ConvolutionValidationQuantizedPerChannelFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T, int8_t>; -const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationInfo", -{ - ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) -}); -const auto QuantizedActivationFunctionsSmallDataset = framework::dataset::make("ActivationInfo", -{ - ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) -}); - TEST_SUITE(Quantized) -const auto QuantizationData = framework::dataset::make("QuantizationInfo", +const auto QuantizationData = make("QuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(0.3f, 3), QuantizationInfo(1.1f, 10), }); + +/// @note: Every asymmetric quantized test has a version with or without activation because the quantization info given +/// is ignored when there is no activation. Instead of using the same quantization information for all the tensors, the +/// fixture generates separate quantization info for each input and the output tensor. +/// When we can also support dynamic quantization with the presence of activation, these two versions should be merged +/// again, with the explicitly specified quantization info removed +const auto NoActivation = make("ActivationInfo", ActivationLayerInfo()); + +const auto IgnoredQuantizationInfo = make("IgnoredQuantizationInfo", QuantizationInfo()); + +const auto QuantizedActivationFunctionsSmallDataset = make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) +}); + TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmallCases, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(SmallConvolutionLayerDatasetCases(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - QuantizationData), - QuantizedActivationFunctionsSmallDataset)) + combine(SmallConvolutionLayerDatasetCases(), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + IgnoredQuantizationInfo, + NoActivation)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + +FIXTURE_DATA_TEST_CASE(RunSmallCasesWithActivation, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, + combine(SmallConvolutionLayerDatasetCases(), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + QuantizationData, + QuantizedActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - QuantizationData), - QuantizedActivationFunctionsSmallDataset)) + combine(datasets::SmallConvolutionLayerDataset(), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + IgnoredQuantizationInfo, + NoActivation)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, + combine(datasets::SmallConvolutionLayerDataset(), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + QuantizationData, + QuantizedActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLGEMMConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), - framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), - framework::dataset::make("Bias", TensorShape(2U))), - framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1, 1))), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - QuantizationData), - QuantizedActivationFunctionsSmallDataset)) + combine( + make("Input", TensorShape(23U, 27U, 5U)), + make("Weights", TensorShape(3U, 3U, 5U, 2U)), + make("Bias", TensorShape(2U)), + make("Output", TensorShape(11U, 25U, 2U)), + make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0)), + make("Dilation", Size2D(1, 1)), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + IgnoredQuantizationInfo, + NoActivation)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, CLGEMMConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::ALL, + combine( + make("Input", TensorShape(23U, 27U, 5U)), + make("Weights", TensorShape(3U, 3U, 5U, 2U)), + make("Bias", TensorShape(2U)), + make("Output", TensorShape(11U, 25U, 2U)), + make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0)), + make("Dilation", Size2D(1, 1)), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + QuantizationData, + QuantizedActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); @@ -307,44 +364,78 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLGEMMConvolutionLayerQuantizedMixedD TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - QuantizationData), - QuantizedActivationFunctionsSmallDataset)) + combine(datasets::SmallConvolutionLayerDataset(), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8_SIGNED), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + IgnoredQuantizationInfo, + NoActivation)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, CLGEMMConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, + combine(datasets::SmallConvolutionLayerDataset(), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8_SIGNED), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + QuantizationData, + QuantizedActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLGEMMConvolutionLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), - framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), - framework::dataset::make("Bias", TensorShape(2U))), - framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1, 1))), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - QuantizationData), - QuantizedActivationFunctionsSmallDataset)) + combine( + make("Input", TensorShape(23U, 27U, 5U)), + make("Weights", TensorShape(3U, 3U, 5U, 2U)), + make("Bias", TensorShape(2U)), + make("Output", TensorShape(11U, 25U, 2U)), + make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0)), + make("Dilation", Size2D(1, 1)), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8_SIGNED), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + IgnoredQuantizationInfo, + NoActivation)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, CLGEMMConvolutionLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::ALL, + combine( + make("Input", TensorShape(23U, 27U, 5U)), + make("Weights", TensorShape(3U, 3U, 5U, 2U)), + make("Bias", TensorShape(2U)), + make("Output", TensorShape(11U, 25U, 2U)), + make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0)), + make("Dilation", Size2D(1, 1)), + make("ReshapeWeights", { true }), + make("DataType", DataType::QASYMM8_SIGNED), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + QuantizationData, + QuantizedActivationFunctionsSmallDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE(QSYMM8_PER_CHANNEL) +const auto QuantizedActivationFunctionsSmallPerChannelDataset = make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) +}); + FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLGEMMConvolutionLayerQuantizedPerChannelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", { DataType::QASYMM8_SIGNED })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ReshapeWeights", { true })), + make("DataType", { DataType::QASYMM8_SIGNED })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), QuantizationData), - QuantizedActivationFunctionsSmallDataset), - framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) + QuantizedActivationFunctionsSmallPerChannelDataset), + make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); @@ -352,12 +443,12 @@ FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLGEMMConvolutionLayerQuantizedPerChannel FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", { DataType::QASYMM8 })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ReshapeWeights", { true })), + make("DataType", { DataType::QASYMM8 })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), QuantizationData), - QuantizedActivationFunctionsSmallDataset), - framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) + QuantizedActivationFunctionsSmallPerChannelDataset), + make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); @@ -376,9 +467,7 @@ TEST_SUITE(Float) TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMGroupedConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallGroupedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), + make("ReshapeWeights", { true })), make("DataType", DataType::F32)), make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsSmallDataset)) { // Validate output @@ -387,9 +476,9 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMGroupedConvolutionLayerFixture<float>, fr FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMGroupedConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeGroupedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), + make("ReshapeWeights", { true })), + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output @@ -400,9 +489,7 @@ TEST_SUITE_END() // FP32 TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMGroupedConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallGroupedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), + make("ReshapeWeights", { true })), make("DataType", DataType::F16)), make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsSmallDataset)) { // Validate output @@ -411,9 +498,9 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMGroupedConvolutionLayerFixture<half>, fra FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMGroupedConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeGroupedConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), + make("ReshapeWeights", { true })), + make("DataType", DataType::F16)), + make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output diff --git a/tests/validation/CL/DeconvolutionLayer.cpp b/tests/validation/CL/DeconvolutionLayer.cpp index 15962b588d..d1508fd902 100644 --- a/tests/validation/CL/DeconvolutionLayer.cpp +++ b/tests/validation/CL/DeconvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -53,27 +53,29 @@ const auto data9x9_small_asymm = framework::dataset::make("InputShape", TensorSh *framework::dataset::make("PadLeft", 3) *framework::dataset::make("PadRight", 4) *framework::dataset::make("PadTop", 3) *framework::dataset::make("PadBottom", 4) *framework::dataset::make("NumKernels", { 1 }); -const auto data9x9_large_asymm = framework::dataset::make("InputShape", TensorShape{ 640U, 360U, 56U, 1U }) *framework::dataset::make("StrideX", 2) *framework::dataset::make("StrideY", - 2) - *framework::dataset::make("PadLeft", 3) - *framework::dataset::make("PadRight", 4) *framework::dataset::make("PadTop", 3) *framework::dataset::make("PadBottom", 4) *framework::dataset::make("NumKernels", { 1 }); - -const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 3) - * framework::dataset::make("PadY", 0, 3) * framework::dataset::make("NumKernels", { 3 }); - -const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2) +const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4, 2) * framework::dataset::make("StrideY", 2, 4) * framework::dataset::make("PadX", 1, 3) * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 }); +const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 2, 4) * framework::dataset::make("StrideY", 1, 4, 2) * framework::dataset::make("PadX", 1, 2) + * framework::dataset::make("PadY", 1, 3) * framework::dataset::make("NumKernels", { 3 }); + const auto data3x3_asymm = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadLeft", 0, 1) * framework::dataset::make("PadRight", 0, 1) * framework::dataset::make("PadTop", 0, 1) * framework::dataset::make("PadBottom", 0, 1) * framework::dataset::make("NumKernels", { 3 }); const auto data3x3_precommit = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadX", 0, 2) * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 }); +const auto data3x3_precommit_large_channels = datasets::SmallDeconvolutionShapesWithLargerChannels() * framework::dataset::make("StrideX", 2) * framework::dataset::make("StrideY", 2) + * framework::dataset::make("PadX", 1) + * framework::dataset::make("PadY", 2) * framework::dataset::make("NumKernels", { 5 }); + const auto data2x2_precommit = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 2) * framework::dataset::make("StrideY", 2) * framework::dataset::make("PadX", 1) * framework::dataset::make("PadY", 1) * framework::dataset::make("NumKernels", { 3 }); -const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 1) +const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4, 2) * framework::dataset::make("StrideY", 2, 4) * framework::dataset::make("PadX", 0, 1) + * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("NumKernels", { 3 }); + +const auto data5x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 2, 4) * framework::dataset::make("StrideY", 1, 4, 2) * framework::dataset::make("PadX", 0, 1) * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("NumKernels", { 3 }); const auto data_layouts_dataset = framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }); @@ -90,9 +92,15 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid weights shape TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16), // Non supported data type - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid bias shape + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid bias shape TensorInfo(TensorShape(13U, 11U, 4U, 3U), 1, DataType::F32), // Window shrink TensorInfo(TensorShape(32U, 16U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 13U, 27U), 1, DataType::F32, DataLayout::NHWC), // Mismatching data type + TensorInfo(TensorShape(2U, 13U, 27U), 1, DataType::F32, DataLayout::NHWC), // Invalid weights shape + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F16, DataLayout::NHWC), // Non supported data type + TensorInfo(TensorShape(2U, 13U, 27U), 1, DataType::F32, DataLayout::NHWC), // Invalid bias shape + TensorInfo(TensorShape(4U, 11U, 13U, 3U), 1, DataType::F32, DataLayout::NHWC), // Window shrink + TensorInfo(TensorShape(2U, 16U, 32U), 1, DataType::F32, DataLayout::NHWC), }), framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F16), TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), @@ -100,6 +108,12 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(3U, 2U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 4U), 1, DataType::F32), TensorInfo(TensorShape(1U, 1U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 2U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 2U, 2U, 4U), 1, DataType::F32, DataLayout::NHWC), })), framework::dataset::make("BiasInfo", { TensorInfo(TensorShape(1U), 1, DataType::F16), TensorInfo(TensorShape(1U), 1, DataType::F32), @@ -107,6 +121,12 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(25U, 11U), 1, DataType::F32), TensorInfo(TensorShape(1U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), + TensorInfo(TensorShape(1U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(1U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(1U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(25U, 11U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(1U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), })), framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F16), TensorInfo(TensorShape(25U, 10U, 2U), 1, DataType::F32), @@ -114,6 +134,12 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(13U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(11U, 9U, 1U, 3U), 1, DataType::F32), TensorInfo(TensorShape(32U, 16U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 11U, 25U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 10U, 25U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 13U, 13U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(1U, 9U, 11U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U, 43U, 91U), 1, DataType::F32, DataLayout::NHWC), })), framework::dataset::make("PadStrideInfo", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), @@ -121,8 +147,15 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(3, 3, 2, 2), })), - framework::dataset::make("Expected", { false, false, false, false, false, true })), + framework::dataset::make("Expected", { false, false, false, false, false, true, // NCHW + false, false, false, false, false, true })), // NHWC input_info, weights_info, bias_info, output_info, pad_info, expected) { bool is_valid = bool(CLDeconvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pad_info)); @@ -149,6 +182,9 @@ using CLDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<CLTensor, template <typename T> using CLDeconvolutionLayerAsymmFixture9x9 = DeconvolutionValidationAsymmFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 9, 9>; +template <typename T> +using CLDeconvolutionLayerFixture5x1 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 5, 1>; + TEST_SUITE(Float) TEST_SUITE(FP32) @@ -171,6 +207,17 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerFixture3x3<float>, framewor // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } + +FIXTURE_DATA_TEST_CASE(RunSmallWithLargeChannels, CLDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data3x3_precommit_large_channels, + framework::dataset::make("DataType", + DataType::F32)), + data_layouts_dataset), + framework::dataset::make("AddBias", { true }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_fp32); +} + FIXTURE_DATA_TEST_CASE(RunAsymm, CLDeconvolutionLayerAsymmFixture3x3<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3_asymm, framework::dataset::make("DataType", DataType::F32)), data_layouts_dataset), @@ -180,8 +227,8 @@ FIXTURE_DATA_TEST_CASE(RunAsymm, CLDeconvolutionLayerAsymmFixture3x3<float>, fra validate(CLAccessor(_target), _reference, tolerance_fp32); } FIXTURE_DATA_TEST_CASE(RunLarge, CLDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::F32)), - data_layouts_dataset), - add_bias_dataset)) + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); @@ -202,7 +249,7 @@ TEST_SUITE_END() // W2x2 TEST_SUITE(W1x1) FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture1x1<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data1x1, framework::dataset::make("DataType", DataType::F32)), data_layouts_dataset), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); @@ -218,6 +265,17 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerAsymmFixture9x9<float>, fra validate(CLAccessor(_target), _reference, tolerance_fp32); } TEST_SUITE_END() // W9x9 + +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture5x1<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data5x1, framework::dataset::make("DataType", DataType::F32)), + data_layouts_dataset), + framework::dataset::make("AddBias", { true }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // FP32 TEST_SUITE(FP16) @@ -225,7 +283,7 @@ TEST_SUITE(FP16) TEST_SUITE(W4x4) FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture4x4<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data4x4, framework::dataset::make("DataType", DataType::F16)), data_layouts_dataset), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); @@ -242,8 +300,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerFixture3x3<half>, framework validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } FIXTURE_DATA_TEST_CASE(RunLarge, CLDeconvolutionLayerFixture3x3<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::F16)), - data_layouts_dataset), - add_bias_dataset)) + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); @@ -264,13 +322,23 @@ TEST_SUITE_END() // W2x2 TEST_SUITE(W1x1) FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture1x1<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data1x1, framework::dataset::make("DataType", DataType::F16)), data_layouts_dataset), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture5x1<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data5x1, framework::dataset::make("DataType", DataType::F16)), + data_layouts_dataset), + framework::dataset::make("AddBias", { true }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float @@ -287,6 +355,9 @@ template <typename T> using CLDeconvolutionLayerQuantizedFixture1x1 = DeconvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 1, 1>; template <typename T> +using CLDeconvolutionLayerQuantizedFixture5x1 = DeconvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 5, 1>; + +template <typename T> using CLDeconvolutionLayerQuantizedPerChannelFixture4x4 = DeconvolutionValidationQuantizedPerChannelFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, int8_t, 4, 4>; template <typename T> @@ -298,6 +369,9 @@ using CLDeconvolutionLayerQuantizedPerChannelFixture2x2 = DeconvolutionValidatio template <typename T> using CLDeconvolutionLayerQuantizedPerChannelFixture1x1 = DeconvolutionValidationQuantizedPerChannelFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, int8_t, 1, 1>; +template <typename T> +using CLDeconvolutionLayerQuantizedPerChannelFixture5x1 = DeconvolutionValidationQuantizedPerChannelFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, int8_t, 5, 1>; + TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) @@ -307,7 +381,7 @@ FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerQuantizedFixture4x4<uint8_t>, fr data_layouts_dataset), framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10), QuantizationInfo(2.f / 255.f, 5) })), framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 5), QuantizationInfo(4.f / 255.f, 10) })), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); @@ -327,12 +401,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedFixture3x3<uint8_t validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } FIXTURE_DATA_TEST_CASE(RunLarge, CLDeconvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data3x3, - framework::dataset::make("DataType", - DataType::QASYMM8)), - data_layouts_dataset), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10), QuantizationInfo(2.f / 255.f, 128) })), framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 128), QuantizationInfo(4.f / 255.f, 128) })), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); @@ -358,13 +431,26 @@ FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerQuantizedFixture1x1<uint8_t>, fr data_layouts_dataset), framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 0), QuantizationInfo(2.f / 255.f, 0) })), framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 0), QuantizationInfo(4.f / 255.f, 0) })), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerQuantizedFixture5x1<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data5x1, framework::dataset::make("DataType", + DataType::QASYMM8)), + data_layouts_dataset), + framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10), QuantizationInfo(2.f / 255.f, 5) })), + framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 5), QuantizationInfo(4.f / 255.f, 10) })), + framework::dataset::make("AddBias", { true }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) @@ -378,7 +464,7 @@ FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerQuantizedFixture4x4<int8_t>, fra data_layouts_dataset), framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10), QuantizationInfo(2.f / 255.f, 5) })), framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 5), QuantizationInfo(4.f / 255.f, 10) })), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); @@ -400,12 +486,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedFixture3x3<int8_t> } FIXTURE_DATA_TEST_CASE(RunLarge, CLDeconvolutionLayerQuantizedFixture3x3<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data3x3, - framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - data_layouts_dataset), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, -10), QuantizationInfo(2.f / 255.f, 127) })), framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 64), QuantizationInfo(4.f / 255.f, -128) })), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); @@ -431,13 +516,26 @@ FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerQuantizedFixture1x1<int8_t>, fra data_layouts_dataset), framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 0), QuantizationInfo(2.f / 255.f, 0) })), framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 0), QuantizationInfo(4.f / 255.f, 0) })), - add_bias_dataset)) + framework::dataset::make("AddBias", { true }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerQuantizedFixture5x1<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data5x1, framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED)), + data_layouts_dataset), + framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10), QuantizationInfo(2.f / 255.f, 5) })), + framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 5), QuantizationInfo(4.f / 255.f, 10) })), + framework::dataset::make("AddBias", { true }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // QASYMM8_SIGNED const auto input_qinfo_dataset = framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10) }); @@ -448,23 +546,23 @@ const auto output_signed_qinfo_dataset = framework::dataset::make("OutputQuantiz TEST_SUITE(QSYMM8_PER_CHANNEL) TEST_SUITE(W4x4) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture4x4<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data4x4, +FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture4x4<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data4x4, framework::dataset::make("DataType", DataType::QASYMM8)), data_layouts_dataset), input_qinfo_dataset), output_qinfo_dataset), - add_bias_dataset), + framework::dataset::make("AddBias", { true })), framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture4x4<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data4x4, +FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture4x4<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data4x4, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), data_layouts_dataset), input_signed_qinfo_dataset), output_signed_qinfo_dataset), - add_bias_dataset), + framework::dataset::make("AddBias", { true })), framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output @@ -473,32 +571,57 @@ FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFi TEST_SUITE_END() // W4x4 TEST_SUITE(W3x3) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture3x3<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data3x3, +FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::QASYMM8)), data_layouts_dataset), input_qinfo_dataset), output_qinfo_dataset), - add_bias_dataset), + framework::dataset::make("AddBias", { true })), framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture3x3<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data3x3, +FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture3x3<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), data_layouts_dataset), input_signed_qinfo_dataset), output_signed_qinfo_dataset), - add_bias_dataset), + framework::dataset::make("AddBias", { true })), framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } + +FIXTURE_DATA_TEST_CASE(RunSmallSignedPrecommit, CLDeconvolutionLayerQuantizedPerChannelFixture2x2<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(data3x3_precommit, + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + data_layouts_dataset), + input_signed_qinfo_dataset), + output_signed_qinfo_dataset), + add_bias_dataset), + framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); +} TEST_SUITE_END() // W3x3 +FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture2x2<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(data3x3_precommit, + framework::dataset::make("DataType", DataType::QASYMM8)), + data_layouts_dataset), + input_qinfo_dataset), + output_qinfo_dataset), + add_bias_dataset), + framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); +} + TEST_SUITE(W2x2) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture2x2<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data2x2_precommit, +FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture2x2<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(data2x2_precommit, framework::dataset::make("DataType", DataType::QASYMM8)), data_layouts_dataset), input_qinfo_dataset), @@ -509,7 +632,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture2 // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture2x2<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data2x2_precommit, +FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture2x2<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(data2x2_precommit, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), data_layouts_dataset), input_signed_qinfo_dataset), @@ -523,23 +646,23 @@ FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFi TEST_SUITE_END() // W2x2 TEST_SUITE(W1x1) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture1x1<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data1x1, +FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture1x1<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data1x1, framework::dataset::make("DataType", DataType::QASYMM8)), data_layouts_dataset), input_qinfo_dataset), output_qinfo_dataset), - add_bias_dataset), + framework::dataset::make("AddBias", { false })), framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture1x1<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data1x1, +FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture1x1<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data1x1, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), data_layouts_dataset), input_signed_qinfo_dataset), output_signed_qinfo_dataset), - add_bias_dataset), + framework::dataset::make("AddBias", { true })), framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output @@ -547,6 +670,31 @@ FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFi } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDeconvolutionLayerQuantizedPerChannelFixture5x1<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data5x1, + framework::dataset::make("DataType", DataType::QASYMM8)), + data_layouts_dataset), + input_qinfo_dataset), + output_qinfo_dataset), + framework::dataset::make("AddBias", { true })), + framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); +} +FIXTURE_DATA_TEST_CASE(RunSmallSigned, CLDeconvolutionLayerQuantizedPerChannelFixture5x1<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(data5x1, + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + data_layouts_dataset), + input_signed_qinfo_dataset), + output_signed_qinfo_dataset), + framework::dataset::make("AddBias", { false })), + framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8, tolerance_num); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // QSYMM8_PER_CHANNEL TEST_SUITE_END() // Quantized diff --git a/tests/validation/CL/DepthConvertLayer.cpp b/tests/validation/CL/DepthConvertLayer.cpp index 8f14337b27..490b38ccf6 100644 --- a/tests/validation/CL/DepthConvertLayer.cpp +++ b/tests/validation/CL/DepthConvertLayer.cpp @@ -62,7 +62,7 @@ TEST_SUITE(DepthConvertLayer) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Invalid data type combination + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Support upcasting from QASYMM8 to S16 TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Mismatching shapes TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8), // Invalid shift @@ -84,7 +84,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( ConvertPolicy::WRAP, })), framework::dataset::make("Shift",{ 0, 0, 0, 1, 1, 0, })), - framework::dataset::make("Expected", { false, false, false, false, false, true})), + framework::dataset::make("Expected", { true, false, false, false, false, true})), input_info, output_info, policy, shift, expected) { ARM_COMPUTE_EXPECT(bool(CLDepthConvertLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), policy, shift)) == expected, framework::LogLevel::ERRORS); diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp index c88f7c1624..d4dbcec9d9 100644 --- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp +++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -41,6 +41,9 @@ namespace test { namespace validation { + +using framework::dataset::make; + namespace { RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ @@ -48,16 +51,47 @@ constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */ constexpr float tolerance_num = 0.05f; /**< Tolerance number */ -const auto depth_multipliers = framework::dataset::make("DepthMultiplier", { 1, 2, 5 }); -const auto large_depth_multipliers = framework::dataset::make("DepthMultiplier", { 1, 2, 5, 8 }); +const auto depth_multipliers = make("DepthMultiplier", { 1, 4 }); +const auto large_depth_multipliers = make("DepthMultiplier", { 2, 5, 8 }); -//Activation Functions -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +// Activation Functions +const auto NoActivation = make("ActivationInfo", ActivationLayerInfo()); + +const auto ActivationFunctionsSmallDataset = make("ActivationInfo", { ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 2.f, 0.f) +}); + +const auto ActivationFunctionsDataset = make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.8f, -0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SOFT_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ABS), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SQUARE), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::HARD_SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 2.f, 1.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::GELU) +}); + +const auto ActivationFunctionsQuantizedSmallDataset = make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 2.f, 0.f) +}); + +const auto ActivationFunctionsQuantizedDataset = make("ActivationInfo", +{ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f, 0.f) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 2.3f, -1.5f), }); + +const auto IgnoredQuantizationInfo = make("IgnoredQuantizationInfo", QuantizationInfo()); + } // namespace TEST_SUITE(CL) @@ -65,85 +99,85 @@ TEST_SUITE(DepthwiseConvolutionLayer) // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights - TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32), // Mismatching input feature maps - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching depth multiplier - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases dimensions - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid output size - TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // patch size bigger than input width - TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // dilation < 1 - TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::QASYMM8), - }), - framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16), - TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::QASYMM8), - })), - framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32), - TensorInfo(TensorShape(2U), 1, DataType::F32), - TensorInfo(TensorShape(2U), 1, DataType::F32), - TensorInfo(TensorShape(4U), 1, DataType::F32), - TensorInfo(TensorShape(2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(2U), 1, DataType::F32), - TensorInfo(TensorShape(16U), 1, DataType::F32), - TensorInfo(TensorShape(16U), 1, DataType::F32), - TensorInfo(TensorShape(16U), 1, DataType::F32), - TensorInfo(TensorShape(24U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 11U, 24U), 1, DataType::QASYMM8), - })), - framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 1, 0), - })), - framework::dataset::make("DepthMultiplier", { 1, - 1, - 3, - 1, - 1, - 1, - 2, - 2, - 2, - 3, - })), - framework::dataset::make("Dilation", { Size2D(1U, 1U), - Size2D(1U, 1U), - Size2D(1U, 1U), - Size2D(1U, 1U), - Size2D(1U, 1U), - Size2D(1U, 1U), - Size2D(20U, 1U), - Size2D(0U, 1U), - Size2D(1U, 1U), - Size2D(1U, 1U), - })), - framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, true, true })), +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights + TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32), // Mismatching input feature maps + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching depth multiplier + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases dimensions + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid output size + TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // patch size bigger than input width + TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // dilation < 1 + TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::QASYMM8), + }), + make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::QASYMM8), + }), + make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(4U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(16U), 1, DataType::F32), + TensorInfo(TensorShape(16U), 1, DataType::F32), + TensorInfo(TensorShape(16U), 1, DataType::F32), + TensorInfo(TensorShape(24U), 1, DataType::S32), + }), + make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 11U, 24U), 1, DataType::QASYMM8), + }), + make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 1, 0), + }), + make("DepthMultiplier", { 1, + 1, + 3, + 1, + 1, + 1, + 2, + 2, + 2, + 3, + }), + make("Dilation", { Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(20U, 1U), + Size2D(0U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + }), + make("Expected", { false, false, false, false, false, false, false, false, true, true })), input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier, dilation, expected) { bool is_valid = bool(CLDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(true), &weights_info.clone()->set_is_resizable(true), &biases_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), conv_info, depth_multiplier,ActivationLayerInfo(), dilation)); @@ -156,48 +190,32 @@ template <typename T> using CLDepthwiseConvolutionLayerFixture = DepthwiseConvolutionLayerValidationFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer, T>; template <typename T> using CLDepthwiseConvolutionLayerMixedDataLayoutFixture = DepthwiseConvolutionLayerValidationFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer, T, true>; +template <typename T> +using CLDepthwiseConvolutionLayerInPlaceFixture = DepthwiseConvolutionLayerValidationFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer, T, false, true>; TEST_SUITE(Float) TEST_SUITE(FP16) TEST_SUITE(W3x3) TEST_SUITE(NCHW) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(framework::dataset::concat(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), - datasets::SmallDepthwiseConvolutionLayerDataset3x3NCHW()), - depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", DataLayout::NCHW)), - ActivationFunctionsDataset)) -{ - validate(CLAccessor(_target), _reference, tolerance_f16); -} -FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", DataLayout::NCHW)), - ActivationFunctionsDataset)) + combine( + framework::dataset::concat(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + datasets::SmallDepthwiseConvolutionLayerDataset3x3NCHW()), + depth_multipliers, + make("DataType", DataType::F16), + make("DataLayout", DataLayout::NCHW), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16); } TEST_SUITE(Dilation) -FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), - depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) -{ - validate(CLAccessor(_target), _reference, tolerance_f16); -} -FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, + combine( + datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::F16), + make("DataLayout", { DataLayout::NCHW }), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16); } @@ -206,41 +224,42 @@ TEST_SUITE_END() // NCHW TEST_SUITE(NHWC) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), - depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - ActivationFunctionsDataset)) + combine( + datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::F16), + make("DataLayout", DataLayout::NHWC), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, + combine( + datasets::LargeDepthwiseConvolutionLayerDataset3x3Fp16Subset(), + large_depth_multipliers, + make("DataType", DataType::F16), + make("DataLayout", DataLayout::NHWC), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f16); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3Fp16Subset(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f16); } @@ -251,19 +270,33 @@ TEST_SUITE_END() // W3x3 TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDatasetFp16Subset(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 2 }), + make("DataType", DataType::F16), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } @@ -271,25 +304,38 @@ FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, f TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), + combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDatasetFp16Subset(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } TEST_SUITE_END() // Dilation TEST_SUITE_END() // Generic + +TEST_SUITE(InPlace) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerInPlaceFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(datasets::SmallInPlaceDepthwiseConvolutionLayerDataset(), + make("DepthMultiplier", { 1 })), + make("DataType", + DataType::F16)), + make("DataLayout", { DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) +{ + validate(CLAccessor(_src), _reference, tolerance_f16, tolerance_num); +} +TEST_SUITE_END() // InPlace TEST_SUITE_END() // FP16 TEST_SUITE(FP32) @@ -299,94 +345,76 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, combine(combine(combine(combine(framework::dataset::concat(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), datasets::SmallDepthwiseConvolutionLayerDataset3x3NCHW()), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NCHW)), - ActivationFunctionsDataset)) -{ - validate(CLAccessor(_target), _reference, tolerance_f32); -} -FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", - DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NCHW)), - ActivationFunctionsDataset)) + make("DataLayout", DataLayout::NCHW)), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", - DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NCHW)), - ActivationFunctionsDataset)) -{ - validate(CLAccessor(_target), _reference, tolerance_f32); -} -FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NCHW)), - ActivationFunctionsDataset)) + make("DataLayout", DataLayout::NCHW)), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } - TEST_SUITE_END() // Dilation TEST_SUITE_END() // NCHW + TEST_SUITE(NHWC) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - ActivationFunctionsDataset)) + make("DataLayout", DataLayout::NHWC)), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout, CLDepthwiseConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", { 2 })), - framework::dataset::make("DataType", + make("DepthMultiplier", { 2 })), + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + make("DataLayout", DataLayout::NHWC)), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - ActivationFunctionsDataset)) + make("DataLayout", DataLayout::NHWC)), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f32); } + TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - ActivationFunctionsDataset)) + make("DataLayout", DataLayout::NHWC)), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", DataLayout::NHWC)), - ActivationFunctionsDataset)) + make("DataLayout", DataLayout::NHWC)), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f32); } @@ -397,19 +425,45 @@ TEST_SUITE_END() // W3x3 TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } + FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE_NEW(RunLargeKernelSize, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(datasets::LargeKernelSizeDepthwiseConvolutionLayerNHWCDataset(), + make("DepthMultiplier", { 1 })), + make("DataType", + DataType::F32)), + make("DataLayout", { DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 2 }), + make("DataType", DataType::F32), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } @@ -417,25 +471,38 @@ FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_f32); } TEST_SUITE_END() // Dilation TEST_SUITE_END() // Generic + +TEST_SUITE(InPlace) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerInPlaceFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(datasets::SmallInPlaceDepthwiseConvolutionLayerDataset(), + make("DepthMultiplier", { 1 })), + make("DataType", + DataType::F32)), + make("DataLayout", { DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) +{ + validate(CLAccessor(_src), _reference, tolerance_f32); +} +TEST_SUITE_END() // InPlace TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float @@ -450,47 +517,84 @@ TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), // NCHW is tested with int8 - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), // NCHW is tested with int8 + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 128), QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(1.f, 128) }), + make("DataLayout", { DataLayout::NHWC }), // NCHW is tested with int8 + ActivationFunctionsQuantizedSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.7f, 2) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseConvolutionLayerDataset(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 2U }), + make("DataType", DataType::QASYMM8), + make("SrcQuantizationInfo", { QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.1f, 128) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.8, 1) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), // NCHW is tested with int8 - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), // NCHW is tested with int8 + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.8, 1) }), + make("DataLayout", { DataLayout::NHWC }), // NCHW is tested with int8 + ActivationFunctionsQuantizedSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(1.3f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.9f, 11) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } @@ -498,58 +602,80 @@ TEST_SUITE_END() // Dilation TEST_SUITE_END() // Generic TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout, CLDepthwiseConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", { 2 })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + make("DepthMultiplier", { 2 }), + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } @@ -560,36 +686,73 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, CLDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) }), + make("DataLayout", { DataLayout::NCHW }), + ActivationFunctionsQuantizedSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout, CLDepthwiseConvolutionLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - framework::dataset::make("DepthMultiplier", { 2 })), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + make("DepthMultiplier", { 2 }), + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 2U }), + make("DataType", DataType::QASYMM8_SIGNED), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.8, 1) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, CLDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(2.2f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.8, 1) }), + make("DataLayout", { DataLayout::NCHW }), + ActivationFunctionsQuantizedSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } @@ -602,24 +765,40 @@ TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.7f, 2) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.7f, 2) })), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 2U }), + make("SrcDataType", DataType::QASYMM8_SIGNED), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } @@ -627,24 +806,24 @@ TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.8, 1) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.8, 1) })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.9f, 11) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.9f, 11) })), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } @@ -654,24 +833,24 @@ TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } @@ -679,24 +858,24 @@ TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsSmallDataset)) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerQuantizedPerChannelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("SrcDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), - framework::dataset::make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("SrcDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("SrcQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(CLAccessor(_target), _reference, tolerance_qasymm8); } diff --git a/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp b/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp index f640ee2b18..012018c0fc 100644 --- a/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp +++ b/tests/validation/CL/DepthwiseConvolutionLayerNative.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,6 +42,7 @@ namespace test { namespace validation { +using framework::dataset::make; using namespace arm_compute::misc::shape_calculator; // Create function for CLDepthwiseConvolutionLayerNativeKernel @@ -62,64 +63,89 @@ RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.01f)); constexpr float abs_tolerance_f16(0.03f); /** Width values to test - Precommit */ -const auto width_values_precommit = framework::dataset::make("width", { 1U, 17U, 32U } ); +const auto width_values_precommit = make("width", { 1U, 33U } ); /** Width values to test - Nightly */ -const auto width_values_nightly = framework::dataset::make("width", { 53U, 47U } ); +const auto width_values_nightly = make("width", { 53U, 47U } ); /** Height values to test - Precommit */ -const auto height_values_precommit = framework::dataset::make("height", { 19U } ); +const auto height_values_precommit = make("height", { 19U } ); /** Height values to test - Nightly */ -const auto height_values_nightly = framework::dataset::make("height", { 39U, 43U } ); +const auto height_values_nightly = make("height", { 39U, 43U } ); /** Channel values to test - Precommit */ -const auto channel_values_precommit = framework::dataset::make("channels", { 15U }); +const auto channel_values_precommit = make("channels", { 15U }); /** Channel values to test - Nightly */ -const auto channel_values_nightly = framework::dataset::make("channels", { 33U, 19U }); +const auto channel_values_nightly = make("channels", { 33U, 19U }); + +/** Channel values to test with cl_image support - Precommit */ +const auto channel_values_export_to_cl_image_precommit = make("channels", { 16U }); + +/** Channel values to test with cl_image support - Nightly */ +const auto channel_values_export_to_cl_image_nightly = make("channels", { 32U }); /** Batch values to test - Precommit */ -const auto batch_values_precommit = framework::dataset::make("batch", { 1U, 2U }); +const auto batch_values_precommit = make("batch", { 1U, 2U }); /** Batch values to test - Nightly */ -const auto batch_values_nightly = framework::dataset::make("batch", { 1U, 3U }); +const auto batch_values_nightly = make("batch", { 3U }); /** Kernel size values to test - Precommit */ -const auto kernel_sz_values_precommit = framework::dataset::make("kernel_size", { Size2D(1U, 1U), Size2D(1U, 3U), Size2D(5U, 5U) }); +const auto kernel_sz_values_precommit = make("kernel_size", { Size2D(1U, 1U), Size2D(1U, 3U), Size2D(5U, 5U) }); /** Kernel size values to test - Nightly */ -const auto kernel_sz_values_nightly = framework::dataset::make("kernel_size", { Size2D(3U, 5U), Size2D(5U, 1U), Size2D(1U, 7U), Size2D(9U, 7U) }); +const auto kernel_sz_values_nightly = make("kernel_size", { Size2D(3U, 5U), Size2D(5U, 1U), Size2D(1U, 7U), Size2D(9U, 7U) }); /** Depth multiplier values to test - All */ -const auto depth_multiplier_values = framework::dataset::make("depth_multiplier", {3U}); +const auto depth_multiplier_values = make("depth_multiplier", {3U}); /** Dilation values to test - All */ -const auto dilation_values = framework::dataset::make("dilation", { Size2D(1U, 1U), Size2D(3U, 3U) }); +const auto dilation_values = make("dilation", { Size2D(1U, 1U), Size2D(3U, 3U) }); /** Stride values to test - All */ -const auto stride_values = framework::dataset::make("stride", { Size2D(1U, 1U), Size2D(3U, 2U) }); +const auto stride_values = make("stride", { Size2D(1U, 1U), Size2D(3U, 2U) }); -/** Padding values to test - All */ -const auto padding_valid_values = framework::dataset::make("padding_valid", { true, false }); +/** Padding values to test - Precommit */ +const auto padding_valid_values = make("padding_valid", { true, false }); -/** Data type values to test - All */ -const auto data_type_values = framework::dataset::make("data_type", { DataType::F32, DataType::F16 }); +/** Padding values to test - Nightly */ +const auto padding_valid_values_nightly = make("padding_valid", { false }); /** Data layout values to test - All */ -const auto data_layout_values = framework::dataset::make("data_layout", { DataLayout::NHWC }); +const auto data_layout_values = make("data_layout", { DataLayout::NHWC }); /** N0 values to test - Precommit */ -const auto n0_values_precommit = framework::dataset::make("N0", {2, 4}); +const auto n0_values_precommit = make("N0", {2, 4}); /** N0 values to test - Nightly */ -const auto n0_values_nightly = framework::dataset::make("N0", {3, 8}); +const auto n0_values_nightly = make("N0", {3, 8}); + +/** N0 values to test with cl_image support - Precommit */ +const auto n0_values_export_to_cl_image_precommit = make("N0", {4}); + +/** N0 values to test with cl_image support - Nightly */ +const auto n0_values_export_to_cl_image_nightly = make("N0", {8}); -/** Activation values to test */ -const auto act_values = framework::dataset::make("Activation", +/** Activation values to test in precommit */ +const auto act_values = make("Activation", { ActivationLayerInfo() }); + +const auto activations_rest = make("Activation", { - ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f, 0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.8f, -0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SOFT_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ABS), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SQUARE), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::HARD_SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 2.f, 1.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::GELU) }); } // namespace @@ -129,94 +155,264 @@ TEST_SUITE(DepthwiseConvolutionLayerNative) TEST_SUITE(Float) TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( width_values_precommit, height_values_precommit), channel_values_precommit), batch_values_precommit), kernel_sz_values_precommit), - framework::dataset::make("depth_multiplier", 1)), + make("depth_multiplier", 1)), dilation_values), stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F32)), + make("DataType", DataType::F32)), data_layout_values), act_values), - n0_values_precommit)) + n0_values_precommit), + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( width_values_nightly, height_values_nightly), channel_values_nightly), batch_values_nightly), kernel_sz_values_nightly), - framework::dataset::make("depth_multiplier", 1)), + make("depth_multiplier", 1)), dilation_values), stride_values), - padding_valid_values), - framework::dataset::make("DataType", DataType::F32)), + padding_valid_values_nightly), + make("DataType", DataType::F32)), data_layout_values), - act_values), - n0_values_nightly)) + make("Activation", { ActivationLayerInfo() })), + n0_values_nightly), + make("ExportToCLImage", false))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::ALL, + combine( + make("width", { 33U } ), + height_values_precommit, + channel_values_precommit, + make("batch", { 2U } ), + make("kernel_size", { Size2D(5U, 5U) }), + make("depth_multiplier", 1), + make("dilation", Size2D(3U, 3U)), + make("stride", Size2D(3U, 2U)), + padding_valid_values_nightly, + make("DataType", DataType::F32), + data_layout_values, + activations_rest, + n0_values_precommit, + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); } + +TEST_SUITE(ExportWeightsToCLImage) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_precommit, + height_values_precommit), + channel_values_export_to_cl_image_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + make("depth_multiplier", 1)), + dilation_values), + stride_values), + padding_valid_values), + make("DataType", DataType::F32)), + data_layout_values), + act_values), + n0_values_export_to_cl_image_precommit), + make("ExportToCLImage", true))) +{ + // Validate output + if(_validate_output) + { + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_nightly, + height_values_nightly), + channel_values_export_to_cl_image_nightly), + batch_values_nightly), + kernel_sz_values_nightly), + make("depth_multiplier", 1)), + dilation_values), + stride_values), + padding_valid_values_nightly), + make("DataType", DataType::F32)), + data_layout_values), + make("Activation", { ActivationLayerInfo() })), + n0_values_export_to_cl_image_nightly), + make("ExportToCLImage", true))) +{ + // Validate output + if(_validate_output) + { + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +TEST_SUITE_END() // ExportWeightsToCLImage TEST_SUITE_END() // FP32 TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( width_values_precommit, height_values_precommit), channel_values_precommit), batch_values_precommit), kernel_sz_values_precommit), - framework::dataset::make("depth_multiplier", 1)), + make("depth_multiplier", 1)), dilation_values), stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F16)), + make("DataType", DataType::F16)), data_layout_values), act_values), - n0_values_precommit)) + n0_values_precommit), + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16); } - FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - width_values_nightly, - height_values_nightly), - channel_values_nightly), + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 47U } ), + make("height", { 39U } )), + make("channels", { 19U } )), batch_values_nightly), - kernel_sz_values_nightly), - framework::dataset::make("depth_multiplier", 1)), + make("kernel_size", { Size2D(5U, 5U) })), + make("depth_multiplier", 1)), + make("dilation", { Size2D(3U, 3U) })), + make("stride", { Size2D(3U, 2U) })), + padding_valid_values_nightly), + make("DataType", DataType::F16)), + data_layout_values), + make("Activation", { ActivationLayerInfo() })), + n0_values_nightly), + make("ExportToCLImage", false))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::ALL, + combine( + make("width", { 33U } ), + height_values_precommit, + channel_values_precommit, + make("batch", { 2U } ), + make("kernel_size", { Size2D(5U, 5U) }), + make("depth_multiplier", 4), + make("dilation", Size2D(3U, 3U)), + make("stride", Size2D(3U, 2U)), + padding_valid_values_nightly, + make("DataType", DataType::F16), + data_layout_values, + activations_rest, + n0_values_precommit, + make("ExportToCLImage", false))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); +} + +TEST_SUITE(ExportWeightsToCLImage) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + width_values_precommit, + height_values_precommit), + channel_values_export_to_cl_image_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + make("depth_multiplier", 1)), dilation_values), stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F16)), + make("DataType", DataType::F16)), data_layout_values), act_values), - n0_values_nightly)) + n0_values_export_to_cl_image_precommit), + make("ExportToCLImage", true))) { - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + // Validate output + if(_validate_output) + { + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 47U } ), + make("height", { 39U } )), + channel_values_export_to_cl_image_nightly), + batch_values_nightly), + make("kernel_size", { Size2D(5U, 5U) })), + make("depth_multiplier", 1)), + make("dilation", { Size2D(3U, 3U) })), + make("stride", { Size2D(3U, 2U) })), + padding_valid_values_nightly), + make("DataType", DataType::F16)), + data_layout_values), + make("Activation", { ActivationLayerInfo() })), + n0_values_export_to_cl_image_nightly), + make("ExportToCLImage", true))) +{ + // Validate output + if(_validate_output) + { + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } +TEST_SUITE_END() // ExportWeightsToCLImage TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float + TEST_SUITE(DepthMultiplier) TEST_SUITE(Float) TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - width_values_precommit, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 33U } ), height_values_precommit), channel_values_precommit), batch_values_precommit), @@ -225,18 +421,19 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<fl dilation_values), stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F32)), + make("DataType", DataType::F32)), data_layout_values), act_values), - framework::dataset::make("N0", 1))) + make("N0", 1)), + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - width_values_nightly, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 53U } ), height_values_nightly), channel_values_nightly), batch_values_nightly), @@ -244,21 +441,77 @@ FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<fl depth_multiplier_values), dilation_values), stride_values), + padding_valid_values_nightly), + make("DataType", DataType::F32)), + data_layout_values), + make("Activation", { ActivationLayerInfo() })), + make("N0", 1)), + make("ExportToCLImage", false))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); +} + +TEST_SUITE(DepthMultiplierMultipleOfOutputChannels) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 33U } ), + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + make("depth_multiplier", 2)), + dilation_values), + stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F32)), + make("DataType", DataType::F32)), data_layout_values), act_values), - framework::dataset::make("N0", 1))) + make("N0", {2})), + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); } + +TEST_SUITE(ExportWeightsToCLImage) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 33U } ), + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + make("depth_multiplier", 4)), + dilation_values), + stride_values), + padding_valid_values), + make("DataType", DataType::F32)), + data_layout_values), + act_values), + make("N0", {4})), + make("ExportToCLImage", true))) +{ + // Validate output + if(_validate_output) + { + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +TEST_SUITE_END() // ExportWeightsToCLImage +TEST_SUITE_END() // DepthMultiplierMultipleOfOutputChannels TEST_SUITE_END() // FP32 TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - width_values_precommit, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 33U } ), height_values_precommit), channel_values_precommit), batch_values_precommit), @@ -267,18 +520,19 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<ha dilation_values), stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F16)), + make("DataType", DataType::F16)), data_layout_values), act_values), - framework::dataset::make("N0", 1))) + make("N0", 1)), + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - width_values_nightly, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 53U } ), height_values_nightly), channel_values_nightly), batch_values_nightly), @@ -286,15 +540,71 @@ FIXTURE_DATA_TEST_CASE_NEW(RunLarge, CLDepthwiseConvolutionLayerNativeFixture<ha depth_multiplier_values), dilation_values), stride_values), + padding_valid_values_nightly), + make("DataType", DataType::F16)), + data_layout_values), + make("Activation", { ActivationLayerInfo() })), + make("N0", 1)), + make("ExportToCLImage", false))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); +} + +TEST_SUITE(DepthMultiplierMultipleOfOutputChannels) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 33U } ), + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + make("depth_multiplier", 2)), + dilation_values), + stride_values), padding_valid_values), - framework::dataset::make("DataType", DataType::F16)), + make("DataType", DataType::F16)), data_layout_values), act_values), - framework::dataset::make("N0", 1))) + make("N0", {2})), + make("ExportToCLImage", false))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); } + +TEST_SUITE(ExportWeightsToCLImage) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CLDepthwiseConvolutionLayerNativeFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + make("width", { 33U } ), + height_values_precommit), + channel_values_precommit), + batch_values_precommit), + kernel_sz_values_precommit), + make("depth_multiplier", 4)), + dilation_values), + stride_values), + padding_valid_values), + make("DataType", DataType::F16)), + data_layout_values), + act_values), + make("N0", {4})), + make("ExportToCLImage", true))) +{ + // Validate output + if(_validate_output) + { + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +TEST_SUITE_END() // ExportWeightsToCLImage +TEST_SUITE_END() // DepthMultiplierMultipleOfOutputChannels TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float TEST_SUITE_END() // DepthMultiplier diff --git a/tests/validation/CL/DilatedConvolutionLayer.cpp b/tests/validation/CL/DilatedConvolutionLayer.cpp index 9a9df2c7e4..776bf34151 100644 --- a/tests/validation/CL/DilatedConvolutionLayer.cpp +++ b/tests/validation/CL/DilatedConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -167,13 +167,18 @@ template <typename T> using CLGEMMDilatedConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(Quantized) +/// @note: Every asymmetric quantized test where there's no fused activation will have its quantization info ignored +/// This is because instead of using the same quantization information for all the tensors, the fixture generates +/// separate quantization info for each input and the output tensor. +/// When we can also support dynamic quantization with the presence of activation, we can remove the explicit +/// quantization info. TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMDilatedConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("IgnoredQuantizationInfo", { QuantizationInfo() })), framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() }))) { // Validate output @@ -185,7 +190,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMDilatedConvolutionLayerQuantizedFixture<u framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 0) })), + framework::dataset::make("IgnoredQuantizationInfo", { QuantizationInfo() })), framework::dataset::make("ActivationLayerInfo", { ActivationLayerInfo() }))) { // Validate output diff --git a/tests/validation/CL/DirectConvolutionLayer.cpp b/tests/validation/CL/DirectConvolutionLayer.cpp index a057f48c87..ff22ae5ef0 100644 --- a/tests/validation/CL/DirectConvolutionLayer.cpp +++ b/tests/validation/CL/DirectConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -35,6 +35,9 @@ #include "tests/validation/Validation.h" #include "tests/validation/fixtures/DirectConvolutionLayerFixture.h" +/** Synced with tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp + * Please check there for any differences in the coverage + */ namespace arm_compute { namespace test @@ -43,10 +46,12 @@ namespace validation { namespace { -RelativeTolerance<half> tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */ -RelativeTolerance<float> tolerance_fp32(0.05f); /**< Tolerance for floating point tests */ -constexpr float tolerance_num = 0.07f; /**< Tolerance number */ -constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance for quantized tests */ +RelativeTolerance<half> tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */ +RelativeTolerance<float> tolerance_fp32(0.05f); /**< Tolerance for floating point tests */ +constexpr float abs_tolerance_f32(0.0001f); /**< Absolute tolerance for FP32 tests*/ + +constexpr float tolerance_num = 0.07f; /**< Tolerance number */ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance for quantized tests */ const auto data_strides = combine(framework::dataset::make("StrideX", 1, 3), framework::dataset::make("StrideY", 1, 3)); const auto data_strides_small = combine(framework::dataset::make("StrideX", 1), framework::dataset::make("StrideY", 1)); @@ -130,55 +135,89 @@ TEST_CASE(NoBias, framework::DatasetMode::PRECOMMIT) validate(CLAccessor(dst), ref_dst); } +/** Check whether the case of rectangle kernels i.e. when width and height of the weight_shape are not equal + * would lead to successful run + */ +TEST_CASE(NonSquareKernel, framework::DatasetMode::PRECOMMIT) +{ + auto src_shape = TensorShape(33U, 27U, 3U); + auto weights_shape = TensorShape(5U, 7U, 3U, 4U); // non-square kernel + const auto bias_shape = TensorShape(4U); + auto dst_shape = TensorShape(11U, 12U, 4U); + constexpr auto dt = DataType::F32; + + TensorShape src_shape_nhwc(src_shape); + TensorShape weights_shape_nhwc(weights_shape); + TensorShape dst_shape_nhwc(dst_shape); + + // Non-square shapes are only allowed for NHWC + permute(src_shape_nhwc, PermutationVector(2U, 0U, 1U)); + permute(weights_shape_nhwc, PermutationVector(2U, 0U, 1U)); + permute(dst_shape_nhwc, PermutationVector(2U, 0U, 1U)); + + auto src = create_tensor<CLTensor>(src_shape_nhwc, dt, 1, QuantizationInfo(), DataLayout::NHWC); + auto weights = create_tensor<CLTensor>(weights_shape_nhwc, dt, 1, QuantizationInfo(), DataLayout::NHWC); + auto dst = create_tensor<CLTensor>(dst_shape_nhwc, dt, 1, QuantizationInfo(), DataLayout::NHWC); + const auto conv_info = PadStrideInfo(3, 2, 1, 1, 2, 0, DimensionRoundingType::FLOOR); + + // Create direct convolution function + CLDirectConvolutionLayer conv{}; + conv.configure(&src, &weights, nullptr, &dst, conv_info); + + src.allocator()->allocate(); + weights.allocator()->allocate(); + dst.allocator()->allocate(); + + library->fill_tensor_value(CLAccessor(src), 1.f); + library->fill_tensor_value(CLAccessor(weights), 1.f); + + conv.run(); + + // Compute reference to compare + SimpleTensor<float> ref_src{ src_shape, dt }; + SimpleTensor<float> ref_weights{ weights_shape, dt }; + SimpleTensor<float> ref_bias{ bias_shape, dt }; + library->fill_tensor_value(ref_src, 1.f); + library->fill_tensor_value(ref_weights, 1.f); + // No bias + library->fill_tensor_value(ref_bias, 0.f); + auto ref_dst = reference::convolution_layer<float>(ref_src, ref_weights, ref_bias, dst_shape, conv_info); + + validate(CLAccessor(dst), ref_dst); +} // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching input feature maps - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported kernel width - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Non-rectangular weights dimensions + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid: Mismatching data type input/weights + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid: Mismatching input feature maps TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid weights dimensions - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid stride - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases dimensions + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported biases size + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported biases dimensions TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid output size - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Window shrink TensorInfo(TensorShape(32U, 16U, 2U), 1, DataType::F32), }), framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F16), TensorInfo(TensorShape(3U, 3U, 3U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(11U, 11U, 2U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(5U, 3U, 2U, 4U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 4U, 3U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), TensorInfo(TensorShape(1U, 1U, 2U, 4U), 1, DataType::F32), })), framework::dataset::make("BiasesInfo",{ TensorInfo(TensorShape(4U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), - TensorInfo(TensorShape(4U), 1, DataType::F32), - TensorInfo(TensorShape(4U), 1, DataType::F32), - TensorInfo(TensorShape(4U), 1, DataType::F32), TensorInfo(TensorShape(3U), 1, DataType::F32), TensorInfo(TensorShape(4U, 2U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), - TensorInfo(TensorShape(4U), 1, DataType::F32), })), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), TensorInfo(TensorShape(26U, 11U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32), TensorInfo(TensorShape(32U, 16U, 4U), 1, DataType::F32), })), framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), @@ -186,23 +225,27 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(3, 3, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), - PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), })), framework::dataset::make("ActivationInfo", { + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) })), - framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, false, true })), + framework::dataset::make("Expected", { false, false, false, false, false, false, true })), input_info, weights_info, biases_info, output_info, conv_info, act_info, expected) { bool is_valid = bool(CLDirectConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, act_info)); ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); } +// clang-format on +// *INDENT-ON* template <typename T> using CLDirectConvolutionLayerFixture = DirectConvolutionValidationFixture<CLTensor, CLAccessor, CLDirectConvolutionLayer, T>; @@ -218,6 +261,46 @@ template <typename T> using CLDirectConvolutionValidationWithTensorShapesQuantizedFixture = DirectConvolutionValidationWithTensorShapesQuantizedFixture<CLTensor, CLAccessor, CLDirectConvolutionLayer, T>; TEST_SUITE(NHWC) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Arbitrary weight sizes for NHWC are supported + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Non-rectangular weights dimensions for NHWC are supported + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Strides > 2 for any kernel sizes for NHWC are supported + }), + framework::dataset::make("WeightsInfo",{ + TensorInfo(TensorShape(2U, 13U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 5U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("BiasesInfo",{ + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("OutputInfo",{ + TensorInfo(TensorShape(4U, 15U, 1U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U, 23U, 11U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U, 9U, 4U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("ConvInfo", { + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(3, 3, 0, 0), + })), + framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), +})), + framework::dataset::make("Expected", { true, true, true })), + input_info, weights_info, biases_info, output_info, conv_info, act_info, expected) +{ + bool is_valid = bool(CLDirectConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, act_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(zip(zip(zip(zip(zip(zip( @@ -273,7 +356,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerFixture<float>, framewo framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )), framework::dataset::make("DataLayout", DataLayout::NHWC))) { - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(zip(zip(zip(zip(zip(zip( @@ -291,7 +374,7 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerMixedDataLayo framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )), framework::dataset::make("DataLayout", DataLayout::NHWC))) { - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(zip(zip(zip(zip(zip(zip( @@ -306,7 +389,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerFixture<float>, framewo framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::IDENTITY) )), framework::dataset::make("DataLayout", DataLayout::NHWC))) { - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } TEST_SUITE_END() // FP32 @@ -427,9 +510,48 @@ TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE_END() // Quantized TEST_SUITE_END() // NHWC +TEST_SUITE(NCHW) +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, DataLayout::NCHW), // Unsupported kernel width + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, DataLayout::NCHW), // Non-rectangular weights dimensions are unsupported + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, DataLayout::NCHW) // Unsupported stride + }), + framework::dataset::make("WeightsInfo",{ + TensorInfo(TensorShape(11U, 11U, 2U, 4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(5U, 3U, 2U, 4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32, DataLayout::NCHW) + })), + framework::dataset::make("BiasesInfo",{ + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NCHW) + })), + framework::dataset::make("OutputInfo",{ + TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(23U, 11U, 4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(25U, 11U, 4U), 1, DataType::F32, DataLayout::NCHW) + })), + framework::dataset::make("ConvInfo", { + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(3, 3, 0, 0) + })), + framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +})), + framework::dataset::make("Expected", { false, false, false})), + input_info, weights_info, biases_info, output_info, conv_info, act_info, expected) +{ + bool is_valid = bool(CLDirectConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, act_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} // clang-format on // *INDENT-ON* -TEST_SUITE(NCHW) + TEST_SUITE(Float) TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit, framework::dataset::make("DataType", DataType::F16)), @@ -454,20 +576,21 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerFixture<float>, framewo ActivationFunctionsDataset), framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit, framework::dataset::make("DataType", - DataType::F32)), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit, + framework::dataset::make("DataType", + DataType::F32)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data_nightly, framework::dataset::make("DataType", DataType::F32)), ActivationFunctionsDataset), framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } TEST_SUITE_END() // FP32 @@ -477,107 +600,202 @@ FIXTURE_DATA_TEST_CASE(Run, CLDirectConvolutionValidationWithTensorShapesFixture ActivationFunctionsDataset)) { // Validate output - validate(CLAccessor(_target), _reference, tolerance_fp32); + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); } TEST_SUITE_END() // FP32_CustomDataset TEST_SUITE_END() // Float +/// @note: Every quantized test has a version with or without activation because the quantization info given is +/// ignored when there is no activation. Instead of using the same quantization information for all the tensors, the +/// fixture generates separate quantization info for each input and the output tensor. +/// When we can also support dynamic quantization with the presence of activation, these two versions should be merged +/// again, with the explicitly specified quantization info removed const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationInfo", { - ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) }); +const auto NoActivation = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo() +}); +const auto IgnoredQuantizationInfo = framework::dataset::make("IgnoredQuantizationInfo", +{ + QuantizationInfo() +}); TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(data_precommit, - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10)})), - QuantizedActivationFunctionsDataset), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(data_precommit, + framework::dataset::make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + NoActivation, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(data_precommit, - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, CLDirectConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(data_precommit, + framework::dataset::make("DataType", DataType::QASYMM8), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10) }), + QuantizedActivationFunctionsDataset, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(data_precommit_9x9, +FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(data_precommit, + framework::dataset::make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + NoActivation, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(data_precommit, + framework::dataset::make("DataType", DataType::QASYMM8), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(data_precommit_9x9, + framework::dataset::make("DataType", + DataType::QASYMM8), + IgnoredQuantizationInfo, + NoActivation, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmall9x9WithActivation, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(data_precommit_9x9, framework::dataset::make("DataType", - DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(3.f / 255, 10), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), + DataType::QASYMM8), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(3.f / 255, 10), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(data_nightly, framework::dataset::make("DataType", - DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), +FIXTURE_DATA_TEST_CASE(RunLarge, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(data_nightly, framework::dataset::make("DataType", + DataType::QASYMM8), + IgnoredQuantizationInfo, + NoActivation, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge9x9, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(data_nightly_9x9, +FIXTURE_DATA_TEST_CASE(RunLargeWithActivation, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(data_nightly, framework::dataset::make("DataType", + DataType::QASYMM8), + framework::dataset::make("QuantizationInfoIf", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunLarge9x9, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(data_nightly_9x9, framework::dataset::make("DataType", - DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(3.f / 255, 10), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), + DataType::QASYMM8), + IgnoredQuantizationInfo, + NoActivation, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } - -TEST_SUITE_END() // QASYMM8 - -TEST_SUITE(QASYMM8_CustomDataset) -FIXTURE_DATA_TEST_CASE(Run, CLDirectConvolutionValidationWithTensorShapesQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::DirectConvolutionLayerDataset(), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 127), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), +FIXTURE_DATA_TEST_CASE(RunLarge9x9WithActivation, CLDirectConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(data_nightly_9x9, + framework::dataset::make("DataType", + DataType::QASYMM8), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(3.f / 255, 10), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(CustomDataset, CLDirectConvolutionValidationWithTensorShapesQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(datasets::DirectConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + NoActivation, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(CustomDatasetWithActivation, CLDirectConvolutionValidationWithTensorShapesQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(datasets::DirectConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 127), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -TEST_SUITE_END() // QASYMM8_CustomDataset +TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(data_precommit, framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, -10) })), - QuantizedActivationFunctionsDataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLDirectConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(data_precommit, framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + NoActivation, framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(data_precommit, framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.1f, -10) })), - QuantizedActivationFunctionsDataset), +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, CLDirectConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(data_precommit, framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, -10) }), + QuantizedActivationFunctionsDataset, framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(data_precommit_9x9, +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLDirectConvolutionLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::ALL, combine(data_precommit, framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), + DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + NoActivation, + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, CLDirectConvolutionLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::ALL, combine(data_precommit, + framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.1f, -10) }), + QuantizedActivationFunctionsDataset, + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(data_precommit_9x9, + framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + NoActivation, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmall9x9WithActivation, CLDirectConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(data_precommit_9x9, + framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 10), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output @@ -585,10 +803,21 @@ FIXTURE_DATA_TEST_CASE(RunSmall9x9, CLDirectConvolutionLayerQuantizedFixture<int } FIXTURE_DATA_TEST_CASE(RunCustomDataset, CLDirectConvolutionValidationWithTensorShapesQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::DirectConvolutionLayerDataset(), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 127), QuantizationInfo(1.1f, 10) })), - QuantizedActivationFunctionsDataset), + combine(datasets::DirectConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + NoActivation, + framework::dataset::make("DataLayout", { DataLayout::NCHW }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + +FIXTURE_DATA_TEST_CASE(RunCustomDatasetWithActivation, CLDirectConvolutionValidationWithTensorShapesQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine(datasets::DirectConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255, 127), QuantizationInfo(1.1f, 10) }), + QuantizedActivationFunctionsDataset, framework::dataset::make("DataLayout", { DataLayout::NCHW }))) { // Validate output diff --git a/tests/validation/CL/ElementwiseMax.cpp b/tests/validation/CL/ElementwiseMax.cpp index b9444b2795..bd47c23256 100644 --- a/tests/validation/CL/ElementwiseMax.cpp +++ b/tests/validation/CL/ElementwiseMax.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -58,7 +58,7 @@ const auto ElementwiseMaxQASYMM8SignedDataset = combine(combine(framework::datas const auto ElementwiseMaxQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)), framework::dataset::make("DataType", DataType::QSYMM16)); -const auto ElementwiseMaxS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), +const auto ElementwiseMaxS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), framework::dataset::make("DataType", DataType::S16)); const auto ElementwiseMaxFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataType", DataType::F16)); @@ -71,6 +71,8 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(CL) @@ -80,21 +82,18 @@ TEST_SUITE(ElementwiseMax) // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes }), framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false})), + framework::dataset::make("Expected", { true, false, false})), input1_info, input2_info, output_info, expected) { ARM_COMPUTE_EXPECT(bool(CLElementwiseMax::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS); @@ -107,7 +106,8 @@ using CLElementwiseMaxFixture = ElementwiseMaxValidationFixture<CLTensor, CLAcce TEST_SUITE(Integer) TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), ElementwiseMaxU8Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ElementwiseMaxU8Dataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -115,7 +115,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFixture<uint8_t>, framework::Da TEST_SUITE_END() TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMaxS16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMaxS16Dataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -128,33 +129,36 @@ using CLElementwiseMaxQuantizedFixture = ElementwiseMaxValidationQuantizedFixtur TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxQASYMM8Dataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32, 0.01); } TEST_SUITE_END() TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxQASYMM8SignedDataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE_END() TEST_SUITE(QSYMM16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxQSYMM16Dataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -167,13 +171,16 @@ using CLElementwiseMaxFloatFixture = ElementwiseMaxValidationFloatFixture<CLTens TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMaxFP16Dataset), EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxFP16Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMaxFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwiseMaxFP16Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMaxFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwiseMaxFP16Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); @@ -181,14 +188,16 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMaxFloatFixture<half>, fr TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMaxFP32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMaxFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxFP32Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMaxFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwiseMaxFP32Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMaxFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwiseMaxFP32Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); @@ -197,16 +206,18 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMaxFloatFixture<float>, f template <typename T> using CLElementwiseMaxBroadcastFloatFixture = ElementwiseMaxBroadcastValidationFloatFixture<CLTensor, CLAccessor, CLElementwiseMax, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwiseMaxBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwiseMaxBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapesBroadcast(), ElementwiseMaxFP32Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwiseMaxBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwiseMaxBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), ElementwiseMaxFP32Dataset), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); diff --git a/tests/validation/CL/ElementwiseMin.cpp b/tests/validation/CL/ElementwiseMin.cpp index 8f53b241ab..ee229a0941 100644 --- a/tests/validation/CL/ElementwiseMin.cpp +++ b/tests/validation/CL/ElementwiseMin.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -58,7 +58,7 @@ const auto ElementwiseMinQASYMM8SignedDataset = combine(combine(framework::datas const auto ElementwiseMinQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)), framework::dataset::make("DataType", DataType::QSYMM16)); -const auto ElementwiseMinS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), +const auto ElementwiseMinS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), framework::dataset::make("DataType", DataType::S16)); const auto ElementwiseMinFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataType", DataType::F16)); @@ -71,6 +71,8 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(CL) @@ -80,21 +82,18 @@ TEST_SUITE(ElementwiseMin) // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes }), framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false})), + framework::dataset::make("Expected", { true, false, false})), input1_info, input2_info, output_info, expected) { ARM_COMPUTE_EXPECT(bool(CLElementwiseMin::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS); @@ -107,7 +106,8 @@ using CLElementwiseMinFixture = ElementwiseMinValidationFixture<CLTensor, CLAcce TEST_SUITE(Integer) TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), ElementwiseMinU8Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ElementwiseMinU8Dataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -115,7 +115,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFixture<uint8_t>, framework::Da TEST_SUITE_END() TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMinS16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMinS16Dataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -128,33 +129,36 @@ using CLElementwiseMinQuantizedFixture = ElementwiseMinValidationQuantizedFixtur TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMinQASYMM8Dataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32, 0.01); } TEST_SUITE_END() TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMinQASYMM8SignedDataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE_END() TEST_SUITE(QSYMM16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMinQSYMM16Dataset), framework::dataset::make("SrcQInfo0", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), framework::dataset::make("SrcQInfo1", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -167,13 +171,16 @@ using CLElementwiseMinFloatFixture = ElementwiseMinValidationFloatFixture<CLTens TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMinFP16Dataset), EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwiseMinFP16Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMinFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwiseMinFP16Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMinFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwiseMinFP16Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); @@ -181,14 +188,16 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMinFloatFixture<half>, fr TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMinFP32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseMinFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwiseMinFP32Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMinFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwiseMinFP32Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMinFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwiseMinFP32Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); @@ -196,16 +205,18 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseMinFloatFixture<float>, f template <typename T> using CLElementwiseMinBroadcastFloatFixture = ElementwiseMinBroadcastValidationFloatFixture<CLTensor, CLAccessor, CLElementwiseMin, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwiseMinBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwiseMinBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapesBroadcast(), ElementwiseMinFP32Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwiseMinBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwiseMinBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), ElementwiseMinFP32Dataset), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); diff --git a/tests/validation/CL/ElementwisePower.cpp b/tests/validation/CL/ElementwisePower.cpp index a2d3ba6c09..c2aeb6e045 100644 --- a/tests/validation/CL/ElementwisePower.cpp +++ b/tests/validation/CL/ElementwisePower.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -57,6 +57,8 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(CL) @@ -96,29 +98,33 @@ using CLElementwisePowerBroadcastFloatFixture = ElementwisePowerBroadcastValidat TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwisePowerFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwisePowerFP16Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwisePowerFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwisePowerFP16Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwisePowerFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwisePowerFP16Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwisePowerFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwisePowerFP16Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwisePowerBroadcastFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwisePowerBroadcastFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapesBroadcast(), ElementwisePowerFP16Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwisePowerBroadcastFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwisePowerBroadcastFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), ElementwisePowerFP16Dataset), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); @@ -126,29 +132,33 @@ FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwisePowerBroadcastFl TEST_SUITE_END() //FP16 TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwisePowerFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwisePowerFP32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwisePowerFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwisePowerFP32Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwisePowerFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwisePowerFP32Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwisePowerFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwisePowerFP32Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwisePowerBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwisePowerBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapesBroadcast(), ElementwisePowerFP32Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwisePowerBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwisePowerBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), ElementwisePowerFP32Dataset), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); diff --git a/tests/validation/CL/ElementwiseSquaredDiff.cpp b/tests/validation/CL/ElementwiseSquaredDiff.cpp index 0a4ab6627b..ee0279df33 100644 --- a/tests/validation/CL/ElementwiseSquaredDiff.cpp +++ b/tests/validation/CL/ElementwiseSquaredDiff.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -57,7 +57,7 @@ const auto ElementwiseSquaredDiffQASYMM8Dataset = combine(combine(framework::dat const auto ElementwiseSquaredDiffQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)), framework::dataset::make("DataType", DataType::QSYMM16)); -const auto ElementwiseSquaredDiffS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), +const auto ElementwiseSquaredDiffS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), framework::dataset::make("DataType", DataType::S16)); const auto ElementwiseSquaredDiffFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataType", DataType::F16)); @@ -70,6 +70,8 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(CL) @@ -79,21 +81,18 @@ TEST_SUITE(ElementwiseSquaredDiff) // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes }), framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false})), + framework::dataset::make("Expected", { true, false, false})), input1_info, input2_info, output_info, expected) { ARM_COMPUTE_EXPECT(bool(CLElementwiseSquaredDiff::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS); @@ -106,7 +105,8 @@ using CLElementwiseSquaredDiffFixture = ElementwiseSquaredDiffValidationFixture< TEST_SUITE(Integer) TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), ElementwiseSquaredDiffU8Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffU8Dataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -114,7 +114,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFixture<uint8_t>, frame TEST_SUITE_END() TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseSquaredDiffS16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffS16Dataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference); @@ -127,22 +128,24 @@ using CLElementwiseSquaredDiffQuantizedFixture = ElementwiseSquaredDiffValidatio TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffQASYMM8Dataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32, 0.01); } TEST_SUITE_END() TEST_SUITE(QSYMM16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffQSYMM16Dataset), framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) })), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qsymm16); @@ -155,14 +158,16 @@ using CLElementwiseSquaredDiffFloatFixture = ElementwiseSquaredDiffValidationFlo TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP16Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP16Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseSquaredDiffFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwiseSquaredDiffFP16Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseSquaredDiffFloatFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwiseSquaredDiffFP16Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp16, 0.01); @@ -170,14 +175,16 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseSquaredDiffFloatFixture<h TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP32Dataset), - EmptyActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLElementwiseSquaredDiffFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP32Dataset), + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseSquaredDiffFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), ElementwiseSquaredDiffFP32Dataset), - ActivationFunctionsDataset)) +FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseSquaredDiffFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapes(), ElementwiseSquaredDiffFP32Dataset), + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); @@ -185,16 +192,18 @@ FIXTURE_DATA_TEST_CASE(RunWithActivation, CLElementwiseSquaredDiffFloatFixture<f template <typename T> using CLElementwiseSquaredDiffBroadcastFloatFixture = ElementwiseSquaredDiffBroadcastValidationFloatFixture<CLTensor, CLAccessor, CLElementwiseSquaredDiff, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwiseSquaredDiffBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLElementwiseSquaredDiffBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapesBroadcast(), ElementwiseSquaredDiffFP32Dataset), - EmptyActivationFunctionsDataset)) + EmptyActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); } -FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwiseSquaredDiffBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcast(), +FIXTURE_DATA_TEST_CASE(RunWithActivationBroadcast, CLElementwiseSquaredDiffBroadcastFloatFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::TinyShapesBroadcast(), ElementwiseSquaredDiffFP32Dataset), - ActivationFunctionsDataset)) + ActivationFunctionsDataset), + OutOfPlaceDataSet)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fp32); diff --git a/tests/validation/CL/FullyConnectedLayer.cpp b/tests/validation/CL/FullyConnectedLayer.cpp index 09da519c51..2f0c86499b 100644 --- a/tests/validation/CL/FullyConnectedLayer.cpp +++ b/tests/validation/CL/FullyConnectedLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -40,6 +40,7 @@ namespace test { namespace validation { +using framework::dataset::make; namespace { /** Tolerance for float operations */ @@ -51,15 +52,20 @@ constexpr float tolerance_num = 0.07f; /**< Tolerance n /** Tolerance for quantized asymmetric operations */ constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); -const auto FullyConnectedParameters = combine(framework::dataset::make("TransposeWeights", { false, true }), framework::dataset::make("ReshapeWeights", { false, true })); +const auto FullyConnectedParameters = combine(make("TransposeWeights", { false, true }), make("ReshapeWeights", { false, true })); -const auto QuantizationData = framework::dataset::make("QuantizationInfo", +const auto QuantizationData = make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 10), QuantizationInfo(1.1f, 10), }); -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +const auto IgnoredQuantizationData = make("IgnoredQuantizationInfo", +{ + QuantizationInfo(), +}); + +const auto ActivationFunctionsDataset = make("ActivationInfo", { ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), @@ -68,11 +74,16 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH) }); -const auto ActivationFunctionsQuantizedDataset = concat(concat(concat( - framework::dataset::make("ActivationInfo", ActivationLayerInfo()), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f))), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.75f, 0.25f))); +// This dataset case only runs with dynamic quantization +const auto NoActivationFunctionsQuantizedDataset = make("ActivationInfo", +{ + ActivationLayerInfo() +}); + +const auto ActivationFunctionsQuantizedDataset = concat(concat( + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f))), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.75f, 0.25f))); } // namespace TEST_SUITE(CL) @@ -81,33 +92,33 @@ TEST_SUITE(FullyConnectedLayer) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Mismatching data types + make("InputInfo", { TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Mismatching data types TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Invalid weights dimensions TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Wrongly reshaped weights }), - framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(315U, 271U), 1, DataType::F16), + make("WeightsInfo",{ TensorInfo(TensorShape(315U, 271U), 1, DataType::F16), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), TensorInfo(TensorShape(217U, 231U), 1, DataType::F32), TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), })), - framework::dataset::make("BiasInfo",{ TensorInfo(TensorShape(271U), 1, DataType::F32), + make("BiasInfo",{ TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(192U), 1, DataType::F32), TensorInfo(TensorShape(192U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), + make("OutputInfo",{ TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), })), - framework::dataset::make("TransposeWeights",{ true, true, false, true, true })), - framework::dataset::make("ReshapedWeights",{ false, false, false, false, false})), - framework::dataset::make("Expected", { false, true, true, false, false })), + make("TransposeWeights",{ true, true, false, true, true })), + make("ReshapedWeights",{ false, false, false, false, false})), + make("Expected", { false, true, true, false, false })), input_info, weights_info, bias_info, output_info, transpose_weights, reshaped_weights, expected) { // Create Fully Connected layer info @@ -131,54 +142,69 @@ template <typename T> using CLFullyConnectedLayerMixedDataLayoutFixture = FullyConnectedLayerValidationFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T, true>; template <typename T> using CLFullyConnectedLayerDynamicWeightsFixture = FullyConnectedWithDynamicWeightsFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>; +template <typename T> +using CLFullyConnectedNoBiasFixture = FullyConnectedDynamicNoBiasFixture<CLTensor, CLAccessor, CLFullyConnectedLayer, T>; TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F16)), +FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + FullyConnectedParameters, + make("DataType", DataType::F16), ActivationFunctionsDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F16)), +FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeFullyConnectedLayerDataset(), + FullyConnectedParameters, + make("DataType", DataType::F16), ActivationFunctionsDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, CLFullyConnectedLayerDynamicWeightsFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::F16), + make("ActivationInfo", ActivationLayerInfo()), + make("WeightsReshaped", { false, true }))) +{ +} TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), +FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters, + make("DataType", DataType::F32), ActivationFunctionsDataset)) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0, abs_tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLFullyConnectedLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(9U, 5U, 7U)), - framework::dataset::make("Weights", TensorShape(315U, 271U))), - framework::dataset::make("Biases", TensorShape(271U))), - framework::dataset::make("Output", TensorShape(271U))), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLFullyConnectedLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::F32), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0, abs_tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunDynamicWeights, CLFullyConnectedLayerDynamicWeightsFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallFullyConnectedLayerDataset(), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, CLFullyConnectedLayerDynamicWeightsFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::F32), + make("ActivationInfo", ActivationLayerInfo()), + make("WeightsReshaped", { false, true }))) +{ +} +FIXTURE_DATA_TEST_CASE(RunDynamicNoBias, CLFullyConnectedNoBiasFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::F32), + make("ActivationInfo", { ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) }), + make("WeightsReshaped", { false }))) { } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), +FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters, + make("DataType", DataType::F32), ActivationFunctionsDataset)) { // Validate output @@ -194,57 +220,132 @@ using CLFullyConnectedLayerQuantizedMixedDataLayoutFixture = FullyConnectedLayer TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QASYMM8)), QuantizationData), +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, CLFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters, make("DataType", DataType::QASYMM8), QuantizationData, ActivationFunctionsQuantizedDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLFullyConnectedLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(9U, 5U, 7U)), - framework::dataset::make("Weights", TensorShape(315U, 271U))), - framework::dataset::make("Biases", TensorShape(271U))), - framework::dataset::make("Output", TensorShape(271U))), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8)), - QuantizationData), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, CLFullyConnectedLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8), + QuantizationData, + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QASYMM8)), QuantizationData), +FIXTURE_DATA_TEST_CASE(RunLargeWithActivation, CLFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters, make("DataType", DataType::QASYMM8), QuantizationData, ActivationFunctionsQuantizedDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } + +// Dynamic Quantization Tests +FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters, make("DataType", DataType::QASYMM8), IgnoredQuantizationData, + NoActivationFunctionsQuantizedDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters, make("DataType", DataType::QASYMM8), IgnoredQuantizationData, + NoActivationFunctionsQuantizedDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, CLFullyConnectedLayerDynamicWeightsFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8), + NoActivationFunctionsQuantizedDataset, + make("WeightsReshaped", { false /* COMPMID-6000: Support FullyConnected with quantized dynamic weights already reshaped */ }))) +{ +} + +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLFullyConnectedLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationData, + NoActivationFunctionsQuantizedDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + TEST_SUITE_END() /* QASYMM8 */ TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), QuantizationData), +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, CLFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters, make("DataType", DataType::QASYMM8_SIGNED), QuantizationData, ActivationFunctionsQuantizedDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, CLFullyConnectedLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8_SIGNED), + QuantizationData, + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + +// Dynamic Quantization tests below +FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters, make("DataType", DataType::QASYMM8_SIGNED), IgnoredQuantizationData, + NoActivationFunctionsQuantizedDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLFullyConnectedLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(9U, 5U, 7U)), - framework::dataset::make("Weights", TensorShape(315U, 271U))), - framework::dataset::make("Biases", TensorShape(271U))), - framework::dataset::make("Output", TensorShape(271U))), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - QuantizationData), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationData, + NoActivationFunctionsQuantizedDataset)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } + +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, CLFullyConnectedLayerDynamicWeightsFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8_SIGNED), + make("ActivationInfo", ActivationLayerInfo()), + make("WeightsReshaped", { false /* COMPMID-6000: Support FullyConnected with quantized dynamic weights already reshaped */ }))) +{ +} +FIXTURE_DATA_TEST_CASE(RunDynamicNoBias, CLFullyConnectedNoBiasFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8_SIGNED), + make("ActivationInfo", ActivationLayerInfo()), + make("WeightsReshaped", { false /* COMPMID-6000: Support FullyConnected with quantized dynamic weights already reshaped */ }))) +{ +} TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE_END() // Quantized TEST_SUITE_END() // FullyConnectedLayer diff --git a/tests/validation/CL/GEMM.cpp b/tests/validation/CL/GEMM.cpp index 838920c29d..16ca14f1d6 100644 --- a/tests/validation/CL/GEMM.cpp +++ b/tests/validation/CL/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -62,6 +62,29 @@ const auto CNNDataTypes = framework::dataset::make("DataType", TEST_SUITE(CL) TEST_SUITE(GEMM) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("LhsInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::S32), // Unsupported data type + TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), + }), + framework::dataset::make("RhsInfo",{ TensorInfo(TensorShape(8U, 27U), 1, DataType::S32), + TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), + })), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(8U, 13U), 1, DataType::S32), + TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), + })), + framework::dataset::make("Expected", { false, true })), + lhs_info, rhs_info, output_info, expected) +{ + constexpr float alpha = 1.0; + constexpr float beta = 0.0; + const auto gemm_info = GEMMInfo(); + bool is_valid = bool(CLGEMM::validate(&lhs_info.clone()->set_is_resizable(true), &rhs_info.clone()->set_is_resizable(true), nullptr, &output_info.clone()->set_is_resizable(true), alpha, beta, gemm_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* template <typename T> using CLGEMMFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T>; @@ -71,6 +94,9 @@ using CLGEMMOutput3DFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM template <typename T> using CLGEMMInputOutput3DFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T, false, true, true>; +template <typename T> +using CLBatchedMatMulFixture = GEMMValidationFixture<CLTensor, CLAccessor, CLGEMM, T, true, false, false, false, false, true>; + TEST_SUITE(Float) TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), @@ -181,10 +207,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMOutput3DFixture<half>, framework::Dataset validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } TEST_SUITE_END() // FP16 - TEST_SUITE_END() // Float TEST_SUITE_END() // OUTPUT_3D +TEST_SUITE(BATCHED_MATMUL) + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchedMatMulFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallBatchedMatMulDataset(), + framework::dataset::make("ReshapeWeights", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, tolerance_num); +} +TEST_SUITE_END() + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLBatchedMatMulFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallBatchedMatMulDataset(), + framework::dataset::make("ReshapeWeights", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} +TEST_SUITE_END() +TEST_SUITE_END() // BATCHED_MATMUL + TEST_SUITE_END() // GEMM TEST_SUITE_END() // CL } // namespace validation diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp index 5a1971b54c..78d794a9bb 100644 --- a/tests/validation/CL/GEMMLowp.cpp +++ b/tests/validation/CL/GEMMLowp.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,6 +44,9 @@ namespace test { namespace validation { + +using framework::dataset::make; + namespace { constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ @@ -53,6 +56,7 @@ TEST_SUITE(GEMMLowp) TEST_SUITE(MatrixMultiplyCore) using CLGEMMLowpMatrixMultiplyCoreFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore>; +using CLGEMMLowpBatchedMatMulFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, false, true>; FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset()) { @@ -66,18 +70,74 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFixture, framework: validate(CLAccessor(_target), _reference); } +using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedUnsigned = + GEMMLowpBatchedMatrixMultiplyCoreFusedOffsetOutputFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, false, uint8_t, uint8_t, true>; +TEST_SUITE(BatchedMatMul) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedUnsigned, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedBatchedMatMulDataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { false }))) +{ + validate(CLAccessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // QASYMM8 + +using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedSigned = + GEMMLowpBatchedMatrixMultiplyCoreFusedOffsetOutputFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, false, int8_t, int8_t, true>; +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedSigned, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedBatchedMatMulDataset(), + make("DataType", { DataType::QASYMM8_SIGNED }), + make("reshape_b_only_on_first_run", { false }))) +{ + validate(CLAccessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // BatchedMatMul + TEST_SUITE(FusedOffsetOutput) TEST_SUITE(QASYMM8) using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(), - framework::dataset::make("DataType", { DataType::QASYMM8 }))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_quant); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMLowpFusedOffsetOutputUint8Dataset(), - framework::dataset::make("DataType", { DataType::QASYMM8 }))) +TEST_SUITE(Output3D) +using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputOutput3DUint8Fixture = + GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, true>; +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputOutput3DUint8Fixture, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedOffsetOutputOutput3DUint8Dataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { true, false }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // Output3D + +TEST_SUITE(InputOutput3D) +using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInputOutput3DUint8Fixture = + GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, true, true>; +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInputOutput3DUint8Fixture, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedOffsetOutputInputOutput3DUint8Dataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { true, false }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // InputOutput3D + +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputUint8Fixture, framework::DatasetMode::NIGHTLY, + combine(datasets::LargeGEMMLowpFusedOffsetOutputUint8Dataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_quant); @@ -86,8 +146,10 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInt8Fixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyCore, false, false, int8_t, int8_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInt8Fixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpFusedOffsetOutputInt8Dataset(), - framework::dataset::make("DataType", { DataType::QASYMM8_SIGNED }))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputInt8Fixture, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedOffsetOutputInt8Dataset(), + make("DataType", { DataType::QASYMM8_SIGNED }), + make("reshape_b_only_on_first_run", { false }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_quant); @@ -130,24 +192,24 @@ TEST_SUITE(QuantizeDownInt32Scale) TEST_SUITE(QASYMM8) -const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2, - 3) - * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true }); +const auto quantize_down_int32_to_uint8_scale_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2) * make("result_shift", 2, 3) + * make("min", 0) * make("max", 255) * make("addBias", { false, true }); -const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, - 2) - * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 173) * framework::dataset::make("addBias", { false, true }); +const auto quantize_down_int32_to_uint8_scale_relu_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2) + * make("result_shift", 2, 3) * make("min", 0, 2) * make("max", 171, 173) * make("addBias", { false, true }); using CLGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, + combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, + combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases)) { // Validate output validate(CLAccessor(_target), _reference); @@ -158,24 +220,24 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -const auto quantize_down_int32_to_int8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2, - 3) - * framework::dataset::make("min", -128) * framework::dataset::make("max", 127) * framework::dataset::make("addBias", { false, true }); +const auto quantize_down_int32_to_int8_scale_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2) * make("result_shift", 2, 3) + * make("min", -128) * make("max", 127) * make("addBias", { false, true }); -const auto quantize_down_int32_to_int8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, - 2) - * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", -100, -98) * framework::dataset::make("max", 71, 73) * framework::dataset::make("addBias", { false, true }); +const auto quantize_down_int32_to_int8_scale_relu_cases = make("result_offset", -2, 1) * make("result_mult_int", 1, 2) + * make("result_shift", 2, 3) * make("min", -100, -98) * make("max", 71, 73) * make("addBias", { false, true }); using CLGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, + combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases)) { // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases)) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, + combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases)) { // Validate output validate(CLAccessor(_target), _reference); @@ -185,140 +247,6 @@ TEST_SUITE_END() // BoundedReLu TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE_END() // QuantizeDownInt32Scale -TEST_SUITE(QuantizeDownInt32ScaleByFixedPoint) - -TEST_SUITE(QASYMM8) - -const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true }); -using CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture<CLTensor, CLAccessor, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint>; - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} - -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} - -TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} - -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} -TEST_SUITE_END() // BoundedReLu -TEST_SUITE_END() // QASYMM8 -TEST_SUITE(QASYMM8_SIGNED) -const auto quantize_down_int32_to_int8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -128) * framework::dataset::make("max", 127) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -128, -126) * framework::dataset::make("max", 110, 112) * framework::dataset::make("addBias", { false, true }); -using CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture<CLTensor, CLAccessor, CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint>; - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int8_scale_by_fixedpoint_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} - -TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} - -TEST_SUITE_END() // BoundedReLu -TEST_SUITE_END() // QASYMM8_SIGNED -TEST_SUITE(QSYMM16) - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("min", -32768) * framework::dataset::make("max", 32767) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_cases = framework::dataset::make("result_fixedpoint_multiplier", 1073741823, - 1073741825) - * framework::dataset::make("result_shift", -3, - -2) - * framework::dataset::make("min", -32768) * framework::dataset::make("max", 32767) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, - 254601602) - * framework::dataset::make("result_shift", -3, - -1) - * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true }); - -using CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<CLTensor, CLAccessor, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>; - -TEST_SUITE(NoRelu) -TEST_SUITE(MultSmallerEq1) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} -TEST_SUITE_END() // MultSmallerEq1 -TEST_SUITE(MultGreater1) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} -TEST_SUITE_END() // MultGreater1 -TEST_SUITE_END() // NoRelu -TEST_SUITE(BoundedReLu) -TEST_SUITE(MultSmallerEq1) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} -TEST_SUITE_END() // MultSmallerEq1 -TEST_SUITE(MultGreater1) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_relu_cases)) -{ - // Validate output - validate(CLAccessor(_target), _reference); -} -TEST_SUITE_END() // MultGreater1 -TEST_SUITE_END() // BoundedReLu -TEST_SUITE_END() // QSYMM16 -TEST_SUITE_END() // QuantizeDownInt32ScaleByFixedPoint - TEST_SUITE(QuantizeDownInt32ScaleByFloat) TEST_SUITE(QASYMM8) @@ -326,13 +254,14 @@ using CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture = GEMMLowpQuantizeDownInt32ScaleByFloatValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage, uint8_t>; FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), - datasets::TinyShapes()), - framework::dataset::make("result_real_multiplier", 0.33f)), - framework::dataset::make("result_offset", 2, 3)), - framework::dataset::make("min", 0)), - framework::dataset::make("max", 255)), - framework::dataset::make("addBias", { false, true }))) + combine( + make("DataType", DataType::QASYMM8), + datasets::TinyShapes(), + make("result_real_multiplier", 0.33f), + make("result_offset", 2, 3), + make("min", 0), + make("max", 255), + make("addBias", { false, true }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -343,13 +272,14 @@ TEST_SUITE(QASYMM8_SIGNED) using CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture_Signed = GEMMLowpQuantizeDownInt32ScaleByFloatValidationFixture<CLTensor, CLAccessor, CLGEMMLowpOutputStage, int8_t>; FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMLowpQuantizeDownInt32ScaleByFloatFixture_Signed, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), - datasets::TinyShapes()), - framework::dataset::make("result_real_multiplier", 0.33f)), - framework::dataset::make("result_offset", 2, 3)), - framework::dataset::make("min", -128)), - framework::dataset::make("max", 127)), - framework::dataset::make("addBias", { false, true }))) + combine( + make("DataType", DataType::QASYMM8_SIGNED), + datasets::TinyShapes(), + make("result_real_multiplier", 0.33f), + make("result_offset", 2, 3), + make("min", -128), + make("max", 127), + make("addBias", { false, true }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -363,4 +293,4 @@ TEST_SUITE_END() // GEMMLowp TEST_SUITE_END() // CL } // namespace validation } // namespace test -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp index 1057af95f2..d0d06a8ddb 100644 --- a/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp +++ b/tests/validation/CL/GEMMLowpMatrixMultiplyNative.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,7 +23,7 @@ */ #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h" +#include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyNativeKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/framework/Asserts.h" @@ -41,7 +41,7 @@ namespace validation using namespace arm_compute::misc::shape_calculator; // Create function for CLGEMMMatrixMultiplyNativeKernel -using CLGEMMLowpMatrixMultiplyNative = CLSynthetizeFunction<CLGEMMLowpMatrixMultiplyNativeKernel>; +using CLGEMMLowpMatrixMultiplyNative = CLSynthetizeOperator<opencl::kernels::ClGemmLowpMatrixMultiplyNativeKernel>; // Fixture for CLGEMMLowpMatrixMultiplyNative using CLGEMMLowpMatrixMultiplyNativeFixture = GEMMLowpMatrixMultiplyNativeValidationFixture<CLTensor, CLAccessor, CLGEMMLowpMatrixMultiplyNative>; diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp index 68a7d055ad..88455bdeb8 100644 --- a/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp +++ b/tests/validation/CL/GEMMLowpMatrixMultiplyReshaped.cpp @@ -23,9 +23,9 @@ */ #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/framework/Asserts.h" @@ -49,7 +49,7 @@ using CLGEMMReshapeLHSMatrix = CLSynthetizeOperator<opencl::kernels::ClGemmResha using CLGEMMReshapeRHSMatrix = CLSynthetizeOperator<opencl::kernels::ClGemmReshapeRhsMatrixKernel>; // Create function for CLGEMMLowpMatrixMultiplyReshapedKernel -using CLGEMMLowpMatrixMultiplyReshaped = CLSynthetizeFunction<CLGEMMLowpMatrixMultiplyReshapedKernel>; +using CLGEMMLowpMatrixMultiplyReshaped = CLSynthetizeOperator<opencl::kernels::ClGemmLowpMatrixMultiplyReshapedKernel>; // Fixture for CLGEMMLowpMatrixMultiplyReshaped using CLGEMMLowpMatrixMultiplyReshapedFixture = GEMMLowpMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshaped>; diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp index 43b86b51e8..c56901effc 100644 --- a/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp +++ b/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRHS.cpp @@ -25,8 +25,8 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" @@ -49,7 +49,7 @@ using namespace arm_compute::misc::shape_calculator; using CLGEMMReshapeRHSMatrix = CLSynthetizeOperator<opencl::kernels::ClGemmReshapeRhsMatrixKernel>; // Create function for CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel -using CLGEMMLowpMatrixMultiplyReshapedOnlyRHS = CLSynthetizeFunction<CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel>; +using CLGEMMLowpMatrixMultiplyReshapedOnlyRHS = CLSynthetizeOperator<opencl::kernels::ClGemmLowpMatrixMultiplyReshapedOnlyRhsKernel>; // Fixture for CLGEMMLowpMatrixMultiplyReshapedOnlyRHS using CLGEMMLowpMatrixMultiplyReshapedOnlyRHSFixture = GEMMLowpMatrixMultiplyReshapedOnlyRHSValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshapedOnlyRHS>; @@ -157,7 +157,7 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned // Create and configure function CLGEMMLowpMatrixMultiplyReshapedOnlyRHS gemm; - gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info); + gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info); } } // namespace diff --git a/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRhsMMUL.cpp b/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRhsMMUL.cpp new file mode 100644 index 0000000000..a0d13c3e39 --- /dev/null +++ b/tests/validation/CL/GEMMLowpMatrixMultiplyReshapedOnlyRhsMMUL.cpp @@ -0,0 +1,206 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/CL/functions/CLCast.h" +#include "arm_compute/runtime/CL/functions/CLReductionOperation.h" +#include "src/gpu/cl/kernels/ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "tests/CL/CLAccessor.h" +#include "tests/CL/Helper.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/fixtures/GEMMLowpFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::opencl::kernels; + +// Create function for CLGEMMReshapeRHSMatrixKernel +using CLGEMMReshapeRHSMatrix = CLSynthetizeOperator<opencl::kernels::ClGemmReshapeRhsMatrixKernel>; + +// Create function for CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel +using CLGEMMLowpMatrixMultiplyReshapedOnlyRHS = CLSynthetizeOperator<opencl::kernels::ClGemmLowpMatrixMultiplyReshapedOnlyRhsMMULKernel>; + +// Fixture for CLGEMMLowpMatrixMultiplyReshapedOnlyRHS +using CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULFixture = + GEMMLowpMatrixMultiplyReshapedOnlyRHSMMULValidationFixture<CLTensor, CLAccessor, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshapedOnlyRHS>; + +// Fixture for CLGEMMLowpMatrixMultiplyReshapedOnlyRHS +using CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageFixtureSigned = + GEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageValidationFixture<int8_t, CLTensor, CLAccessor, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshapedOnlyRHS, CLReductionOperation, CLCast>; + +using CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageFixtureUnsigned = + GEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageValidationFixture<uint8_t, CLTensor, CLAccessor, CLGEMMReshapeRHSMatrix, CLGEMMLowpMatrixMultiplyReshapedOnlyRHS, CLReductionOperation, CLCast>; + +namespace +{ +// *INDENT-OFF* +// clang-format off + +/** M values to test */ +const auto m_values = framework::dataset::make("M", {16, 49}); + +/** N values to test */ +const auto n_values = framework::dataset::make("N", {16, 259}); + +/** K values to test */ +const auto k_values = framework::dataset::make("K", {192}); + +/** Batch size values to test */ +const auto b_values = framework::dataset::make("batch_size", {1, 2}); + +/** M0 values to test - Precommit */ +const auto m0 = framework::dataset::make("M0", {1, 2, 4}); + +/** N0 values to test - Precommit */ +const auto n0 = framework::dataset::make("N0", { 1, 4, 8}); + +/** K0 values to test - Precommit */ +const auto k0 = framework::dataset::make("K0", { 4 }); + +/** H0 values to test - Precommit */ +const auto h0 = framework::dataset::make("H0", 1); + +/** Interleave values to test with RHS matrix */ +const auto i_values_rhs = framework::dataset::make("interleave_rhs", { false }); + +/** Transpose values to test with RHS matrix */ +const auto t_values_rhs = framework::dataset::make("transpose_rhs", { true }); + +const auto broadcast_bias = framework::dataset::make("broadcast_bias", {true, false}); + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(GEMMLowpMatrixMultiplyReshapedOnlyRhsMMUL) +FIXTURE_DATA_TEST_CASE(Signed, CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0), + n0), + k0), + h0), + i_values_rhs), + t_values_rhs), + framework::dataset::make("DataType", { DataType::QASYMM8_SIGNED }))) +{ + // Validate output + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + validate(CLAccessor(_target), _reference); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +FIXTURE_DATA_TEST_CASE(Unsigned, CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0), + n0), + k0), + h0), + i_values_rhs), + t_values_rhs), + framework::dataset::make("DataType", { DataType::QASYMM8}))) +{ + // Validate output + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + validate(CLAccessor(_target), _reference); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +FIXTURE_DATA_TEST_CASE(OutputStageSigned, CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageFixtureSigned, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0), + n0), + k0), + h0), + i_values_rhs), + t_values_rhs), + broadcast_bias), + framework::dataset::make("DataType", { DataType::QASYMM8_SIGNED}))) +{ + // Validate output + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + validate(CLAccessor(_target), _reference); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +FIXTURE_DATA_TEST_CASE(OutputStageUnsigned, CLGEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageFixtureUnsigned, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0), + n0), + k0), + h0), + i_values_rhs), + t_values_rhs), + broadcast_bias), + framework::dataset::make("DataType", { DataType::QASYMM8}))) +{ + // Validate output + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + validate(CLAccessor(_target), _reference); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +TEST_SUITE_END() // GEMMLowpMatrixMultiplyReshapedOnlyRhsMMUL +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute
\ No newline at end of file diff --git a/tests/validation/CL/GEMMMatrixMultiply.cpp b/tests/validation/CL/GEMMMatrixMultiply.cpp deleted file mode 100644 index 21e085087d..0000000000 --- a/tests/validation/CL/GEMMMatrixMultiply.cpp +++ /dev/null @@ -1,339 +0,0 @@ -/* - * Copyright (c) 2019-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/core/KernelDescriptors.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/runtime/CL/CLTensor.h" -#include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h" -#include "tests/CL/CLAccessor.h" -#include "tests/CL/Helper.h" -#include "tests/PaddingCalculator.h" -#include "tests/datasets/ShapeDatasets.h" -#include "tests/framework/Asserts.h" -#include "tests/framework/Macros.h" -#include "tests/framework/datasets/Datasets.h" -#include "tests/validation/Validation.h" -#include "tests/validation/fixtures/GEMMFixture.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -using namespace arm_compute::misc::shape_calculator; -using namespace arm_compute::opencl::kernels; - -// Create function for CLGEMMMatrixMultiplyKernel -using CLGEMMMatrixMultiplyNative = CLSynthetizeOperator<ClGemmMatrixMultiplyKernel>; - -// Fixture for GEMMMatrixMultiplyValidationFixture -template <typename T> -using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; - -// Fixture for GEMMMatrixMultiply3DValidationFixture -template <typename T> -using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiply3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; - -namespace -{ -// *INDENT-OFF* -// clang-format off -RelativeTolerance<float> rel_tolerance_f32(0.001f); -constexpr float abs_tolerance_f32(0.0001f); - -RelativeTolerance<half> rel_tolerance_f16(half(0.2)); -constexpr float tolerance_num_f16 = 0.02f; - -/** Alpha values to test */ -const auto alpha_values = framework::dataset::make("alpha", {1.0f, -0.75f} ); - -/** Beta values to test */ -const auto beta_values = framework::dataset::make("beta", {-0.35f, 0.0f} ); - -/** M, N combinations to test - * 1: Special 1x1 case - * 2: Special multples of processor size in both dimensions - * 3: Non multiples of processor size in both dimensions - * 4: Special 1x1003 case -*/ -const auto m_n_values = zip( - framework::dataset::make("M", {1, 16, 37, 1}), - framework::dataset::make("N", {1, 16, 51, 1003}) - ); - -/** N values to test */ -const auto n_values = framework::dataset::make("N", {51, 1003}); - -/** K values to test */ -const auto k_values = framework::dataset::make("K", 23); - -/** M_W values to test */ -const auto m_w_values = framework::dataset::make("M_W", 5); - -/** M_H values to test */ -const auto m_h_values = framework::dataset::make("M_H", 7); - -/** Batch size values to test */ -const auto b_values = framework::dataset::make("batch_size", 1, 3); - -/** Activation values to test */ -const auto act_values = framework::dataset::make("Activation", -{ - ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), -}); - -/** Broadcast bias from vector to matrix */ -const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", { false, true } ); - -/** GPU architectures values to test */ -const auto gpu_arch_values = framework::dataset::make("GPUArch", -{ - GPUTarget::MIDGARD, - GPUTarget::BIFROST -}); - -/** Data types values to test in the configuration */ -const auto data_type_values = framework::dataset::make("DataType", -{ - DataType::F32, - DataType::F16 -}); - -/** M values to test */ -const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false}); -} // namespace - -TEST_SUITE(CL) -TEST_SUITE(GEMMMatrixMultiply) -TEST_CASE(Negative, framework::DatasetMode::ALL) -{ - // Unsupported QASYMM8 data type - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::QASYMM8); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::QASYMM8); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::QASYMM8); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Unsupported SIZE_T data type - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::SIZET); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::SIZET); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::SIZET); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Mixed precision with F32 - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = true; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Max number of dimensions LHS matrix - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U, 4U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Max number of dimensions RHS matrix - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 4U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 4U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 4U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Broadcast bias - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F16); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F16); - // The correct shape should be bias = TensorInfo(TensorShape(14U, 1U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F16); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F16); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, true); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Invalid dimensions for the bias - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32); - // The correct shape should be bias = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(14U, 8U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Invalid dimensions for the output - { - const auto lhs = TensorInfo(TensorShape(13U, 12U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(14U, 13U, 1U, 1U), 1, DataType::F32); - // The correct shape should be out = TensorInfo(TensorShape(14U, 12U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(14U, 7U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = false; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(12, 14, 13, 1, 1, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, nullptr, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } -} - -TEST_SUITE(Float) -TEST_SUITE(FP32) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_n_values, - k_values), - b_values), - alpha_values), - beta_values), - broadcast_bias_values), - framework::dataset::make("fp16_mixed_precision", false)), - act_values), - framework::dataset::make("DataType", DataType::F32)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); -} - -FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_w_values, - m_h_values), - n_values), - k_values), - b_values), - alpha_values), - beta_values), - broadcast_bias_values), - framework::dataset::make("fp16_mixed_precision", false)), - act_values), - framework::dataset::make("DataType", DataType::F32)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); -} - -TEST_SUITE_END() // FP32 - -TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_n_values, - k_values), - b_values), - alpha_values), - beta_values), - broadcast_bias_values), - fp16_mixed_precision_values), - act_values), - framework::dataset::make("DataType", DataType::F16)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); -} - -FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyNative3DFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_w_values, - m_h_values), - n_values), - k_values), - b_values), - alpha_values), - beta_values), - broadcast_bias_values), - fp16_mixed_precision_values), - act_values), - framework::dataset::make("DataType", DataType::F16)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); -} - -TEST_SUITE_END() // FP16 -TEST_SUITE_END() // Float -TEST_SUITE_END() // GEMMMatrixMuliplty -TEST_SUITE_END() // CL -} // namespace validation -} // namespace test -} // namespace arm_compute
\ No newline at end of file diff --git a/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp b/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp deleted file mode 100644 index e47518ad7d..0000000000 --- a/tests/validation/CL/GEMMMatrixMultiplyInterleavedTransposed.cpp +++ /dev/null @@ -1,334 +0,0 @@ -/* - * Copyright (c) 2019-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/core/KernelDescriptors.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/runtime/CL/CLTensor.h" -#include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" -#include "tests/CL/CLAccessor.h" -#include "tests/CL/Helper.h" -#include "tests/PaddingCalculator.h" -#include "tests/datasets/ShapeDatasets.h" -#include "tests/framework/Asserts.h" -#include "tests/framework/Macros.h" -#include "tests/framework/datasets/Datasets.h" -#include "tests/validation/Validation.h" -#include "tests/validation/fixtures/GEMMFixture.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -using namespace arm_compute::misc::shape_calculator; -using namespace arm_compute::opencl::kernels; - -// Create function for ClGemmReshapeLhsMatrixKernel -using CLGEMMReshapeLHSMatrix = CLSynthetizeOperator<ClGemmReshapeLhsMatrixKernel>; - -// Create function for ClGemmReshapeRhsMatrixKernel -using CLGEMMReshapeRHSMatrix = CLSynthetizeOperator<ClGemmReshapeRhsMatrixKernel>; - -// Create function for ClGemmMatrixMultiplyKernel -using CLGEMMMatrixMultiplyReshaped = CLSynthetizeOperator<ClGemmMatrixMultiplyKernel>; - -// Fixture for GEMMMatrixMultiplyInterleavedTransposedValidationFixture -template <typename T> -using CLGEMMMatrixMultiplyReshapedFixture = - GEMMMatrixMultiplyInterleavedTransposedValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>; - -// Fixture for GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture -template <typename T> -using CLGEMMMatrixMultiplyReshaped3DFixture = - GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>; - -namespace -{ -// *INDENT-OFF* -// clang-format off -RelativeTolerance<float> rel_tolerance_f32(0.001f); -constexpr float abs_tolerance_f32(0.0001f); - -RelativeTolerance<half> rel_tolerance_f16(half(0.2)); -constexpr float tolerance_num_f16 = 0.02f; - -/** Alpha values to test */ -const auto alpha_values = framework::dataset::make("alpha", {1.0f, -0.75f} ); - -/** Beta values to test */ -const auto beta_values = framework::dataset::make("beta", {-0.35f, 0.0f} ); - -/** M, N combinations to test - * 1: Special 1x1 case - * 2: Special multples of processor size in both dimensions - * 3: Non multiples of processor size in both dimensions -*/ -const auto m_n_values = zip( - framework::dataset::make("M", {1, 16, 37}), - framework::dataset::make("N", {1, 16, 51}) - ); - -/** N values to test */ -const auto n_values = framework::dataset::make("N", 51); - -/** K values to test */ -const auto k_values = framework::dataset::make("K", 23); - -/** M_W values to test */ -const auto m_w_values = framework::dataset::make("M_W", 5); - -/** M_H values to test */ -const auto m_h_values = framework::dataset::make("M_H", 7); - -/** Batch size values to test */ -const auto b_values = framework::dataset::make("batch_size", 1, 3); - -/** Activation values to test */ -const auto act_values = framework::dataset::make("Activation", -{ - ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), -}); - -/** V0 values to test */ -const auto v0_values = framework::dataset::make("V0", 2); - -/** H0 values to test */ -const auto h0_values = framework::dataset::make("H0", 4); - -/** Broadcast bias from vector to matrix */ -const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", {false, true} ); - -/** GPU architectures values to test */ -const auto gpu_arch_values = framework::dataset::make("GPUArch", -{ - GPUTarget::MIDGARD, - GPUTarget::BIFROST -}); - -/** Data types values to test in the configuration */ -const auto data_type_values = framework::dataset::make("DataType", -{ - DataType::F32, - DataType::F16 -}); - -/** M values to test */ -const auto fp16_mixed_precision_values = framework::dataset::make("fp16_mixed_precision", {true, false}); -} // namespace - -TEST_SUITE(CL) -TEST_SUITE(GEMMMatrixMultiplyInterleavedTransposed) -TEST_CASE(Negative, framework::DatasetMode::ALL) -{ - // The following tests are already integrated in the GEMMMatrixMultiply validation because - // in common with this validation - // - Unsupported QASYMM8 data type - // - Unsupported SIZE_T data type - // - Mixed precision with F32 - // - Max number of dimensions LHS matrix - // - Max number of dimensions RHS matrix - - // Invalid LHS dimensions - { - // The correct shape should be: lhs = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32); - const auto lhs = TensorInfo(TensorShape(256U, 2U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = true; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Invalid RHS dimensions - { - const auto lhs = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32); - // The correct shape should be rhs = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(104U, 4U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = true; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Broadcast bias - { - const auto lhs = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32); - // The correct shape should be bias = TensorInfo(TensorShape(24U, 1U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = true; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, true); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Invalid dimensions for the bias - { - const auto lhs = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32); - // The correct shape should be bias = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(25U, 16U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = true; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } - - // Invalid dimensions for the output - { - const auto lhs = TensorInfo(TensorShape(256U, 1U, 1U, 1U), 1, DataType::F32); - const auto rhs = TensorInfo(TensorShape(104U, 3U, 1U, 1U), 1, DataType::F32); - const auto bias = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - // The correct shape should be out = TensorInfo(TensorShape(24U, 16U, 1U, 1U), 1, DataType::F32); - const auto out = TensorInfo(TensorShape(24U, 13U, 1U, 1U), 1, DataType::F32); - constexpr float alpha = 1.3f; - constexpr float beta = 0.7f; - const bool is_interleaved_transposed = true; - const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(16, 24, 13, 2, 4, 0, false, false); - const GPUTarget gpu_target = GPUTarget::MIDGARD; - const bool fp_mixed_precision = false; - const auto status = ClGemmMatrixMultiplyKernel::validate(&lhs, &rhs, &bias, &out, alpha, beta, is_interleaved_transposed, reshape_info, gpu_target, fp_mixed_precision); - ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); - } -} - -TEST_SUITE(Float) -TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_n_values, - k_values), - b_values), - alpha_values), - beta_values), - v0_values), - h0_values), - broadcast_bias_values), - framework::dataset::make("fp16_mixed_precision", false)), - act_values), - framework::dataset::make("DataType", DataType::F32)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); -} - -FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_w_values, - m_h_values), - n_values), - k_values), - b_values), - alpha_values), - beta_values), - v0_values), - h0_values), - broadcast_bias_values), - framework::dataset::make("fp16_mixed_precision", false)), - act_values), - framework::dataset::make("DataType", DataType::F32)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); -} - -TEST_SUITE_END() // FP32 - -TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_n_values, - k_values), - b_values), - alpha_values), - beta_values), - v0_values), - h0_values), - broadcast_bias_values), - fp16_mixed_precision_values), - act_values), - framework::dataset::make("DataType", DataType::F16)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); -} - -FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_w_values, - m_h_values), - n_values), - k_values), - b_values), - alpha_values), - beta_values), - v0_values), - h0_values), - broadcast_bias_values), - fp16_mixed_precision_values), - act_values), - framework::dataset::make("DataType", DataType::F16)), - gpu_arch_values)) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16); -} - -TEST_SUITE_END() // FP16 -TEST_SUITE_END() // Float -TEST_SUITE_END() // GEMMMatrixMulipltyInterleavedTransposed -TEST_SUITE_END() // CL -} // namespace validation -} // namespace test -} // namespace arm_compute
\ No newline at end of file diff --git a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp index a737c687c4..0ddf43766f 100644 --- a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp +++ b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,7 +26,7 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h" +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyNativeKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" @@ -91,8 +91,8 @@ const auto b_values = framework::dataset::make("batch_size", 1, 3); /** Activation values to test */ const auto act_values = framework::dataset::make("Activation", { - ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), }); /** M0 values to test - Precommit */ @@ -323,6 +323,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, f // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); } + TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float TEST_SUITE_END() // GEMMMatrixMulipltyNative diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp index 6f368a9650..b06e4bf213 100644 --- a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp +++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,9 +26,9 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" @@ -110,6 +110,7 @@ const auto b_values = framework::dataset::make("batch_size", 2, 3); const auto act_values = framework::dataset::make("Activation", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), }); /** Alpha values to test - Precommit */ @@ -328,7 +329,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zi framework::dataset::make("Expected", { true, true, false, false, false, true, true,true})), input0_info ,input1_info, input2_info, output_info, lhs_info, rhs_info, gemm_info, expected) { - ARM_COMPUTE_EXPECT(bool(ClGemmMatrixMultiplyReshapedKernel::validate(&input0_info.clone()->set_is_resizable(true), + ARM_COMPUTE_EXPECT(bool(ClGemmMatrixMultiplyReshapedKernel::validate(&input0_info.clone()->set_is_resizable(true), &input1_info.clone()->set_is_resizable(true), &input2_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true),1.f,1.f, @@ -336,6 +337,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zi rhs_info, gemm_info)) == expected, framework::LogLevel::ERRORS); } + TEST_SUITE(Float) TEST_SUITE(FP32) @@ -361,7 +363,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<float>, fra act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<float>, framework::DatasetMode::DISABLED, @@ -386,7 +396,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<float>, fra act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::ALL, @@ -411,7 +429,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::DatasetMode::DISABLED, @@ -436,8 +462,17 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } + TEST_SUITE(ExportToCLImage) DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( framework::dataset::make("Input0Info", { TensorInfo(TensorShape(256U, 16U, 2U), 1, DataType::F32), // OK or incorrect if cl_khr_image2d_from_buffer not supported @@ -560,7 +595,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zi true, true, false, - false})), + true})), input0_info ,input1_info, input2_info, output_info, lhs_info, rhs_info, gemm_info, expected) { ARM_COMPUTE_EXPECT(bool(ClGemmMatrixMultiplyReshapedKernel::validate(&input0_info.clone()->set_is_resizable(true), @@ -704,6 +739,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::ARM_COMPUTE_PRINT_INFO(); } } + TEST_SUITE_END() // ExportToCLImage TEST_SUITE_END() // FP32 @@ -731,7 +767,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedFixture<half>, fram act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<half>, framework::DatasetMode::DISABLED, @@ -756,7 +800,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedFixture<half>, fram act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::ALL, @@ -781,7 +833,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::DatasetMode::DISABLED, @@ -806,7 +866,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } TEST_SUITE(ExportToCLImage) @@ -931,7 +999,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zi true, true, false, - false})), + true})), input0_info ,input1_info, input2_info, output_info, lhs_info, rhs_info, gemm_info, expected) { ARM_COMPUTE_EXPECT(bool(ClGemmMatrixMultiplyReshapedKernel::validate(&input0_info.clone()->set_is_resizable(true), @@ -1075,6 +1143,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::ARM_COMPUTE_PRINT_INFO(); } } + TEST_SUITE_END() // ExportToCLImage TEST_SUITE_END() // FP16 @@ -1102,7 +1171,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedMixedPrecisionFixtu act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedMixedPrecisionFixture<half>, framework::DatasetMode::DISABLED, @@ -1127,7 +1204,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMMatrixMultiplyReshapedMixedPrecisionFixtu act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DMixedPrecisionFixture<half>, framework::DatasetMode::ALL, @@ -1152,7 +1237,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall3D, CLGEMMMatrixMultiplyReshaped3DMixedPrecisionF act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DMixedPrecisionFixture<half>, framework::DatasetMode::DISABLED, @@ -1177,8 +1270,17 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DMixedPrecisionF act_values)) { // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16_mixed_precision, 0.f, abs_tolerance_f16_mixed_precision); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } } + TEST_SUITE_END() // MixedPrecision TEST_SUITE_END() // Float TEST_SUITE_END() // GEMMMatrixMultiplyReshaped diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp index 88e99bcfef..dafc8dc5ec 100644 --- a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp +++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,8 +26,8 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" @@ -99,6 +99,7 @@ const auto b_values = framework::dataset::make("batch_size", 2); const auto act_values = framework::dataset::make("Activation", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 10.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), }); /** M0 values to test - precommit */ @@ -211,6 +212,7 @@ bool validate_configuration(unsigned int m_value, unsigned int n_value, unsigned CLGEMMMatrixMultiplyReshapedOnlyRHS gemm; return bool(gemm.validate(&lhs, &rhs_reshaped, &bias, &dst, alpha, beta, lhs_info, rhs_info, kernel_info)); } + } // namespace TEST_SUITE(CL) @@ -462,6 +464,7 @@ FIXTURE_DATA_TEST_CASE(RunNightly3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixtur framework::ARM_COMPUTE_PRINT_INFO(); } } + TEST_SUITE_END() // FP32 TEST_SUITE(FP16) @@ -590,6 +593,7 @@ FIXTURE_DATA_TEST_CASE(RunNightly3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixtur framework::ARM_COMPUTE_PRINT_INFO(); } } + TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float diff --git a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRhsMMUL.cpp b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRhsMMUL.cpp new file mode 100644 index 0000000000..3b3cf85317 --- /dev/null +++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRhsMMUL.cpp @@ -0,0 +1,232 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "src/gpu/cl/kernels/ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "tests/CL/CLAccessor.h" +#include "tests/CL/Helper.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/GEMMFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::opencl::kernels; + +// Create function for ClGemmReshapeRhsMatrixKernel +using CLGEMMReshapeRHSMatrix = CLSynthetizeOperator<ClGemmReshapeRhsMatrixKernel>; + +// Create function for ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel +using CLGEMMMatrixMultiplyReshapedOnlyRhsMMUL = CLSynthetizeOperator<ClGemmMatrixMultiplyReshapedOnlyRhsMMULKernel>; + +// Fixture for CLGEMMMatrixMultiplyReshapedOnlyRhsMMUL +template <typename T> +using CLGEMMMatrixMultiplyReshapedOnlyRhsMMULFixture = GEMMMatrixMultiplyReshapedOnlyRhsMMULValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRhsMMUL>; + +namespace +{ +// *INDENT-OFF* +// clang-format off +RelativeTolerance<float> rel_tolerance_f32(0.001f); +constexpr float abs_tolerance_f32(0.0001f); +RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.001f)); +constexpr float abs_tolerance_f16(0.3f); + +/** Alpha values to test - Precommit */ +const auto a_values = framework::dataset::make("alpha", {1.0f, 0.75f} ); + +/** Beta values to test - Precommit */ +const auto beta_values = framework::dataset::make("beta", {0.0f, -0.75f} ); + +/** M values to test */ +const auto m_values = framework::dataset::make("M", {49}); + +/** N values to test */ +const auto n_values = framework::dataset::make("N", {257}); + +/** K values to test */ +/** The test case requires this to be multiple of 4*/ +const auto k_values = framework::dataset::make("K", {192}); + +/** Batch size values to test */ +const auto b_values = framework::dataset::make("batch_size", {1, 2}); + +/** Activation values to test */ +const auto act_values = framework::dataset::make("Activation", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), +}); + +/** M0 values to test - Precommit */ +const auto m0_values_precommit = framework::dataset::make("M0", { 1, 2, 4 }); + +/** N0 values to test - Precommit */ +const auto n0_values_precommit = framework::dataset::make("N0", { 4, 8 }); + +/** K0 values to test - Precommit */ +const auto k0_values_precommit = framework::dataset::make("K0", { 1 }); + +/** Broadcast bias from vector to matrix */ +const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", { false, true } ); + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(GEMMMatrixMultiplyReshapedOnlyRhsMMUL) +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRhsMMULFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportToCLImage", false)), + framework::dataset::make("DataType", DataType::F32)), + a_values), + beta_values), + broadcast_bias_values), + act_values)) +{ + // Validate output + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRhsMMULFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportToCLImage", false)), + framework::dataset::make("DataType", DataType::F16)), + a_values), + beta_values), + broadcast_bias_values), + act_values)) +{ + // Validate output + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +TEST_SUITE_END() // FP16 + +TEST_SUITE(ExportToCLImage) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRhsMMULFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportToCLImage", true)), + framework::dataset::make("DataType", DataType::F32)), + a_values), + beta_values), + broadcast_bias_values), + act_values)) +{ + // Validate output + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedOnlyRhsMMULFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + m_values, + n_values), + k_values), + b_values), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportToCLImage", true)), + framework::dataset::make("DataType", DataType::F16)), + a_values), + beta_values), + broadcast_bias_values), + act_values)) +{ + // Validate output + if(validate_result) + { + validate(CLAccessor(_target), _reference, rel_tolerance_f16, 0.f, abs_tolerance_f16); + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // ExportToCLImage +TEST_SUITE_END() // Float +TEST_SUITE_END() // GEMMMatrixMultiplyReshapedOnlyRhsMMUL +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/GEMMReshapeLHSMatrix.cpp b/tests/validation/CL/GEMMReshapeLHSMatrix.cpp index f995608308..0dd9b811f6 100644 --- a/tests/validation/CL/GEMMReshapeLHSMatrix.cpp +++ b/tests/validation/CL/GEMMReshapeLHSMatrix.cpp @@ -25,7 +25,7 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeLhsMatrixKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" @@ -66,8 +66,10 @@ const auto b_values = framework::dataset::make("batchsize", 1, 3); /** M0 values to test */ const auto m0_values_s32 = framework::dataset::make("M0", { 2, 3 }); -const auto m0_values_s16 = framework::dataset::make("M0", { 4, 5 }); -const auto m0_values_s8 = framework::dataset::make("M0", { 6, 7, 8 }); +const auto m0_values_s16 = framework::dataset::make("M0", { 4 }); +const auto m0_values_s16_nt = framework::dataset::make("M0", { 5 }); +const auto m0_values_s8_nt = framework::dataset::make("M0", { 6,7 }); +const auto m0_values_s8 = framework::dataset::make("M0", { 8 }); /** K0 values to test */ const auto k0_values_s32 = framework::dataset::make("K0", { 2, 3 }); @@ -101,6 +103,7 @@ FIXTURE_DATA_TEST_CASE(S32, CLGEMMReshapeLHSMatrixFixture<int>, framework::Datas // Validate output validate(CLAccessor(_target), _reference); } + FIXTURE_DATA_TEST_CASE(S16, CLGEMMReshapeLHSMatrixFixture<short>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallGEMMReshape2DShapes(), b_values), @@ -114,6 +117,7 @@ FIXTURE_DATA_TEST_CASE(S16, CLGEMMReshapeLHSMatrixFixture<short>, framework::Dat // Validate output validate(CLAccessor(_target), _reference); } + FIXTURE_DATA_TEST_CASE(S8, CLGEMMReshapeLHSMatrixFixture<char>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallGEMMReshape2DShapes(), b_values), @@ -128,6 +132,37 @@ FIXTURE_DATA_TEST_CASE(S8, CLGEMMReshapeLHSMatrixFixture<char>, framework::Datas validate(CLAccessor(_target), _reference); } +TEST_SUITE(NotTransposed) +FIXTURE_DATA_TEST_CASE(S16, CLGEMMReshapeLHSMatrixFixture<short>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::SmallGEMMReshape2DShapes(), + b_values), + framework::dataset::make("DataType", DataType::S16)), + m0_values_s16_nt), + k0_values_s16), + v0_values), + i_values), + framework::dataset::make("transpose", { false }))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(S8, CLGEMMReshapeLHSMatrixFixture<char>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::SmallGEMMReshape2DShapes(), + b_values), + framework::dataset::make("DataType", DataType::S8)), + m0_values_s8_nt), + k0_values_s8), + v0_values), + i_values), + framework::dataset::make("transpose", { false }))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +TEST_SUITE_END() + TEST_SUITE(ReinterpretInputAs3D) FIXTURE_DATA_TEST_CASE(S32, CLGEMMReshapeLHSMatrix3DFixture<int>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallGEMMReshape3DShapes(), diff --git a/tests/validation/CL/GEMMReshapeRHSMatrix.cpp b/tests/validation/CL/GEMMReshapeRHSMatrix.cpp index ff1240ea2e..f8462058a6 100644 --- a/tests/validation/CL/GEMMReshapeRHSMatrix.cpp +++ b/tests/validation/CL/GEMMReshapeRHSMatrix.cpp @@ -25,7 +25,7 @@ #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" -#include "src/core/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" +#include "src/gpu/cl/kernels/ClGemmReshapeRhsMatrixKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" diff --git a/tests/validation/CL/Gather.cpp b/tests/validation/CL/Gather.cpp index f0b87d7d9f..7619baae1e 100644 --- a/tests/validation/CL/Gather.cpp +++ b/tests/validation/CL/Gather.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,19 +48,21 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 27U), 1, DataType::F16), TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid Indices data type - TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid Indices dimensionality - TensorInfo(TensorShape(5U, 5U, 5U, 5U, 5U), 1, DataType::F32), // Invalid Input dimensionality - TensorInfo(TensorShape(27U, 27U), 1, DataType::F16), // Mismatching data type input/output - TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid positive axis value - TensorInfo(TensorShape(27U, 27U), 1, DataType::F16), // Invalid negative axis value + TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid Output shape + TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid Indices data type + TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid Indices dimensionality + TensorInfo(TensorShape(5U, 5U, 5U, 5U, 5U), 1, DataType::F32), // Invalid Input dimensionality + TensorInfo(TensorShape(27U, 27U), 1, DataType::F16), // Mismatching data type input/output + TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), // Invalid positive axis value + TensorInfo(TensorShape(27U, 27U), 1, DataType::F16), // Invalid negative axis value }), framework::dataset::make("IndicesInfo", { TensorInfo(TensorShape(10U), 1, DataType::U32), TensorInfo(TensorShape(10U), 1, DataType::U32), TensorInfo(TensorShape(10U), 1, DataType::U32), - TensorInfo(TensorShape(10U), 1, DataType::U8), TensorInfo(TensorShape(10U, 10U), 1, DataType::U32), + TensorInfo(TensorShape(10U), 1, DataType::U8), + TensorInfo(TensorShape(10U, 10U, 10U, 10U), 1, DataType::U32), TensorInfo(TensorShape(10U), 1, DataType::U32), TensorInfo(TensorShape(10U), 1, DataType::U32), TensorInfo(TensorShape(10U), 1, DataType::U32), @@ -71,7 +73,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( TensorInfo(TensorShape(27U, 10U), 1, DataType::F32), TensorInfo(TensorShape(10U, 27U), 1, DataType::F32), TensorInfo(TensorShape(10U, 27U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 10U), 1, DataType::F32), + TensorInfo(TensorShape(10U, 27U), 1, DataType::F32), + TensorInfo(TensorShape(27U, 10U, 10U, 10U, 10U), 1, DataType::F32), TensorInfo(TensorShape(10U, 5U, 5U, 5U, 5U), 1, DataType::F32), TensorInfo(TensorShape(27U, 10U), 1, DataType::F32), TensorInfo(TensorShape(27U, 27U), 1, DataType::F32), @@ -82,13 +85,14 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( 1, -2, 0, + 0, 1, 0, 1, 2, -3, })), - framework::dataset::make("Expected", { true, true, true, false, false, false, false, false, false })), + framework::dataset::make("Expected", { true, true, true, false, false, false, false, false, false, false })), input_info, indices_info, output_info, axis, expected) { const Status status = CLGather::validate(&input_info.clone()->set_is_resizable(true), &indices_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), axis); @@ -111,6 +115,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmallMultiDimIndices, + CLGatherFixture<half>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::CLSmallGatherMultiDimIndicesDataset(), framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, CLGatherFixture<half>, framework::DatasetMode::NIGHTLY, @@ -131,6 +144,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmallMultiDimIndices, + CLGatherFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::CLSmallGatherMultiDimIndicesDataset(), framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, CLGatherFixture<float>, framework::DatasetMode::NIGHTLY, @@ -152,6 +174,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmallMultiDimIndices, + CLGatherFixture<uint8_t>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::CLSmallGatherMultiDimIndicesDataset(), framework::dataset::make("DataType", DataType::U8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, CLGatherFixture<uint8_t>, framework::DatasetMode::NIGHTLY, @@ -172,6 +203,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmallMultiDimIndices, + CLGatherFixture<uint16_t>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::CLSmallGatherMultiDimIndicesDataset(), framework::dataset::make("DataType", DataType::U16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + + FIXTURE_DATA_TEST_CASE(RunLarge, CLGatherFixture<uint16_t>, framework::DatasetMode::NIGHTLY, diff --git a/tests/validation/CL/Im2Col.cpp b/tests/validation/CL/Im2Col.cpp index c6006efcba..1f5b781690 100644 --- a/tests/validation/CL/Im2Col.cpp +++ b/tests/validation/CL/Im2Col.cpp @@ -22,7 +22,7 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "src/core/CL/kernels/CLIm2ColKernel.h" +#include "src/gpu/cl/kernels/ClIm2ColKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/framework/Asserts.h" @@ -40,7 +40,7 @@ namespace validation TEST_SUITE(CL) TEST_SUITE(Im2Col) -using CLIm2Col = CLSynthetizeFunction<CLIm2ColKernel>; +using ClIm2Col = ClSynthetizeOperatorWithBorder<opencl::kernels::ClIm2ColKernel>; /** Negative tests * @@ -63,7 +63,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto output = TensorInfo(TensorShape(9U, 10U, 12U, 2U), 1, DataType::F32); const auto conv_size = Size2D(3, 3); const bool has_bias = false; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -73,7 +73,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto output = TensorInfo(TensorShape(9U, 80U, 2U), 1, DataType::QASYMM8); const auto conv_size = Size2D(3, 3); const bool has_bias = true; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -84,7 +84,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto conv_size = Size2D(3, 3); const auto dilation = Size2D(0, 1); const bool has_bias = false; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -96,7 +96,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto dilation = Size2D(1, 1); const bool has_bias = false; const unsigned int num_groups = 2; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -108,7 +108,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto dilation = Size2D(1, 1); const bool has_bias = false; const unsigned int num_groups = 2; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias, dilation, num_groups); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -118,7 +118,7 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto output = TensorInfo(TensorShape(9U, 81U, 2U), 1, DataType::F32); const auto conv_size = Size2D(3, 3); const bool has_bias = false; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } @@ -128,13 +128,13 @@ TEST_CASE(Negative, framework::DatasetMode::ALL) const auto output = TensorInfo(TensorShape(1U, 1U, 1U, 2U), 1, DataType::F32, DataLayout::NHWC); const auto conv_size = Size2D(9, 9); const bool has_bias = false; - const auto status = CLIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); + const auto status = opencl::kernels::ClIm2ColKernel::validate(&input, &output, conv_size, PadStrideInfo(), has_bias); ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); } } template <typename T> -using CLIm2ColFixture = Im2ColValidationFixture<CLTensor, CLAccessor, CLIm2Col, T, true>; +using ClIm2ColFixture = Im2ColOpValidationFixture<CLTensor, CLAccessor, ClIm2Col, T, true>; TEST_SUITE(NHWC) @@ -150,7 +150,7 @@ TEST_SUITE(NHWC) * Kernel tested im2col3x3_nhwc */ FIXTURE_DATA_TEST_CASE(W3x3, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", @@ -180,7 +180,7 @@ framework::dataset::make("Groups", 1))) * Kernel tested im2col9x9_nhwc */ FIXTURE_DATA_TEST_CASE(W9x9, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", @@ -210,7 +210,7 @@ framework::dataset::make("Groups", 1))) * Kernel tested im2col_generic_nhwc */ FIXTURE_DATA_TEST_CASE(Generic, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", @@ -243,7 +243,7 @@ TEST_SUITE(NCHW) * Kernel tested im2col1x1_stridex1_nchw */ FIXTURE_DATA_TEST_CASE(W1x1_Stride1_NoPad, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", { TensorShape(4U, 4U, 3U, 2U), TensorShape(5U, 4U, 3U, 2U), TensorShape(3U, 4U, 3U, 2U) }), @@ -267,7 +267,7 @@ FIXTURE_DATA_TEST_CASE(W1x1_Stride1_NoPad, * Kernel tested im2col3x3_nchw */ FIXTURE_DATA_TEST_CASE(W3x3, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(4U, 4U, 3U, 2U)), @@ -291,7 +291,7 @@ FIXTURE_DATA_TEST_CASE(W3x3, * Kernel tested im2col5x5_nchw */ FIXTURE_DATA_TEST_CASE(W5x5, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(7U, 4U, 3U, 2U)), @@ -317,7 +317,7 @@ FIXTURE_DATA_TEST_CASE(W5x5, * Kernel tested im2col11x11_padx0_pady0_nchw */ FIXTURE_DATA_TEST_CASE(W11x11_NoPad, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", { TensorShape(11U, 11U, 2U, 2U), TensorShape(14U, 13U, 1U, 2U) }), @@ -341,7 +341,7 @@ FIXTURE_DATA_TEST_CASE(W11x11_NoPad, * Kernel tested im2col_generic_padx0_pady0_nchw */ FIXTURE_DATA_TEST_CASE(GenericZeroPad, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(13U, 11U, 2U, 2U)), @@ -367,7 +367,7 @@ TEST_SUITE_END() // NCHW * Kernel tested im2col_generic_(nchw|nhwc) */ FIXTURE_DATA_TEST_CASE(Generic, - CLIm2ColFixture<float>, + ClIm2ColFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(13U, 11U, 5U, 2U)), @@ -393,7 +393,7 @@ FIXTURE_DATA_TEST_CASE(Generic, * - im2col9x9_nhwc */ FIXTURE_DATA_TEST_CASE(Quantized, - CLIm2ColFixture<uint8_t>, + ClIm2ColFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(13U, 11U, 11U, 2U)), @@ -419,7 +419,7 @@ FIXTURE_DATA_TEST_CASE(Quantized, * - im2col9x9_nhwc */ FIXTURE_DATA_TEST_CASE(FP16, - CLIm2ColFixture<half>, + ClIm2ColFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( framework::dataset::make("InputShape", TensorShape(13U, 11U, 11U, 2U)), diff --git a/tests/validation/CL/IndirectConv2dAddressPrecalculation.cpp b/tests/validation/CL/IndirectConv2dAddressPrecalculation.cpp new file mode 100644 index 0000000000..67f70685d1 --- /dev/null +++ b/tests/validation/CL/IndirectConv2dAddressPrecalculation.cpp @@ -0,0 +1,89 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "src/gpu/cl/kernels/ClIndirectConv2dAddressPrecalculationKernel.h" +#include "tests/CL/CLAccessor.h" +#include "tests/CL/Helper.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::misc::shape_calculator; +using namespace arm_compute::opencl::kernels; + +using CLIndirectConv2dAddressPrecalculation = CLSynthetizeOperator<ClIndirectConv2dAddressPrecalculationKernel>; + +using CLIndirectConv2dAddressPrecalculationFixture = IndirectConv2dAddressPrecalculationValidationFixture<CLTensor, CLAccessor, CLIndirectConv2dAddressPrecalculation>; + +// *INDENT-OFF* +// clang-format off +/** Data types */ + +namespace +{ +const auto src_w_values = framework::dataset::make("src_w", {91}); +const auto src_h_values = framework::dataset::make("src_h", {103}); +const auto src_b_values = framework::dataset::make("src_b", {1, 2}); +const auto wei_w_values = framework::dataset::make("wei_w", {3, 5}); +const auto wei_h_values = framework::dataset::make("wei_h", {1, 6}); +const auto pad_values = framework::dataset::make("pad", {1, 2, 3}); +const auto stride_values = framework::dataset::make("stride", {1, 2}); +const auto m0_values = framework::dataset::make("M0", { 1, 2, 4, 5, 7 }); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(IndirectConv2dAddressPrecalculation) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLIndirectConv2dAddressPrecalculationFixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(src_w_values, + src_h_values), + src_b_values), + wei_w_values), + wei_h_values), + pad_values), + stride_values), + m0_values)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +TEST_SUITE_END() // IndirectConv2dAddressPrecalculation +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/IndirectConvolutionLayer.cpp b/tests/validation/CL/IndirectConvolutionLayer.cpp new file mode 100644 index 0000000000..aedf070e6b --- /dev/null +++ b/tests/validation/CL/IndirectConvolutionLayer.cpp @@ -0,0 +1,268 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLIndirectConvolutionLayer.h" +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/DirectConvolutionLayerFixture.h" + +// Note: Since the interface of indirect convolution is the same of direct convolution, we can reuse +// the direct convolution fixture + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<half> tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */ +RelativeTolerance<float> tolerance_fp32(0.05f); /**< Tolerance for floating point tests */ +constexpr float abs_tolerance_f32(0.0001f); /**< Absolute tolerance for FP32 tests*/ +constexpr float tolerance_num = 0.07f; /**< Tolerance number */ + +/** Activation function Dataset*/ +const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +{ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) }); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(IndirectConvolutionLayer) + +/** Check whether the configuration of a indirect convolution layer with no + * bias leads to a successful run. + */ +TEST_CASE(NoBias, framework::DatasetMode::PRECOMMIT) +{ + const TensorShape src_shape_nhwc = TensorShape(8U, 27U, 13U); + const TensorShape wei_shape_nhwc = TensorShape(8U, 3U, 3U, 4U); + const TensorShape bia_shape = TensorShape(4U); + const TensorShape dst_shape_nhwc = TensorShape(4U, 25U, 11U); + constexpr DataType dt = DataType::F32; + constexpr DataLayout data_layout = DataLayout::NHWC; + + auto src_nhwc = create_tensor<CLTensor>(src_shape_nhwc, dt, 1, QuantizationInfo(), data_layout); + auto wei_nhwc = create_tensor<CLTensor>(wei_shape_nhwc, dt, 1, QuantizationInfo(), data_layout); + auto dst_nhwc = create_tensor<CLTensor>(dst_shape_nhwc, dt, 1, QuantizationInfo(), data_layout); + + TensorShape src_shape_nchw = src_shape_nhwc; + TensorShape wei_shape_nchw = wei_shape_nhwc; + TensorShape dst_shape_nchw = dst_shape_nhwc; + + permute(src_shape_nchw, PermutationVector(1U, 2U, 0U)); + permute(wei_shape_nchw, PermutationVector(1U, 2U, 0U, 3U)); + permute(dst_shape_nchw, PermutationVector(1U, 2U, 0U)); + + const PadStrideInfo conv_info = PadStrideInfo(1, 1, 0, 0); + + // Create indirect Convolution function + CLIndirectConvolutionLayer conv{}; + conv.configure(&src_nhwc, &wei_nhwc, nullptr, &dst_nhwc, conv_info); + + src_nhwc.allocator()->allocate(); + wei_nhwc.allocator()->allocate(); + dst_nhwc.allocator()->allocate(); + + library->fill_tensor_value(CLAccessor(src_nhwc), 1.f); + library->fill_tensor_value(CLAccessor(wei_nhwc), 1.f); + + conv.run(); + + // Compute reference to compare + SimpleTensor<float> ref_src{ src_shape_nchw, dt }; + SimpleTensor<float> ref_wei{ wei_shape_nchw, dt }; + SimpleTensor<float> ref_bia{ bia_shape, dt }; + library->fill_tensor_value(ref_src, 1.f); + library->fill_tensor_value(ref_wei, 1.f); + // No bias + library->fill_tensor_value(ref_bia, 0.f); + auto ref_dst = reference::convolution_layer<float>(ref_src, ref_wei, ref_bia, dst_shape_nchw, conv_info); + + validate(CLAccessor(dst_nhwc), ref_dst); +} + +/** Check whether the case of rectangle kernels i.e. when width and height of the weight_shape are not equal + * would lead to successful run + */ +TEST_CASE(NonSquareKernel, framework::DatasetMode::PRECOMMIT) +{ + const TensorShape src_shape_nhwc = TensorShape(3U, 33U, 27U); + const TensorShape wei_shape_nhwc = TensorShape(3U, 5U, 7U, 4U); // non-square kernel + const TensorShape bia_shape = TensorShape(4U); + const TensorShape dst_shape_nhwc = TensorShape(4U, 11U, 12U); + constexpr DataType dt = DataType::F32; + constexpr DataLayout data_layout = DataLayout::NHWC; + + auto src_nhwc = create_tensor<CLTensor>(src_shape_nhwc, dt, 1, QuantizationInfo(), data_layout); + auto wei_nhwc = create_tensor<CLTensor>(wei_shape_nhwc, dt, 1, QuantizationInfo(), data_layout); + auto dst_nhwc = create_tensor<CLTensor>(dst_shape_nhwc, dt, 1, QuantizationInfo(), data_layout); + + TensorShape src_shape_nchw = src_shape_nhwc; + TensorShape wei_shape_nchw = wei_shape_nhwc; + TensorShape dst_shape_nchw = dst_shape_nhwc; + + permute(src_shape_nchw, PermutationVector(1U, 2U, 0U)); + permute(wei_shape_nchw, PermutationVector(1U, 2U, 0U, 3U)); + permute(dst_shape_nchw, PermutationVector(1U, 2U, 0U)); + + const PadStrideInfo conv_info = PadStrideInfo(3, 2, 1, 1, 2, 0, DimensionRoundingType::FLOOR); + + // Create indirect convolution function + CLIndirectConvolutionLayer conv{}; + conv.configure(&src_nhwc, &wei_nhwc, nullptr, &dst_nhwc, conv_info); + + src_nhwc.allocator()->allocate(); + wei_nhwc.allocator()->allocate(); + dst_nhwc.allocator()->allocate(); + + library->fill_tensor_value(CLAccessor(src_nhwc), 1.f); + library->fill_tensor_value(CLAccessor(wei_nhwc), 1.f); + + conv.run(); + + // Compute reference to compare + SimpleTensor<float> ref_src{ src_shape_nchw, dt }; + SimpleTensor<float> ref_wei{ wei_shape_nchw, dt }; + SimpleTensor<float> ref_bia{ bia_shape, dt }; + library->fill_tensor_value(ref_src, 1.f); + library->fill_tensor_value(ref_wei, 1.f); + // No bias + library->fill_tensor_value(ref_bia, 0.f); + auto ref_dst = reference::convolution_layer<float>(ref_src, ref_wei, ref_bia, dst_shape_nchw, conv_info); + + validate(CLAccessor(dst_nhwc), ref_dst); +} +// *INDENT-OFF* +// clang-format off +// Note: Since the interface of indirect convolution is the same of direct convolution, we can reuse +// the direct convolution fixture +template <typename T> +using CLIndirectConvolutionLayerFixture = DirectConvolutionValidationFixture<CLTensor, CLAccessor, CLIndirectConvolutionLayer, T>; +template <typename T> +using CLIndirectConvolutionLayerMixedDataLayoutFixture = DirectConvolutionValidationFixture<CLTensor, CLAccessor, CLIndirectConvolutionLayer, T, true>; + +TEST_SUITE(NHWC) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLIndirectConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U), + TensorShape(32U, 37U, 13U) } ), + framework::dataset::make("StrideX", { 1, 3, 1, 1 })), + framework::dataset::make("StrideY", { 1, 3, 2, 1 })), + framework::dataset::make("PadX", { 1, 3, 0, 4 })), + framework::dataset::make("PadY", { 1, 3, 0, 4 })), + framework::dataset::make("KernelSize", { 3, 8, 1, 9 })), + framework::dataset::make("NumKernels", { 17, 3, 1, 19 })), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLIndirectConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(800U, 800U, 3U) } ), + framework::dataset::make("StrideX", { 1 })), + framework::dataset::make("StrideY", { 1 })), + framework::dataset::make("PadX", { 1 })), + framework::dataset::make("PadY", { 1 })), + framework::dataset::make("KernelSize", { 9 })), + framework::dataset::make("NumKernels", { 3 })), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::IDENTITY) )), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp16, tolerance_num); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLIndirectConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U), + TensorShape(32U, 37U, 13U) } ), + framework::dataset::make("StrideX", { 1, 3, 1, 1 })), + framework::dataset::make("StrideY", { 1, 3, 2, 1 })), + framework::dataset::make("PadX", { 1, 3, 0, 4 })), + framework::dataset::make("PadY", { 1, 3, 0, 4 })), + framework::dataset::make("KernelSize", { 3, 8, 1, 9 })), + framework::dataset::make("NumKernels", { 17, 3, 1, 19 })), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLIndirectConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U), + TensorShape(32U, 37U, 13U) } ), + framework::dataset::make("StrideX", { 1 })), + framework::dataset::make("StrideY", { 2 })), + framework::dataset::make("PadX", { 1 })), + framework::dataset::make("PadY", { 3 })), + framework::dataset::make("KernelSize", { 3 })), + framework::dataset::make("NumKernels", { 3 })), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) )), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLIndirectConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(800U, 800U, 3U) } ), + framework::dataset::make("StrideX", { 1 })), + framework::dataset::make("StrideY", { 1 })), + framework::dataset::make("PadX", { 1 })), + framework::dataset::make("PadY", { 1 })), + framework::dataset::make("KernelSize", { 9 })), + framework::dataset::make("NumKernels", { 3 })), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::IDENTITY) )), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_fp32, 0.0, abs_tolerance_f32); +} +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // NHWC +TEST_SUITE_END() // IndirectConvolutionLayer +TEST_SUITE_END() // CL + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/MatMul.cpp b/tests/validation/CL/MatMul.cpp new file mode 100644 index 0000000000..844597f3e9 --- /dev/null +++ b/tests/validation/CL/MatMul.cpp @@ -0,0 +1,232 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/functions/CLMatMul.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ActivationFunctionsDataset.h" +#include "tests/framework/DatasetModes.h" +#include "tests/framework/Macros.h" +#include "tests/framework/TestCase.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" + +#include "tests/datasets/LargeMatMulDataset.h" +#include "tests/datasets/SmallMatMulDataset.h" +#include "tests/validation/fixtures/MatMulFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for fp32 data type */ +constexpr float abs_tolerance_f32( + 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp32 data type in case using relative tolerance fails because of small values */ +constexpr float abs_tolerance_f16( + 0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data type in case using relative tolerance fails because of small values */ +RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for fp16 data type */ +constexpr AbsoluteTolerance<uint8_t> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +} // namespace + +template <typename T> +using CLMatMulFixture = MatMulValidationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>; + +template <typename T> +using CLQuantizedMatMulFixture = QuantizedMatMulValidationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>; + +template <typename T> +using CLMatMulActivationFixture = MatMulValidationWithActivationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>; + +template <typename T> +using CLMatMulActivationAlphaBetaFixture = MatMulValidationWithActivationAlphaBetaFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>; + +template <typename T> +using CLQuantizedMatMulActivationFixture = QuantizedMatMulValidationWithActivationFixture<CLTensor, CLAccessor, CLMatMul, GpuMatMulSettings, T>; + +/* The main act functions matmul (float) is expected to support */ +const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.75f, 0.25f), +}); + +/* (Float datatype only) Larger activation functions dataset, used during some nightly tests. */ +const auto AllActivationsDataset = combine(datasets::ActivationFunctions(), framework::dataset::make("AlphaBeta", { 0.5f, 1.f })); + +// Alpha beta values should be integer values +// This is for testing purposes with quantized datatypes and is not a limitation of the kernel. +// To properly remove this restriction, dst_qinfo should be auto-initialised with consideration for alpha beta values +// The main act functions quantized matmul kernels are expected to support +const auto ActivationFunctionsQuantizedDataset = concat(concat(concat( + framework::dataset::make("ActivationInfo", ActivationLayerInfo()), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU))), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 1.f))), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 2.f, 1.f))); + +TEST_SUITE(CL) +TEST_SUITE(MatMul) + +TEST_SUITE(Float) +TEST_SUITE(FP32) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulActivationFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::F32)), + ActivationFunctionsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLMatMulActivationFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::F32)), + ActivationFunctionsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunAllActivations, CLMatMulActivationAlphaBetaFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::SmallerMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + framework::dataset::make("DataType", DataType::F32)), + AllActivationsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} + +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulActivationFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::F16)), + ActivationFunctionsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLMatMulActivationFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::F16)), + ActivationFunctionsDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} + +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLQuantizedMatMulFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine( + datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + ActivationFunctionsQuantizedDataset), + framework::dataset::make("NumberOfExtraRuns", { 0, 1 })), + framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 50, 1) })), + framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 30, -1) })), + framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLQuantizedMatMulFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine( + datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + ActivationFunctionsQuantizedDataset), + framework::dataset::make("NumberOfExtraRuns", { 0, 1 })), + framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })), + framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })), + framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} + +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLQuantizedMatMulFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine( + datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + ActivationFunctionsQuantizedDataset), + framework::dataset::make("NumberOfExtraRuns", { 0, 1 })), + framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 50, 1) })), + framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 30, -1) })), + framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 2) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLQuantizedMatMulFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine( + datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + ActivationFunctionsQuantizedDataset), + framework::dataset::make("NumberOfExtraRuns", { 0, 1 })), + framework::dataset::make("LhsQInfo", { QuantizationInfo(1.f / 100, 1) })), + framework::dataset::make("RhsQInfo", { QuantizationInfo(1.f / 200, -1) })), + framework::dataset::make("DstQInfo", { QuantizationInfo(1.f, 50) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} + +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE_END() // Quantized + +TEST_SUITE_END() // MatMul +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/MatMulKernel.cpp b/tests/validation/CL/MatMulKernel.cpp new file mode 100644 index 0000000000..b47f8bc924 --- /dev/null +++ b/tests/validation/CL/MatMulKernel.cpp @@ -0,0 +1,650 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/runtime/CL/CLTensor.h" +#include "src/gpu/cl/kernels/ClMatMulNativeKernel.h" +#include "tests/datasets/LargeMatMulDataset.h" +#include "tests/datasets/SmallMatMulDataset.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/MatMulKernelFixture.h" +#include "tests/validation/reference/Permute.h" + +#include <tuple> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +constexpr float abs_tolerance_f32( + 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for floating point data types in case using relative tolerance fails because of small values */ +constexpr float abs_tolerance_f16( + 0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data types in case using relative tolerance fails because of small values */ +RelativeTolerance<half_float::half> tolerance_f16(half(0.01)); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +} // namespace + +/** M0 values to test --precommit*/ +const auto m0_values_precommit = framework::dataset::make("M0", { 1, 3 }); + +/** N0 values to test --precommit*/ +const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 }); + +/** K0 values to test --precommit*/ +const auto k0_values_precommit = framework::dataset::make("K0", { 2, 3 }); + +/** M0 values to test --nightly*/ +const auto m0_values_nightly_lhs_nt = framework::dataset::make("M0", { 1, 2, 3, 4, 5, 6, 7, 8 }); +const auto m0_values_nightly_lhs_t = framework::dataset::make("M0", { 1, 2, 3, 4, 8 }); + +/** N0 values to test --nightly*/ +const auto n0_values_nightly_rhs_nt = framework::dataset::make("N0", { 1, 2, 3, 4, 8, 16 }); +const auto n0_values_nightly_rhs_t = framework::dataset::make("N0", { 1, 2, 3, 4, 8 }); + +/** K0 values to test --nightly*/ +const auto k0_values_nightly_lhs_nt_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 8, 16 }); +const auto k0_values_nightly_rhs_t = framework::dataset::make("K0", { 1, 2, 3, 4, 8 }); +const auto k0_values_nightly_lhs_t_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 5, 6, 7, 8 }); + +template <typename T> +using CLMatMulKernelFixture = MatMulKernelValidationFixture<T, ClMatMulNativeKernel>; + +template <typename T> +using CLMatMulKernelBiasFixture = MatMulKernelWithBiasValidation<T, ClMatMulNativeKernel>; + +TEST_SUITE(CL) +TEST_SUITE(MatMulKernel) +TEST_SUITE(Validate) + +TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL) +{ + using MatMulConfigurationPair = std::pair<MatMulKernelInfo, bool>; + + const std::vector<MatMulConfigurationPair> supported_block_sizes = + { + // MatMulKernelInfo(adj_lhs, adj_rhs, M0, N0, K0, export_rhs_to_cl_image = false) + // Lhs not-transposed, Rhs-not-transposed + { MatMulKernelInfo(false, false, 0, 1, 1), false }, // M0 should be > 0 + { MatMulKernelInfo(false, false, 3, 5, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 6, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 17), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 7), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 9, 1, 2), true }, + { MatMulKernelInfo(false, false, 3, 16, 3), true }, + { MatMulKernelInfo(false, false, 7, 3, 4), true }, + { MatMulKernelInfo(false, false, 7, 3, 4, true), false }, // N0 not in {4, 8, 16} + { MatMulKernelInfo(false, false, 7, 1, 4, true), false }, // N0 not in {4, 8, 16} + { MatMulKernelInfo(false, false, 7, 12, 4, true), false }, // N0 not in {4, 8, 16} + { MatMulKernelInfo(false, false, 7, 4, 4, true), true }, + { MatMulKernelInfo(false, false, 7, 8, 4, true), true }, + { MatMulKernelInfo(false, false, 7, 16, 4, true), true }, + + // Lhs not-transposed, Rhs transposed + { MatMulKernelInfo(false, true, 0, 1, 1), false }, // M0 should be > 0 + { MatMulKernelInfo(false, true, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 3, 3, 12), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 3, 3, 6), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, true, 5, 1, 2), true }, + { MatMulKernelInfo(false, true, 3, 3, 3), true }, + { MatMulKernelInfo(false, true, 2, 4, 8), true }, + { MatMulKernelInfo(false, true, 2, 4, 5, true), false }, // K0 not in {4, 8, 16} + { MatMulKernelInfo(false, true, 2, 4, 9, true), false }, // K0 not in {4, 8, 16} + { MatMulKernelInfo(false, true, 2, 4, 3, true), false }, // K0 not in {4, 8, 16} + { MatMulKernelInfo(false, true, 2, 4, 4, true), true }, + { MatMulKernelInfo(false, true, 2, 4, 8, true), true }, + { MatMulKernelInfo(false, true, 2, 8, 16, true), true }, + + // Lhs transposed, Rhs-not-transposed + { MatMulKernelInfo(true, false, 1, 1, 0), false }, // K0 should be > 0 + { MatMulKernelInfo(true, false, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 6, 3, 12), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 5, 3, 6), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 4, 1, 22), true }, + { MatMulKernelInfo(true, false, 3, 3, 3), true }, + { MatMulKernelInfo(true, false, 2, 4, 8), true }, + { MatMulKernelInfo(true, false, 2, 3, 8, true), false }, // N0 not in {4, 8, 16} + { MatMulKernelInfo(true, false, 2, 7, 8, true), false }, // N0 not in {4, 8, 16} + { MatMulKernelInfo(true, false, 2, 5, 8, true), false }, // N0 not in {4, 8, 16} + { MatMulKernelInfo(true, false, 2, 4, 8, true), true }, + { MatMulKernelInfo(true, false, 2, 8, 8, true), true }, + { MatMulKernelInfo(true, false, 2, 16, 8, true), true }, + + // Lhs transposed, Rhs-transposed + { MatMulKernelInfo(true, true, 2, 1, 5), false }, // K0 should in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 1, 8, 7), false }, // K0 should in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 6, 3, 12), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 5, 3, 6), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 4, 8, 16), true }, + { MatMulKernelInfo(true, true, 3, 3, 4), true }, + { MatMulKernelInfo(true, true, 16, 4, 8), true }, + { MatMulKernelInfo(true, true, 2, 2, 1, true), false }, // K0 not in {4, 8, 16} + { MatMulKernelInfo(true, true, 2, 2, 5, true), false }, // K0 not in {4, 8, 16} + { MatMulKernelInfo(true, true, 2, 4, 7, true), false }, // K0 not in {4, 8, 16} + { MatMulKernelInfo(true, true, 2, 4, 4, true), true }, + { MatMulKernelInfo(true, true, 2, 8, 8, true), true }, + { MatMulKernelInfo(true, true, 2, 8, 16, true), true }, + }; + + // Set big enough shapes so that block sizes are not truncated. Also, set all dimensions equal + // so that it doesn't fail for different NT/T configurations. We aim to test the block sizes here, + // not the shapes themselves. + const TensorInfo lhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32); + + const bool export_to_cl_image_supported = image2d_from_buffer_supported(CLKernelLibrary::get().get_device()); + for(auto &pair : supported_block_sizes) + { + TensorInfo output_info; + Status status = ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &output_info, pair.first); + + if(!pair.first.export_rhs_to_cl_image || export_to_cl_image_supported) + { + ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS); + } + } +} + +TEST_CASE(ExportToCLImage, framework::DatasetMode::ALL) +{ + // We skip this test if the hardware does not support exporting to CL Image + if(image2d_from_buffer_supported(CLKernelLibrary::get().get_device())) + { + constexpr size_t pixel_size = 4; + const size_t max_image_w = pixel_size * CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_WIDTH>(); + const size_t max_image_h = CLKernelLibrary::get().get_device().getInfo<CL_DEVICE_IMAGE2D_MAX_HEIGHT>(); + + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, bool, bool, bool>; + const std::vector<ShapeConfigurationTuple> shape_configurations = + { + // lhs_shape, rhs_shape, adj_lhs, adj_rhs, expected + // Lhs t/Nt, Rhs Nt + // Transposition of Lhs doesn't add any value to the tests, therefore always assumed false below + { TensorShape(5U, 1U), TensorShape(3U, 5U), false, false, false }, // N should be multiple of 4 + { TensorShape(5U, 1U), TensorShape(14U, 5U), false, false, false }, // N should be multiple of 4 + { TensorShape(5U, 1U), TensorShape(12U, 5U), false, false, true }, + { TensorShape(5U, 1U), TensorShape(8U, 5U), false, false, true }, + { TensorShape(5U, 1U), TensorShape(4U, 5U), false, false, true }, + { TensorShape(max_image_h + 1, 1U), TensorShape(4U, max_image_h + 1), false, false, false }, // Cannot fit into CL Image memory's height + { TensorShape(5U, 1U), TensorShape(max_image_w + 1, 5U), false, false, false }, // Cannot fit into CL Image memory's width + { TensorShape(max_image_h, 1U), TensorShape(4U, max_image_h), false, false, true }, // Barely fits into CL Image memory's height + { TensorShape(5U, 1U), TensorShape(max_image_w, 5U), false, false, true }, // Barely fits into CL Image memory's width + + // Lhs Nt/T , Rhs T + { TensorShape(5U, 1U), TensorShape(5U, 3U), false, true, false }, // K should be multiple of 4 + { TensorShape(5U, 1U), TensorShape(5U, 14U), false, true, false }, // K should be multiple of 4 + { TensorShape(4U, 1U), TensorShape(4U, 10U), false, true, true }, + { TensorShape(8U, 1U), TensorShape(8U, 9U), false, true, true }, + { TensorShape(12U, 1U), TensorShape(12U, 6U), false, true, true }, + }; + + for(auto &tuple : shape_configurations) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + + const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::F32); + + const bool adj_lhs = std::get<2>(tuple); + const bool adj_rhs = std::get<3>(tuple); + + // We choose M0, N0, K0 equal to 4 so that they're always valid for CLImage in any combination + const MatMulKernelInfo matmul_kernel_info + { + adj_lhs, adj_rhs, 4, 4, 4, true /* export_rhs_to_cl_image */ + }; + + TensorInfo output_info; + Status status = ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &output_info, matmul_kernel_info); + + const bool expected = std::get<4>(tuple); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } +} + +TEST_CASE(ValidateInputShapes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, TensorShape, bool>; + const std::vector<ShapeConfigurationTuple> shape_configurations = + { + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(3U), true }, + { TensorShape(10U, 12U), TensorShape(3U, 10U), TensorShape(3U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 8U), TensorShape(2U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 5U), TensorShape(2U), false }, // Mismatch in the K dimension + { TensorShape(5U, 0U), TensorShape(2U, 5U), TensorShape(2U), false }, // Invalid dimension + { TensorShape(5U, 4U, 3U, 4U, 5U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), true }, + { TensorShape(5U, 4U, 3U, 4U, 5U, 1U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // no batch broadcasting + { TensorShape(5U, 4U, 3U, 4U, 9U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // mismatch in batch dimension + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(1U), false }, // Unsupported bias broadcasting. + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(3U, 3U), false }, // 2D bias is unsupported. + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(6U), false }, // bias first dimension != dst first dimension + }; + + for(auto &tuple : shape_configurations) + { + const bool expected = std::get<3>(tuple); + + for(bool adj_lhs : + { + false, true + }) + { + for(bool adj_rhs : + { + false, true + }) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + TensorShape bia_shape = std::get<2>(tuple); + + if(adj_lhs) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + + if(adj_rhs) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::F32); + const TensorInfo bia_info = TensorInfo(bia_shape, 1, DataType::F32); + TensorInfo output_info; + + MatMulKernelInfo matmul_kernel_info{ adj_lhs, adj_rhs, 1, 1, 1, false /* export_rhs_to_cl_image */ }; + + Status status = ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + } +} + +TEST_CASE(ValidateDataTypes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using DataTypeConfigurationTuple = std::tuple<DataType, DataType, DataType, bool>; + const std::vector<DataTypeConfigurationTuple> data_type_configurations = + { + { DataType::F32, DataType::F32, DataType::F32, true }, + { DataType::F16, DataType::F16, DataType::F16, true }, + { DataType::F16, DataType::F32, DataType::F32, false }, // no mixed precision + { DataType::F64, DataType::F64, DataType::F64, false }, // no double precision + { DataType::QASYMM8, DataType::QASYMM8, DataType::QASYMM8, false }, // no quantized types + { DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, false }, // no quantized types + { DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, false }, // no quantized types + { DataType::QASYMM16, DataType::QASYMM16, DataType::QASYMM16, false }, // no quantized types + { DataType::QSYMM16, DataType::QSYMM16, DataType::QSYMM16, false }, // no quantized types + { DataType::QSYMM8, DataType::QSYMM8, DataType::QSYMM8, false }, // no quantized types + { DataType::S64, DataType::S64, DataType::S64, false }, // no integral types + { DataType::S32, DataType::S32, DataType::S32, false }, // no integral types + { DataType::S16, DataType::S16, DataType::S16, false }, // no integral types + { DataType::S8, DataType::S8, DataType::S8, false }, // no integral types + { DataType::U64, DataType::U64, DataType::U64, false }, // no integral types + { DataType::U32, DataType::U32, DataType::U32, false }, // no integral types + { DataType::U16, DataType::U16, DataType::U16, false }, // no integral types + { DataType::U8, DataType::U8, DataType::U8, false }, // no integral types + }; + + const TensorShape shape = TensorShape(10U, 10U); + const MatMulKernelInfo matmul_kernel_info{ false, false, 1, 1, 1, false }; + for(auto &tuple : data_type_configurations) + { + const bool expected = std::get<3>(tuple); + + const TensorInfo lhs_info(shape, 1, std::get<0>(tuple)); + const TensorInfo rhs_info(shape, 1, std::get<1>(tuple)); + TensorInfo output_info(shape, 1, std::get<2>(tuple)); + + Status status = ClMatMulNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } +} + +TEST_SUITE_END() // Validate + +TEST_SUITE(Float) +TEST_SUITE(FP32) +TEST_SUITE(Buffer) +FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunWithBias, CLMatMulKernelBiasFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) + +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_nt_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_t_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +// Running High Dimensional test is enough for FP32, because we're stressing the number of dimensions, not data type or M0/N0/K0 +// It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels +FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 2 })), + framework::dataset::make("K0", { 2 })), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +TEST_SUITE_END() // Buffer + +TEST_SUITE(ExportRhsToCLImage) +FIXTURE_DATA_TEST_CASE(RunSmallRhsNotTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDatasetRhsExportToCLImageRhsNT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { false })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 4, 8, 16 })), + framework::dataset::make("K0", { 2, 4 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsNotTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDatasetRhsExportToCLImageRhsNT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { false })), + framework::dataset::make("M0", { 2 })), // Choices of M0 does not matter much because it's related to Lhs tensor + framework::dataset::make("N0", { 4, 8, 16 })), + framework::dataset::make("K0", { 1, 2, 3, 4 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunSmallRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDatasetRhsExportToCLImageRhsT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 2, 4 })), + framework::dataset::make("K0", { 4, 8, 16 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDatasetRhsExportToCLImageRhsT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true })), + framework::dataset::make("M0", { 2 })), // Choices of M0 does not matter much because it's related to Lhs tensor + framework::dataset::make("N0", { 1, 2, 3, 4 })), + framework::dataset::make("K0", { 4, 8, 16 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +TEST_SUITE_END() // ExportRhsToCLImage +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +TEST_SUITE(Buffer) +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_nt_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_t_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} +TEST_SUITE_END() // Buffer + +TEST_SUITE(ExportRhsToCLImage) +FIXTURE_DATA_TEST_CASE(RunSmallRhsNotTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDatasetRhsExportToCLImageRhsNT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { false })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 4, 8, 16 })), + framework::dataset::make("K0", { 2, 4 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsNotTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDatasetRhsExportToCLImageRhsNT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { false })), + framework::dataset::make("M0", { 2 })), // Choices of M0 does not matter much because it's related to Lhs tensor + framework::dataset::make("N0", { 4, 8, 16 })), + framework::dataset::make("K0", { 1, 2, 3, 4 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunSmallRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDatasetRhsExportToCLImageRhsT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 2, 4 })), + framework::dataset::make("K0", { 4, 8, 16 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDatasetRhsExportToCLImageRhsT(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true })), + framework::dataset::make("M0", { 2 })), // Choices of M0 does not matter much because it's related to Lhs tensor + framework::dataset::make("N0", { 1, 2, 3, 4 })), + framework::dataset::make("K0", { 4, 8, 16 })), + framework::dataset::make("ExportRhsToCLImage", { true })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_export_to_cl_image) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +TEST_SUITE_END() // ExportRhsToCLImage +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float +TEST_SUITE_END() // MatMulKernel +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/MatMulLowpNativeKernel.cpp b/tests/validation/CL/MatMulLowpNativeKernel.cpp new file mode 100644 index 0000000000..90eee4fb82 --- /dev/null +++ b/tests/validation/CL/MatMulLowpNativeKernel.cpp @@ -0,0 +1,411 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "src/gpu/cl/kernels/ClMatMulLowpNativeKernel.h" + +#include "tests/datasets/LargeMatMulDataset.h" +#include "tests/datasets/SmallMatMulDataset.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/MatMulKernelFixture.h" +#include "tests/validation/reference/Permute.h" + +#include <tuple> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +} +template <typename T> +using CLMatMulLowpNativeKernelFixture = MatMulKernelValidationFixture<T, ClMatMulLowpNativeKernel>; + +template <typename T> +using CLMatMulLowpKernelWithBiasFixture = MatMulKernelWithBiasValidation<T, ClMatMulLowpNativeKernel>; + +/** M0 values to test --precommit*/ +const auto m0_values_precommit = framework::dataset::make("M0", { 1, 3 }); + +/** N0 values to test --precommit*/ +const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 }); + +/** K0 values to test --precommit*/ +const auto k0_values_precommit = framework::dataset::make("K0", { 2, 3 }); + +/** M0 values to test --nightly*/ +const auto m0_values_nightly_lhs_nt = framework::dataset::make("M0", { 1, 2, 3, 4, 5, 6, 7, 8 }); +const auto m0_values_nightly_lhs_t = framework::dataset::make("M0", { 1, 2, 3, 4, 8 }); + +/** N0 values to test --nightly*/ +const auto n0_values_nightly_rhs_nt = framework::dataset::make("N0", { 1, 2, 3, 4, 8, 16 }); +const auto n0_values_nightly_rhs_t = framework::dataset::make("N0", { 1, 2, 3, 4, 8 }); + +/** K0 values to test --nightly*/ +const auto k0_values_nightly_lhs_nt_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 8, 16 }); +const auto k0_values_nightly_rhs_t = framework::dataset::make("K0", { 1, 2, 3, 4, 8 }); +const auto k0_values_nightly_lhs_t_rhs_nt = framework::dataset::make("K0", { 1, 2, 3, 4, 5, 6, 7, 8 }); + +TEST_SUITE(CL) +TEST_SUITE(MatMulLowpNativeKernel) +TEST_SUITE(Validate) + +TEST_CASE(SupportedKernelConfigurations, framework::DatasetMode::ALL) +{ + using MatMulConfigurationPair = std::pair<MatMulKernelInfo, bool>; + + const std::vector<MatMulConfigurationPair> supported_block_sizes = + { + // MatMulKernelInfo(adj_lhs, adj_rhs, M0, N0, K0, export_rhs_to_cl_image = false) + // Lhs not-transposed, Rhs-not-transposed + { MatMulKernelInfo(false, false, 0, 1, 1), false }, // M0 should be > 0 + { MatMulKernelInfo(false, false, 3, 5, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 6, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 17), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 7), false }, // K0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 9, 1, 2), true }, + { MatMulKernelInfo(false, false, 3, 16, 3), true }, + { MatMulKernelInfo(false, false, 7, 3, 4), true }, + { MatMulKernelInfo(false, false, 7, 3, 4, true), true }, // export to CLImage is unsupported for quantized types + }; + + // Set big enough shapes so that block sizes are not truncated. Also, set all dimensions equal + // so that it doesn't fail for different NT/T configurations. We aim to test the block sizes here, + // not the shapes themselves. + const TensorInfo lhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::QASYMM8_SIGNED); + const TensorInfo rhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::QASYMM8_SIGNED); + + for(auto &pair : supported_block_sizes) + { + TensorInfo output_info; + Status status = ClMatMulLowpNativeKernel::validate(&lhs_info, &rhs_info, nullptr, &output_info, pair.first); + + ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS); + } +} + +TEST_CASE(ValidateInputShapes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, TensorShape, bool>; + const std::vector<ShapeConfigurationTuple> shape_configurations = + { + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(3U), true }, + { TensorShape(10U, 12U), TensorShape(3U, 10U), TensorShape(3U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 8U), TensorShape(2U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 5U), TensorShape(2U), false }, // Mismatch in the K dimension + { TensorShape(5U, 0U), TensorShape(2U, 5U), TensorShape(2U), false }, // Invalid dimension + { TensorShape(5U, 4U, 3U, 4U, 5U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), true }, + { TensorShape(5U, 4U, 3U, 4U, 5U, 1U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // no batch broadcasting + { TensorShape(5U, 4U, 3U, 4U, 9U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // mismatch in batch dimension + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(1U), false }, // invalid broadcast of bias + { TensorShape(5U, 1U), TensorShape(3U, 5U), TensorShape(3U, 3U), false }, // 2d bias is invalid + }; + + for(auto &tuple : shape_configurations) + { + const bool expected = std::get<3>(tuple); + + for(bool adj_lhs : + { + false, true + }) + { + for(bool adj_rhs : + { + false, true + }) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + TensorShape bia_shape = std::get<2>(tuple); + + if(adj_lhs) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + + if(adj_rhs) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::QASYMM8_SIGNED); + const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::QASYMM8_SIGNED); + const TensorInfo bia_info = TensorInfo(bia_shape, 1, DataType::S32); + TensorInfo output_info; + + MatMulKernelInfo matmul_kernel_info{ adj_lhs, adj_rhs, 1, 1, 1, false /* export_rhs_to_cl_image */ }; + + Status status = ClMatMulLowpNativeKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + } +} + +TEST_CASE(ValidateDataTypes, framework::DatasetMode::ALL) +{ + using DataTypeConfigurationTuple = std::tuple<DataType, DataType, DataType, DataType, bool>; + const std::vector<DataTypeConfigurationTuple> data_type_configurations = + { + { DataType::F32, DataType::F32, DataType::F32, DataType::F32, false }, // no floating point types + { DataType::F16, DataType::F16, DataType::F16, DataType::F16, false }, // no floating point types + { DataType::F64, DataType::F64, DataType::F64, DataType::F64, false }, // no double precision + { DataType::QASYMM8, DataType::QASYMM8, DataType::S32, DataType::QASYMM8, true }, + { DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, DataType::S32, DataType::QASYMM8_SIGNED, true }, + { DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, DataType::S32, DataType::QSYMM8_PER_CHANNEL, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QASYMM16, DataType::QASYMM16, DataType::S32, DataType::QASYMM16, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QSYMM16, DataType::QSYMM16, DataType::S32, DataType::QSYMM16, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QSYMM8, DataType::QSYMM8, DataType::S32, DataType::QSYMM8, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::QASYMM8, false }, // no mixed data types + { DataType::S64, DataType::S64, DataType::S64, DataType::S64, false }, // no integral types + { DataType::S32, DataType::S32, DataType::S32, DataType::S32, false }, // no integral types + { DataType::S16, DataType::S16, DataType::S16, DataType::S16, false }, // no integral types + { DataType::S8, DataType::S8, DataType::S8, DataType::S8, false }, // no integral types + { DataType::U64, DataType::U64, DataType::U64, DataType::U64, false }, // no integral types + { DataType::U32, DataType::U32, DataType::U32, DataType::U32, false }, // no integral types + { DataType::U16, DataType::U16, DataType::U16, DataType::U16, false }, // no integral types + { DataType::U8, DataType::U8, DataType::U8, DataType::U8, false }, // no integral types + { DataType::QASYMM8, DataType::QASYMM8, DataType::F32, DataType::QASYMM8, false } // Only S32 bias is supported + }; + + // It's enough to test a single shape and block size configuration while checking data types + const TensorShape shape = TensorShape(10U, 10U); + const TensorShape bia_shape = TensorShape(10U); + const MatMulKernelInfo matmul_kernel_info{ false, false, 1, 1, 1, false }; + for(auto &tuple : data_type_configurations) + { + const bool expected = std::get<4>(tuple); + + const TensorInfo lhs_info(shape, 1, std::get<0>(tuple)); + const TensorInfo rhs_info(shape, 1, std::get<1>(tuple)); + const TensorInfo bia_info(bia_shape, 1, std::get<2>(tuple)); + TensorInfo output_info(shape, 1, std::get<3>(tuple)); + + Status status = ClMatMulLowpNativeKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } +} + +TEST_SUITE_END() // Validate + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true, false })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true, false })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunWithBias, CLMatMulLowpKernelWithBiasFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true, false })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_nt_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_t_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +// Running High Dimensional test is enough for qasymm8_signed, because we're stressing the number of dimensions, not data type or M0/N0/K0 +// It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels +FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulLowpNativeKernelFixture<int8_t>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulDataset(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true, false })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 2 })), + framework::dataset::make("K0", { 2 })), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulDataset(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true, false })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulDataset(), + framework::dataset::make("TransposeA", { true, false })), + framework::dataset::make("TransposeB", { true, false })), + m0_values_precommit), + n0_values_precommit), + k0_values_precommit), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_nt_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTransposed, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_values_nightly_lhs_t_rhs_nt), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulLowpNativeKernelFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_values_nightly_rhs_t), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::QASYMM8))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // QASYMM8 +TEST_SUITE_END() // Quantized +TEST_SUITE_END() // MatMulLowpNativeKernel +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/MatMulLowpNativeMMULKernel.cpp b/tests/validation/CL/MatMulLowpNativeMMULKernel.cpp new file mode 100644 index 0000000000..ac46b67c9e --- /dev/null +++ b/tests/validation/CL/MatMulLowpNativeMMULKernel.cpp @@ -0,0 +1,394 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "src/gpu/cl/kernels/ClMatMulLowpNativeMMULKernel.h" + +#include "tests/datasets/MatMulLowpMMULDataset.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/MatMulKernelFixture.h" +#include "tests/validation/reference/Permute.h" + +#include <tuple> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +} +using framework::dataset::make; + +template <typename T> +using CLMatMulLowpNativeMMULKernelFixture = MatMulKernelValidationFixture<T, ClMatMulLowpNativeMMULKernel, true /* use_mmul */>; + +template <typename T> +using CLMatMulLowpNativeMMULKernelWithBiasFixture = MatMulKernelWithBiasValidation<T, ClMatMulLowpNativeMMULKernel, true /* use_mmul */>; + +/** M0 values to test --precommit*/ +const auto m0_values_precommit = framework::dataset::make("M0", { 1, 3 }); + +/** N0 values to test --precommit*/ +const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 }); + +/** M0 values to test --nightly*/ +const auto m0_values_nightly_lhs_nt = framework::dataset::make("M0", { 2, 4, 5, 8 }); +const auto m0_values_nightly_lhs_t = framework::dataset::make("M0", { 2, 4, 8 }); + +/** N0 values to test --nightly*/ +const auto n0_values_nightly = framework::dataset::make("N0", { 1, 3, 8, 16 }); + +TEST_SUITE(CL) +TEST_SUITE(MatMulLowpNativeMMULKernel) +TEST_SUITE(Validate) + +TEST_CASE(SupportedKernelConfigurations, framework::DatasetMode::ALL) +{ + using MatMulConfigurationPair = std::pair<MatMulKernelInfo, bool>; + + const std::vector<MatMulConfigurationPair> supported_block_sizes = + { + // MatMulKernelInfo(adj_lhs, adj_rhs, M0, N0, K0, export_rhs_to_cl_image = false) + { MatMulKernelInfo(false, false, 0, 1, 4), false }, // M0 should be > 0 + { MatMulKernelInfo(false, true, 3, 5, 4), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 6, 4), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 8), false }, // K0 not in 4 + { MatMulKernelInfo(true, false, 5, 3, 4), false }, // M0 not in {1, 2, 3, 4, 8, 16} when Lhs is transposed + { MatMulKernelInfo(false, false, 9, 1, 4), true }, + { MatMulKernelInfo(false, true, 3, 16, 4), true }, + { MatMulKernelInfo(false, false, 7, 3, 4), true }, + { MatMulKernelInfo(true, false, 8, 3, 4), true }, + { MatMulKernelInfo(true, true, 4, 3, 4), true }, + { MatMulKernelInfo(false, false, 7, 3, 4, true), false }, // export to CLImage is unsupported for quantized types + }; + + // Set big enough shapes so that block sizes are not truncated. Also, set all dimensions equal + // so that it doesn't fail for different NT/T configurations. We aim to test the block sizes here, + // not the shapes themselves. + const TensorInfo lhs_info = TensorInfo(TensorShape(64U, 64U), 1, DataType::QASYMM8_SIGNED); + const TensorInfo rhs_info = TensorInfo(TensorShape(64U, 64U), 1, DataType::QASYMM8_SIGNED); + + for(auto &pair : supported_block_sizes) + { + TensorInfo output_info; + Status status = ClMatMulLowpNativeMMULKernel::validate(&lhs_info, &rhs_info, nullptr, &output_info, pair.first); + const bool expected = (pair.second && arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())); + + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } +} + +TEST_CASE(ValidateInputShapes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, TensorShape, bool>; + const std::vector<ShapeConfigurationTuple> shape_configurations = + { + { TensorShape(32U, 1U), TensorShape(3U, 32U), TensorShape(3U), true }, + { TensorShape(16U, 12U), TensorShape(3U, 16U), TensorShape(3U), true }, + { TensorShape(64U, 4U), TensorShape(2U, 64U), TensorShape(2U), true }, + { TensorShape(16U, 4U), TensorShape(2U, 32U), TensorShape(2U), false }, // Mismatch in the K dimension + { TensorShape(16U, 0U), TensorShape(2U, 16U), TensorShape(2U), false }, // Invalid dimension + { TensorShape(32U, 4U, 3U, 4U, 5U, 6U), TensorShape(2U, 32U, 3U, 4U, 5U, 6U), TensorShape(2U), true }, + { TensorShape(32U, 4U, 3U, 4U, 5U, 1U), TensorShape(2U, 32U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // no batch broadcasting + { TensorShape(32U, 4U, 3U, 4U, 9U, 6U), TensorShape(2U, 32U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // mismatch in batch dimension + { TensorShape(32U, 1U), TensorShape(3U, 32U), TensorShape(1U), false }, // invalid broadcast of bias + { TensorShape(32U, 1U), TensorShape(3U, 32U), TensorShape(3U, 3U), false }, // 2d bias is invalid + { TensorShape(12U, 12U), TensorShape(3U, 12U), TensorShape(3U), false }, // K must be multiple of 16 + }; + + for(auto &tuple : shape_configurations) + { + const bool expected = (std::get<3>(tuple) && arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())); + + for(bool adj_lhs : + { + false, true + }) + { + for(bool adj_rhs : + { + false, true + }) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + TensorShape bia_shape = std::get<2>(tuple); + + if(adj_lhs) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + + if(adj_rhs) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::QASYMM8_SIGNED); + const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::QASYMM8_SIGNED); + const TensorInfo bia_info = TensorInfo(bia_shape, 1, DataType::S32); + TensorInfo output_info; + + MatMulKernelInfo matmul_kernel_info{ adj_lhs, adj_rhs, 1, 1, 4, false /* export_rhs_to_cl_image */ }; + + Status status = ClMatMulLowpNativeMMULKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + } +} + +TEST_CASE(ValidateDataTypes, framework::DatasetMode::ALL) +{ + using DataTypeConfigurationTuple = std::tuple<DataType, DataType, DataType, DataType, bool>; + const std::vector<DataTypeConfigurationTuple> data_type_configurations = + { + { DataType::F32, DataType::F32, DataType::F32, DataType::F32, false }, // no floating point types + { DataType::F16, DataType::F16, DataType::F16, DataType::F16, false }, // no floating point types + { DataType::F64, DataType::F64, DataType::F64, DataType::F64, false }, // no double precision + { DataType::QASYMM8, DataType::QASYMM8, DataType::S32, DataType::QASYMM8, true }, + { DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, DataType::S32, DataType::QASYMM8_SIGNED, true }, + { DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, DataType::S32, DataType::QSYMM8_PER_CHANNEL, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QASYMM16, DataType::QASYMM16, DataType::S32, DataType::QASYMM16, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QSYMM16, DataType::QSYMM16, DataType::S32, DataType::QSYMM16, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QSYMM8, DataType::QSYMM8, DataType::S32, DataType::QSYMM8, false }, // only qasymm8/qasymm8_signed is supported + { DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::S32, DataType::QASYMM8, false }, // no mixed data types + { DataType::S64, DataType::S64, DataType::S64, DataType::S64, false }, // no integral types + { DataType::S32, DataType::S32, DataType::S32, DataType::S32, false }, // no integral types + { DataType::S16, DataType::S16, DataType::S16, DataType::S16, false }, // no integral types + { DataType::S8, DataType::S8, DataType::S8, DataType::S8, false }, // no integral types + { DataType::U64, DataType::U64, DataType::U64, DataType::U64, false }, // no integral types + { DataType::U32, DataType::U32, DataType::U32, DataType::U32, false }, // no integral types + { DataType::U16, DataType::U16, DataType::U16, DataType::U16, false }, // no integral types + { DataType::U8, DataType::U8, DataType::U8, DataType::U8, false }, // no integral types + { DataType::QASYMM8, DataType::QASYMM8, DataType::F32, DataType::QASYMM8, false } // Only S32 bias is supported + }; + + // It's enough to test a single shape and block size configuration while checking data types + const TensorShape shape = TensorShape(48U, 48U); + const TensorShape bia_shape = TensorShape(48U); + const MatMulKernelInfo matmul_kernel_info{ false, false, 1, 1, 4, false }; + for(auto &tuple : data_type_configurations) + { + const bool expected = (std::get<4>(tuple) && arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())); + + const TensorInfo lhs_info(shape, 1, std::get<0>(tuple)); + const TensorInfo rhs_info(shape, 1, std::get<1>(tuple)); + const TensorInfo bia_info(bia_shape, 1, std::get<2>(tuple)); + TensorInfo output_info(shape, 1, std::get<3>(tuple)); + + Status status = ClMatMulLowpNativeMMULKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } +} + +TEST_SUITE_END() // Validate + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8_SIGNED) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulLowpNativeMMULKernelFixture<int8_t>, + framework::DatasetMode::ALL, + combine(datasets::SmallMatMulLowpMMULDataset(), + make("TransposeA", { false, true }), + make("TransposeB", { false, true }), + m0_values_precommit, + n0_values_precommit, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8_SIGNED))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +FIXTURE_DATA_TEST_CASE(RunWithBias, CLMatMulLowpNativeMMULKernelWithBiasFixture<int8_t>, + framework::DatasetMode::ALL, + combine(datasets::SmallMatMulLowpMMULWithBiasDataset(), + make("TransposeA", { false, true }), + make("TransposeB", { false, true }), + m0_values_precommit, + n0_values_precommit, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8_SIGNED))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +FIXTURE_DATA_TEST_CASE(RunLargeLhsNotTransposed, CLMatMulLowpNativeMMULKernelFixture<int8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulLowpMMULDataset(), + make("TransposeA", { false }), + make("TransposeB", { false, true }), + m0_values_nightly_lhs_nt, + n0_values_nightly, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8_SIGNED))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulLowpNativeMMULKernelFixture<int8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulLowpMMULDataset(), + make("TransposeA", { true }), + make("TransposeB", { false, true }), + m0_values_nightly_lhs_t, + n0_values_nightly, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8_SIGNED))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +// Running High Dimensional test is enough for qasymm8_signed, because we're stressing the number of dimensions, not data type or M0/N0/K0 +// It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels +FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulLowpNativeMMULKernelFixture<int8_t>, + framework::DatasetMode::ALL, + combine(datasets::HighDimensionalMatMulLowpMMULDataset(), + make("TransposeA", { false, true }), + make("TransposeB", { false, true }), + make("M0", { 2 }), + make("N0", { 2 }), + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8_SIGNED))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE(QASYMM8) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulLowpNativeMMULKernelFixture<uint8_t>, + framework::DatasetMode::ALL, + combine(datasets::SmallMatMulLowpMMULDatasetSubset(), + make("TransposeA", { false, true }), + make("TransposeB", { false, true }), + m0_values_precommit, + n0_values_precommit, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +FIXTURE_DATA_TEST_CASE(RunWithBias, CLMatMulLowpNativeMMULKernelWithBiasFixture<uint8_t>, + framework::DatasetMode::ALL, + combine(datasets::SmallMatMulLowpMMULWithBiasDataset(), + make("TransposeA", { false, true }), + make("TransposeB", { false, true }), + m0_values_precommit, + n0_values_precommit, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +FIXTURE_DATA_TEST_CASE(RunLargeLhsNotTransposed, CLMatMulLowpNativeMMULKernelFixture<uint8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulLowpMMULDataset(), + make("TransposeA", { false }), + make("TransposeB", { false, true }), + m0_values_nightly_lhs_nt, + n0_values_nightly, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulLowpNativeMMULKernelFixture<uint8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulLowpMMULDataset(), + make("TransposeA", { true }), + make("TransposeB", { false, true }), + m0_values_nightly_lhs_t, + n0_values_nightly, + make("K0", { 4 }), + make("ExportRhsToCLImage", { false }), + make("DataType", DataType::QASYMM8))) +{ + if(_device_supports_mmul) + { + // Validate output + validate(CLAccessor(_target), _reference, tolerance_quant); + } +} + +TEST_SUITE_END() // QASYMM8 +TEST_SUITE_END() // Quantized +TEST_SUITE_END() // MatMulLowpNativeMMULKernel +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/MatMulNativeMMULKernel.cpp b/tests/validation/CL/MatMulNativeMMULKernel.cpp new file mode 100644 index 0000000000..655dd354dc --- /dev/null +++ b/tests/validation/CL/MatMulNativeMMULKernel.cpp @@ -0,0 +1,501 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/runtime/CL/CLTensor.h" +#include "src/gpu/cl/kernels/ClMatMulNativeMMULKernel.h" +#include "tests/datasets/LargeMatMulMMULDataset.h" +#include "tests/datasets/SmallMatMulMMULDataset.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/MatMulKernelFixture.h" +#include "tests/validation/reference/Permute.h" + +#include <tuple> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +constexpr float abs_tolerance_f32( + 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for floating point data types in case using relative tolerance fails because of small values */ +constexpr float abs_tolerance_f16( + 0.02f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data types in case using relative tolerance fails because of small values */ +RelativeTolerance<half_float::half> tolerance_f16(half(0.02)); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +} // namespace + +/** M0 values to test --precommit*/ +const auto m0_values_precommit = framework::dataset::make("M0", { 1, 3 }); + +/** N0 values to test --precommit*/ +const auto n0_values_precommit = framework::dataset::make("N0", { 2, 4 }); + +/** M0 values to test --nightly*/ +const auto m0_values_nightly_lhs_nt = framework::dataset::make("M0", { 1, 2, 3, 4, 5, 6, 7, 8 }); +const auto m0_values_nightly_lhs_t = framework::dataset::make("M0", { 1, 2, 3, 4, 8 }); + +/** N0 values to test --nightly*/ +const auto n0_values_nightly_rhs_nt = framework::dataset::make("N0", { 1, 2, 3, 4, 8, 16 }); +const auto n0_values_nightly_rhs_t = framework::dataset::make("N0", { 1, 2, 3, 4, 8 }); + +/** K0 value -- Fixed to 1 */ +const auto k0_value = framework::dataset::make("K0", { 1 }); + +template <typename T> +using CLMatMulNativeMMULKernelFixture = MatMulKernelValidationFixture<T, ClMatMulNativeMMULKernel, true /*use_mmul*/>; + +template <typename T> +using CLMatMulKernelBiasFixture = MatMulKernelWithBiasValidation<T, ClMatMulNativeMMULKernel, true /*use_mmul*/>; + +TEST_SUITE(CL) +TEST_SUITE(MatMulNativeMMULKernel) +TEST_SUITE(Validate) + +TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL) +{ + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + using MatMulConfigurationPair = std::pair<MatMulKernelInfo, bool>; + + const std::vector<MatMulConfigurationPair> supported_block_sizes = + { + // MatMulKernelInfo(adj_lhs, adj_rhs, M0, N0, K0, export_rhs_to_cl_image = false) + // Lhs not-transposed, Rhs not-transposed + { MatMulKernelInfo(false, false, 0, 1, 1), false }, // M0 should be > 0 + { MatMulKernelInfo(false, false, 3, 5, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 6, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(false, false, 3, 3, 4), false }, // K0 not 1 + { MatMulKernelInfo(false, false, 9, 1, 1), true }, + { MatMulKernelInfo(false, false, 3, 16, 1), true }, + { MatMulKernelInfo(false, false, 7, 3, 1), true }, + + // Lhs transposed, Rhs not-transposed + { MatMulKernelInfo(true, false, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 6, 3, 1), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 5, 3, 1), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, false, 2, 2, 2), false }, // K0 is not 1 + { MatMulKernelInfo(true, false, 4, 1, 1), true }, + { MatMulKernelInfo(true, false, 3, 3, 1), true }, + { MatMulKernelInfo(true, false, 2, 4, 1), true }, + + // Lhs not-transposed, Rhs not-transposed + { MatMulKernelInfo(false, true, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8} + { MatMulKernelInfo(false, true, 2, 17, 1), false }, // N0 not in {1, 2, 3, 4, 8} + { MatMulKernelInfo(false, true, 4, 5, 1), false }, // N0 not in {1, 2, 3, 4, 8} + { MatMulKernelInfo(false, true, 4, 4, 7), false }, // K0 is not 1 + { MatMulKernelInfo(false, true, 4, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8} + { MatMulKernelInfo(false, true, 3, 8, 1), true }, + { MatMulKernelInfo(false, true, 8, 16, 1), true }, + { MatMulKernelInfo(false, true, 2, 4, 1), true }, + + // Lhs transposed, Rhs transposed + { MatMulKernelInfo(true, true, 3, 11, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 3, 7, 1), false }, // N0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 6, 3, 1), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 5, 3, 1), false }, // M0 not in {1, 2, 3, 4, 8, 16} + { MatMulKernelInfo(true, true, 4, 8, 2), false }, // K0 is not 1 + { MatMulKernelInfo(true, true, 4, 8, 1), true }, + { MatMulKernelInfo(true, true, 3, 3, 1), true }, + { MatMulKernelInfo(true, true, 16, 4, 1), true }, + }; + + // Set big enough shapes so that block sizes are not truncated. Also, set all dimensions equal + // so that it doesn't fail for different NT/T configurations. We aim to test the block sizes here, + // not the shapes themselves. + const TensorInfo lhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(TensorShape(100U, 100U), 1, DataType::F32); + + for(auto &pair : supported_block_sizes) + { + TensorInfo output_info; + Status status = ClMatMulNativeMMULKernel::validate(&lhs_info, &rhs_info, nullptr, &output_info, pair.first); + ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS); + } + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +TEST_CASE(ValidateInputShapes, framework::DatasetMode::ALL) +{ + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, TensorShape, bool>; // lhs, rhs, bias, result + const std::vector<ShapeConfigurationTuple> shape_configurations = + { + { TensorShape(4U, 1U), TensorShape(3U, 4U), TensorShape(3U), true }, + { TensorShape(12U, 12U), TensorShape(3U, 12U), TensorShape(3U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 8U), TensorShape(2U), true }, + { TensorShape(8U, 4U), TensorShape(2U, 4U), TensorShape(2U), false }, // Mismatch in the K dimension + { TensorShape(5U, 0U), TensorShape(2U, 5U), TensorShape(2U), false }, // Invalid dimension + { TensorShape(5U, 7U), TensorShape(2U, 5U), TensorShape(2U), false }, // K not a multiple of 4 (MMUL_K0) + { TensorShape(8U, 4U, 3U, 4U, 5U, 6U), TensorShape(2U, 8U, 3U, 4U, 5U, 6U), TensorShape(2U), true }, + { TensorShape(5U, 4U, 3U, 4U, 5U, 1U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // No batch broadcasting + { TensorShape(5U, 4U, 3U, 4U, 9U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), TensorShape(2U), false }, // Mismatch in batch dimension + { TensorShape(4U, 1U), TensorShape(3U, 4U), TensorShape(1U), false }, // Bias first dimensions != dst first dimension. + { TensorShape(4U, 1U), TensorShape(3U, 4U), TensorShape(5U, 6U), false }, // Bias is 2d which is invalid. + }; + + for(auto &tuple : shape_configurations) + { + const bool expected = std::get<3>(tuple); + + for(bool adj_lhs : + { + false, true + }) + { + for(bool adj_rhs : + { + false, true + }) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + TensorShape bia_shape = std::get<2>(tuple); + + if(adj_lhs) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + + if(adj_rhs) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + const TensorInfo lhs_info = TensorInfo(lhs_shape, 1, DataType::F32); + const TensorInfo rhs_info = TensorInfo(rhs_shape, 1, DataType::F32); + const TensorInfo bia_info = TensorInfo(bia_shape, 1, DataType::F32); + TensorInfo output_info; + + MatMulKernelInfo matmul_kernel_info{ adj_lhs, adj_rhs, 1, 1, 1, false /* export_rhs_to_cl_image */ }; + + Status status = ClMatMulNativeMMULKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + } + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +TEST_CASE(ValidateDataTypes, framework::DatasetMode::ALL) +{ + if(arm_matrix_multiply_supported(CLKernelLibrary::get().get_device())) + { + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using DataTypeConfigurationTuple = std::tuple<DataType, DataType, DataType, DataType, bool>; + const std::vector<DataTypeConfigurationTuple> data_type_configurations = + { + { DataType::F32, DataType::F32, DataType::F32, DataType::F32, true }, + { DataType::F16, DataType::F16, DataType::F16, DataType::F16, true }, + { DataType::F32, DataType::F32, DataType::F32, DataType::F32, true }, + { DataType::F32, DataType::F32, DataType::F16, DataType::F32, false }, // incorrect bias type + { DataType::F16, DataType::F32, DataType::F32, DataType::F32, false }, // no mixed precision + { DataType::F64, DataType::F64, DataType::F64, DataType::F64, false }, // no double precision + { DataType::QASYMM8, DataType::QASYMM8, DataType::S32, DataType::QASYMM8, false }, // no quantized types + { DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, DataType::S32, DataType::QASYMM8_SIGNED, false }, // no quantized types + { DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, DataType::S32, DataType::QSYMM8_PER_CHANNEL, false }, // no quantized types + { DataType::QASYMM16, DataType::QASYMM16, DataType::S32, DataType::QASYMM16, false }, // no quantized types + { DataType::QSYMM16, DataType::QSYMM16, DataType::S32, DataType::QSYMM16, false }, // no quantized types + { DataType::QSYMM8, DataType::QSYMM8, DataType::S32, DataType::QSYMM8, false }, // no quantized types + { DataType::S64, DataType::S64, DataType::S64, DataType::S64, false }, // no integral types + { DataType::S32, DataType::S32, DataType::S32, DataType::S32, false }, // no integral types + { DataType::S16, DataType::S16, DataType::S16, DataType::S16, false }, // no integral types + { DataType::S8, DataType::S8, DataType::S8, DataType::S8, false }, // no integral types + { DataType::U64, DataType::U64, DataType::U64, DataType::U64, false }, // no integral types + { DataType::U32, DataType::U32, DataType::U32, DataType::U32, false }, // no integral types + { DataType::U16, DataType::U16, DataType::U16, DataType::U16, false }, // no integral types + { DataType::U8, DataType::U8, DataType::U8, DataType::U8, false }, // no integral types + }; + + const TensorShape shape = TensorShape(8U, 8U); + const TensorShape bia_shape = TensorShape(8U); + const MatMulKernelInfo matmul_kernel_info{ false, false, 1, 1, 1, false }; + for(auto &tuple : data_type_configurations) + { + const bool expected = std::get<4>(tuple); + + const TensorInfo lhs_info(shape, 1, std::get<0>(tuple)); + const TensorInfo rhs_info(shape, 1, std::get<1>(tuple)); + const TensorInfo bia_info(bia_shape, 1, std::get<2>(tuple)); + TensorInfo output_info(shape, 1, std::get<3>(tuple)); + + Status status = ClMatMulNativeMMULKernel::validate(&lhs_info, &rhs_info, &bia_info, &output_info, matmul_kernel_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + else + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + } +} + +TEST_SUITE_END() // Validate + +TEST_SUITE(Float) +TEST_SUITE(FP32) +TEST_SUITE(Buffer) +FIXTURE_DATA_TEST_CASE(RunTiny, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::TinyMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunWithBias, CLMatMulKernelBiasFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} + +FIXTURE_DATA_TEST_CASE(RunLargeRhsTranspose, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +// Running High Dimensional test is enough for FP32, because we're stressing the number of dimensions, not data type or M0/N0/K0 +// It's a good idea to test for each Lhs/Rhs T/NT combinations because they're different CL kernels +FIXTURE_DATA_TEST_CASE(RunHighDimensional, CLMatMulNativeMMULKernelFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::HighDimensionalMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + framework::dataset::make("M0", { 2 })), + framework::dataset::make("N0", { 2 })), + framework::dataset::make("K0", { 1 })), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); + } +} +TEST_SUITE_END() // Buffer + +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +TEST_SUITE(Buffer) +FIXTURE_DATA_TEST_CASE(RunSmall, CLMatMulNativeMMULKernelFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false, true })), + framework::dataset::make("TransposeB", { false, true })), + m0_values_precommit), + n0_values_precommit), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeNoTranspose, CLMatMulNativeMMULKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_nt), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeRhsTranspose, CLMatMulNativeMMULKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { false })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_nt), + n0_values_nightly_rhs_t), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposed, CLMatMulNativeMMULKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { false })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_nt), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +FIXTURE_DATA_TEST_CASE(RunLargeLhsTransposedRhsTransposed, CLMatMulNativeMMULKernelFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(combine(datasets::LargeMatMulMMULDataset(), + framework::dataset::make("TransposeA", { true })), + framework::dataset::make("TransposeB", { true })), + m0_values_nightly_lhs_t), + n0_values_nightly_rhs_t), + k0_value), + framework::dataset::make("ExportRhsToCLImage", { false })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + // Validate output + if(_device_supports_mmul) + { + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + } +} +TEST_SUITE_END() // Buffer + +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float +TEST_SUITE_END() // MatMulNativeMMULKernel +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/MaxUnpoolingLayer.cpp b/tests/validation/CL/MaxUnpoolingLayer.cpp index 6cba8b8bd5..cf4fcdda70 100644 --- a/tests/validation/CL/MaxUnpoolingLayer.cpp +++ b/tests/validation/CL/MaxUnpoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020 Arm Limited. + * Copyright (c) 2020-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -51,20 +51,19 @@ const auto PoolingLayerIndicesDatasetFPSmall = combine(combine(framework::datase TEST_SUITE(Float) TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(MaxUnpooling, CLMaxUnpoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerIndicesDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(MaxUnpooling, CLMaxUnpoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerIndicesDatasetFPSmall, framework::dataset::make("DataType", DataType::F32))), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }) )) { - printf("validate\n"); // Validate output validate(CLAccessor(_target), _reference); } TEST_SUITE_END() // FP32 TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(MaxUnpooling, CLMaxUnpoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerIndicesDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(MaxUnpooling, CLMaxUnpoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerIndicesDatasetFPSmall, framework::dataset::make("DataType", DataType::F16))), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }) diff --git a/tests/validation/CL/PReluLayer.cpp b/tests/validation/CL/PReluLayer.cpp index 043262d891..f3f1c8b1b8 100644 --- a/tests/validation/CL/PReluLayer.cpp +++ b/tests/validation/CL/PReluLayer.cpp @@ -56,7 +56,7 @@ const auto PReluLayerQASYMM8Dataset = combine(combine(framework::dataset::make(" const auto PReluLayerQASYMM8SIGNEDDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)); -const auto PReluLayerS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), +const auto PReluLayerS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), framework::dataset::make("DataType", DataType::S16)); const auto PReluLayerFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataType", DataType::F16)); @@ -71,21 +71,18 @@ TEST_SUITE(PReluLayer) // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes }), framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false})), + framework::dataset::make("Expected", { true, false, false})), input1_info, input2_info, output_info, expected) { ARM_COMPUTE_EXPECT(bool(CLPReluLayer::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS); diff --git a/tests/validation/CL/PadLayer.cpp b/tests/validation/CL/PadLayer.cpp index 370195b078..ea0cb32785 100644 --- a/tests/validation/CL/PadLayer.cpp +++ b/tests/validation/CL/PadLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,10 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ +#include "arm_compute/graph/Utils.h" #include "arm_compute/runtime/CL/CLScheduler.h" #include "arm_compute/runtime/CL/functions/CLPadLayer.h" +#include "src/graph/mutators/MutatorUtils.h" #include "tests/CL/CLAccessor.h" #include "tests/Globals.h" #include "tests/datasets/ShapeDatasets.h" @@ -110,6 +112,63 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( ARM_COMPUTE_EXPECT(bool(CLPadLayer::validate(&input_info.clone()->set_is_resizable(true), &output_info.clone()->set_is_resizable(true), padding, PixelValue(), mode)) == expected, framework::LogLevel::ERRORS); } +DATA_TEST_CASE(CheckFusingWithConvolution, framework::DatasetMode::ALL, zip(zip( + framework::dataset::make("DataLayout", { DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NCHW, + DataLayout::NHWC, + DataLayout::NHWC, + DataLayout::NHWC, + DataLayout::NHWC, + DataLayout::NHWC, + DataLayout::NHWC, + DataLayout::NHWC, + DataLayout::UNKNOWN + }), + framework::dataset::make("PaddingList", { PaddingList({{0, 0}, {1, 1}, {1, 1}}), // nchw + PaddingList({{1, 1}, {1, 1}, {0, 0}, {0, 0}}), + PaddingList({{1, 1}, {1, 1}}), + PaddingList({}), + PaddingList({{0, 0}}), + PaddingList({{0, 0}, {0, 0}, {0, 0}, {0, 0}}), + PaddingList({{0, 0}, {0, 0}, {0, 0}, {1, 0}}), + PaddingList({{0, 1}}), + PaddingList({{0, 0}, {1, 1}, {1, 1}}), // nhwc + PaddingList({{0, 0}, {0, 0}, {1, 1}, {1, 1}}), + PaddingList({{0, 0}, {1, 0}, {1, 1}, {0, 0}}), + PaddingList({}), + PaddingList({{0, 0}}), + PaddingList({{0, 1}}), + PaddingList({{0, 0}, {1, 1}}), + PaddingList({{0, 0}}) + })), // unknown + framework::dataset::make("Expected", { false, // nchw + true, + true, + true, + true, + true, + false, + true, + true, // nhwc + false, + true, + true, + true, + false, + true, + false // unknown + })), + data_layout, padding_list, expected) +{ + ARM_COMPUTE_EXPECT(expected == arm_compute::graph::is_padding_in_height_or_width(data_layout, padding_list), framework::LogLevel::ERRORS); +} + // clang-format on // *INDENT-ON* diff --git a/tests/validation/CL/PixelWiseMultiplication.cpp b/tests/validation/CL/PixelWiseMultiplication.cpp index 9e0a6243d7..62ff15a37f 100644 --- a/tests/validation/CL/PixelWiseMultiplication.cpp +++ b/tests/validation/CL/PixelWiseMultiplication.cpp @@ -36,6 +36,9 @@ namespace test { namespace validation { +/** Synced with tests/validation/dynamic_fusion/gpu/cl/Mul.cpp from the dynamic fusion interface. + * Please check there for any differences in the coverage + */ namespace { namespace @@ -50,9 +53,6 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.75f, 0.25f), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.75f, 0.25f) }); -// Since in-place computation on CL-side hasn't been intended to be implemented, they are not tested. -// However, this dataset is required for the shared fixture and it would make extension easier when -// CL-side also starts supporting in-place computation. const auto InPlaceDataSet = framework::dataset::make("InPlace", { false }); } //namespace // *INDENT-OFF* @@ -124,7 +124,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLPixelWiseMultiplicationIntegerFixture<int>, f datasets::SmallShapes(), framework::dataset::make("DataType1", DataType::S32)), framework::dataset::make("DataType2", DataType::S32)), - framework::dataset::make("Scale", {1.f})), + framework::dataset::make("Scale", { 1.f })), datasets::ConvertPolicies()), framework::dataset::make("RoundingPolicy", RoundingPolicy::TO_NEAREST_UP)), EmptyActivationFunctionsDataset), @@ -132,6 +132,18 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLPixelWiseMultiplicationIntegerFixture<int>, f { validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunInplace, CLPixelWiseMultiplicationIntegerFixture<int>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(combine(datasets::TinyShapes(), + framework::dataset::make("DataType1", DataType::S32)), + framework::dataset::make("DataType2", DataType::S32)), + framework::dataset::make("Scale", { 1.f })), + datasets::ConvertPolicies()), + framework::dataset::make("RoundingPolicy", RoundingPolicy::TO_NEAREST_UP)), + EmptyActivationFunctionsDataset), + framework::dataset::make("InPlace", { true }))) +{ + validate(CLAccessor(_target), _reference); +} TEST_SUITE_END() TEST_SUITE(F16toF16) @@ -147,6 +159,19 @@ TEST_SUITE(F32toF32) TEST_SUITE(Scale255) PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToF32Fixture<float>, PRECOMMIT, SmallShapes(), F32, F32, scale_255, TO_NEAREST_UP, EmptyActivationFunctionsDataset, VALIDATE(float, 1.f)) PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunWithActivation, ToF32Fixture<float>, ALL, TinyShapes(), F32, F32, scale_255, TO_NEAREST_UP, ActivationFunctionsDataset, VALIDATE(float, 1.f)) +FIXTURE_DATA_TEST_CASE(RunInplace, CLPixelWiseMultiplicationToF32Fixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(combine(datasets::TinyShapes(), + framework::dataset::make("DataTypeIn1", DataType::F32)), + framework::dataset::make("DataTypeIn2", DataType::F32)), + framework::dataset::make("Scale", { scale_255 })), + datasets::ConvertPolicies()), + framework::dataset::make("RoundingPolicy", RoundingPolicy::TO_NEAREST_UP)), + EmptyActivationFunctionsDataset), + framework::dataset::make("InPlace", { true }))) +{ + // Validate output + VALIDATE(float, 1.f) +} TEST_SUITE_END() // Scale255 TEST_SUITE_END() // F32toF32 @@ -200,6 +225,23 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLPixelWiseMultiplicationQuantizedBroa validate(CLAccessor(_target), _reference, tolerance_qasymm8); } +FIXTURE_DATA_TEST_CASE(RunInplace, CLPixelWiseMultiplicationQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(datasets::TinyShapesBroadcastInplace(), + framework::dataset::make("DataTypeIn1", DataType::QASYMM8)), + framework::dataset::make("DataTypeIn2", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", DataType::QASYMM8)), + framework::dataset::make("Scale", { 1.f, 2.f })), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("RoundingPolicy", RoundingPolicy::TO_NEAREST_EVEN)), + framework::dataset::make("Src0QInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("OUtQInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("InPlace", { true }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) diff --git a/tests/validation/CL/Pooling3dLayer.cpp b/tests/validation/CL/Pooling3dLayer.cpp new file mode 100644 index 0000000000..84d630e6cf --- /dev/null +++ b/tests/validation/CL/Pooling3dLayer.cpp @@ -0,0 +1,345 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/TensorShape.h" +#include "tests/framework/datasets/Datasets.h" + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLPooling3dLayer.h" +#include "tests/CL/CLAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/Pooling3dLayerDataset.h" +#include "tests/datasets/PoolingTypesDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/Pooling3dLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +/** Input data sets for floating-point data types */ +const auto Pooling3dLayerDatasetFP = combine(combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size3D(2, 3, 2) })), + framework::dataset::make("Stride", { Size3D(1, 1, 1), Size3D(2, 1, 1), Size3D(1, 2, 1), Size3D(2, 2, 1) })), + framework::dataset::make("Padding", { Padding3D(0, 1, 0), Padding3D(1, 1, 1) })), + framework::dataset::make("ExcludePadding", { true, false })); + +const auto Pooling3dLayerDatasetFPSmall = combine(combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size3D(2, 2, 2), Size3D(3, 3, 3) })), + framework::dataset::make("Stride", { Size3D(2, 2, 2), Size3D(2, 1, 1) })), + framework::dataset::make("Padding", { Padding3D(0, 0, 0), Padding3D(1, 1, 1), Padding3D(1, 0, 0) })), + framework::dataset::make("ExcludePadding", { true, false })); + +const auto Pooling3DLayerDatasetQuantized = combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), + framework::dataset::make("PoolingSize", { Size3D(2, 3, 2) })), + framework::dataset::make("Stride", { Size3D(1, 1, 1), Size3D(2, 1, 1), Size3D(1, 2, 1), Size3D(1, 1, 2), Size3D(2, 2, 1)})), + framework::dataset::make("Padding", { Padding3D(0, 0, 0), Padding3D(1, 1, 1), Padding3D(1, 0, 0) })), + framework::dataset::make("ExcludePadding", { true })); + +using ShapeDataset = framework::dataset::ContainerDataset<std::vector<TensorShape>>; + +constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */ +constexpr AbsoluteTolerance<float> tolerance_f16(0.1f); /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */ +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); /**< Tolerance value for comparing reference's output against implementation's output for QASYMM8_SIGNED integer datatype*/ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric type */ + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(Pooling3dLayer) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Mismatching data type + TensorInfo(TensorShape(2U, 27U, 13U, 4U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid pad/size combination + TensorInfo(TensorShape(2U, 27U, 13U, 4U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid pad/size combination + TensorInfo(TensorShape(2U, 27U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Invalid output shape + TensorInfo(TensorShape(5U, 13U, 15U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Global Pooling + TensorInfo(TensorShape(13U,13U, 5U, 1U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid output Global Pooling + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 4U), 1, DataType::F32, DataLayout::NDHWC), // Invalid data type + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 5U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(1U, 16U, 1U, 3U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), + }), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(2U, 25U, 11U, 3U, 3U), 1, DataType::F16, DataLayout::NDHWC), + TensorInfo(TensorShape(2U, 30U, 11U, 3U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(2U, 25U, 16U, 3U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(2U, 27U, 13U, 3U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 1U, 1U, 1U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Global pooling applied + TensorInfo(TensorShape(5U, 2U, 2U, 2U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid output Global Pooling + TensorInfo(TensorShape(5U, 12U, 12U, 3U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 12U, 12U, 3U, 4U), 1, DataType::QASYMM8, DataLayout::NDHWC), // Invalid data type + TensorInfo(TensorShape(5U, 1U, 1U, 1U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(1U, 15U, 1U, 2U, 4U), 1, DataType::F32, DataLayout::NDHWC), // Output width larger than input + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), + })), + framework::dataset::make("PoolInfo", { Pooling3dLayerInfo(PoolingType::AVG, 3, Size3D(1, 1, 1), Padding3D(0, 0, 0)), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1, 1, 1), Padding3D(2, 0, 0)), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1, 1, 1), Padding3D(0, 0, 0)), + Pooling3dLayerInfo(PoolingType::L2, 3, Size3D(1, 1, 1), Padding3D(0, 0, 0)), + Pooling3dLayerInfo(PoolingType::AVG), + Pooling3dLayerInfo(PoolingType::MAX), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1U, 1U, 1U), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG), + Pooling3dLayerInfo(PoolingType::MAX, 2, Size3D(1, 1, 2), Padding3D(0, 0, 0), false), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(2U, 2U, 2U), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG, 1, Size3D(2U, 2U, 2U), Padding3D(2, 2, 2), true), // Pool size is smaller than the padding size with padding excluded + Pooling3dLayerInfo(PoolingType::AVG, 1, Size3D(2U, 2U, 2U), Padding3D(2, 2, 2), false), // Pool size is smaller than the padding size with padding included + Pooling3dLayerInfo(PoolingType::AVG, 3, Size3D(2U, 2U, 2U), Padding3D(2,1,2,2,1,2), false, false, DimensionRoundingType::CEIL), // CEIL with asymmetric Padding + })), + framework::dataset::make("Expected", { false, false, false, false, true, false, false, false, true , false, true, false, false, false})), + input_info, output_info, pool_info, expected) +{ + ARM_COMPUTE_EXPECT(bool(CLPooling3dLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pool_info)) == expected, framework::LogLevel::ERRORS); +} + + +template <typename T> +using CLPooling3dLayerFixture = Pooling3dLayerValidationFixture<CLTensor, CLAccessor, CLPooling3dLayer, T>; + +template <typename T> +using CLSpecialPooling3dLayerFixture = SpecialPooling3dLayerValidationFixture<CLTensor, CLAccessor, CLPooling3dLayer, T>; + +template <typename T> +using CLPooling3dLayerGlobalFixture = Pooling3dLayerGlobalValidationFixture<CLTensor, CLAccessor, CLPooling3dLayer, T>; + +template <typename T> +using CLPooling3dLayerQuantizedFixture = Pooling3dLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLPooling3dLayer, T>; + +// clang-format on +// *INDENT-ON* +TEST_SUITE(QUANTIZED) + +TEST_SUITE(QASYMM8) +// Small Dataset Quantized Dataset +FIXTURE_DATA_TEST_CASE(RunSmall, CLPooling3dLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small5dShapes(), + combine(Pooling3DLayerDatasetQuantized, + framework::dataset::make("DataType", DataType::QASYMM8))), + framework::dataset::make("InputQuantInfo", { QuantizationInfo(1.f / 127.f, 10), QuantizationInfo(1.f / 127.f, 10) })), + framework::dataset::make("OutputQuantInfo", { QuantizationInfo(1.f / 127.f, 5), QuantizationInfo(1.f / 127.f, 10) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} + +// Large Dataset Quantized Dataset +FIXTURE_DATA_TEST_CASE(RunLarge, CLPooling3dLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::Large5dShapes(), + combine(Pooling3DLayerDatasetQuantized, + framework::dataset::make("DataType", DataType::QASYMM8))), + framework::dataset::make("InputQuantInfo", { QuantizationInfo(1.f / 127.f, 10), QuantizationInfo(1.f / 127.f, 10) })), + framework::dataset::make("OutputQuantInfo", { QuantizationInfo(1.f / 127.f, 5), QuantizationInfo(1.f / 127.f, 10) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() + +TEST_SUITE(QASYMM8_SIGNED) + +// Large Dataset Quantized Dataset Signed +FIXTURE_DATA_TEST_CASE(RunSmall, CLPooling3dLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small5dShapes(), + combine(Pooling3DLayerDatasetQuantized, + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))), + framework::dataset::make("InputQuantInfo", { QuantizationInfo(1.f / 127.f, -10), QuantizationInfo(1.f / 127.f, -10) })), + framework::dataset::make("OutputQuantInfo", { QuantizationInfo(1.f / 127.f, -5), QuantizationInfo(1.f / 127.f, -10) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8_signed); +} + +// Large Dataset Quantized pooling test +FIXTURE_DATA_TEST_CASE(RunLarge, CLPooling3dLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::Large5dShapes(), + combine(Pooling3DLayerDatasetQuantized, + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))), + framework::dataset::make("InputQuantInfo", { QuantizationInfo(1.f / 127.f, -10), QuantizationInfo(1.f / 127.f, -10) })), + framework::dataset::make("OutputQuantInfo", { QuantizationInfo(1.f / 127.f, -5), QuantizationInfo(1.f / 127.f, -10) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8_signed); +} + +TEST_SUITE_END() +TEST_SUITE_END() + +TEST_SUITE(Float) +TEST_SUITE(FP32) + +FIXTURE_DATA_TEST_CASE(RunSpecial, CLSpecialPooling3dLayerFixture<float>, framework::DatasetMode::ALL, datasets::Pooling3dLayerDatasetSpecial() * framework::dataset::make("DataType", DataType::F32)) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmall, CLPooling3dLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small5dShapes(), combine(Pooling3dLayerDatasetFPSmall, + framework::dataset::make("DataType", DataType::F32)))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLPooling3dLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::Large5dShapes(), combine(Pooling3dLayerDatasetFP, + framework::dataset::make("DataType", DataType::F32)))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE(GlobalPooling) +// *INDENT-OFF* +// clang-format off +FIXTURE_DATA_TEST_CASE(RunSmall, CLPooling3dLayerFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(3U, 27U, 13U, 4U), + TensorShape(4U, 27U, 13U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(27, 13, 4) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", false)), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmallGlobal, CLPooling3dLayerGlobalFixture<float>, framework::DatasetMode::ALL, + combine(combine( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 4U, 3U), + TensorShape(27U, 13U, 4U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLPooling3dLayerFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(4U, 79U, 37U, 11U), + TensorShape(4U, 79U, 37U, 11U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(79, 37, 11) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", false)), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +// clang-format on +// *INDENT-ON* +TEST_SUITE_END() // GlobalPooling +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLPooling3dLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small5x5Shapes(), combine(Pooling3dLayerDatasetFPSmall, + framework::dataset::make("DataType", DataType::F16)))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLPooling3dLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::Large5dShapes(), combine(Pooling3dLayerDatasetFP, + framework::dataset::make("DataType", DataType::F16)))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE(GlobalPooling) +// *INDENT-OFF* +// clang-format off +FIXTURE_DATA_TEST_CASE(RunSmall, CLPooling3dLayerFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(3U, 27U, 13U, 4U), + TensorShape(4U, 27U, 13U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(27, 13, 4) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", false)), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmallGlobal, CLPooling3dLayerGlobalFixture<half>, framework::DatasetMode::ALL, + combine(combine( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 4U, 3U), + TensorShape(27U, 13U, 4U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLPooling3dLayerFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(4U, 79U, 37U, 11U), + TensorShape(4U, 79U, 37U, 11U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(79, 37, 11) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", false)), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +// clang-format on +// *INDENT-ON* +TEST_SUITE_END() // GlobalPooling +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float +TEST_SUITE_END() // Pooling3dLayer +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/PoolingLayer.cpp b/tests/validation/CL/PoolingLayer.cpp index 63dec3910f..9fe28c7acf 100644 --- a/tests/validation/CL/PoolingLayer.cpp +++ b/tests/validation/CL/PoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -153,7 +153,7 @@ FIXTURE_DATA_TEST_CASE(RunSpecial, CLSpecialPoolingLayerFixture<float>, framewor // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetFPSmall, framework::dataset::make("DataType", DataType::F32))), pool_data_layout_dataset)) @@ -161,7 +161,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixture<float>, framework::Datase // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(2, 1, 0, 0) })), @@ -181,10 +181,11 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLPoolingLayerFixture<float>, framework::Datase validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunSmallIndices, CLPoolingLayerIndicesFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFPIndicesSmall, - framework::dataset::make("DataType", - DataType::F32))), - pool_data_layout_dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallIndices, CLPoolingLayerIndicesFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallNoneUnitShapes(), + combine(PoolingLayerDatasetFPIndicesSmall, + framework::dataset::make("DataType", + DataType::F32))), + pool_data_layout_dataset),framework::dataset::make("UseKernelIndices", { false }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); @@ -232,8 +233,9 @@ TEST_SUITE_END() // GlobalPooling TEST_SUITE_END() // FP32 TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLMixedPrecesionPoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFPSmall, - framework::dataset::make("DataType", DataType::F16))), +FIXTURE_DATA_TEST_CASE(RunSmall, CLMixedPrecesionPoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallNoneUnitShapes(), + combine(PoolingLayerDatasetFPSmall, + framework::dataset::make("DataType", DataType::F16))), pool_data_layout_dataset), pool_fp_mixed_precision_dataset)) { @@ -248,10 +250,11 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLMixedPrecesionPoolingLayerFixture<half>, fram // Validate output validate(CLAccessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunSmallIndices, CLPoolingLayerIndicesFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFPIndicesSmall, - framework::dataset::make("DataType", - DataType::F16))), - pool_data_layout_dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallIndices, CLPoolingLayerIndicesFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallNoneUnitShapes(), + combine(PoolingLayerDatasetFPIndicesSmall, + framework::dataset::make("DataType", + DataType::F16))), + pool_data_layout_dataset), framework::dataset::make("UseKernelIndices", { false }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); @@ -307,7 +310,7 @@ template <typename T> using CLPoolingLayerQuantizedMixedDataLayoutFixture = PoolingLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLPoolingLayer, T, true>; TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetQASYMM8Small, framework::dataset::make("DataType", DataType::QASYMM8))), pool_data_layout_dataset), @@ -317,7 +320,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<uint8_t>, framew // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 2, 1, 1) })), @@ -333,7 +336,7 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerQuantizedMixedDataLayou TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetQASYMM8Small, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))), pool_data_layout_dataset), @@ -343,7 +346,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerQuantizedFixture<int8_t>, framewo // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8_s); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLPoolingLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 2, 1, 1) })), diff --git a/tests/validation/CL/ReduceMean.cpp b/tests/validation/CL/ReduceMean.cpp index 947f84af49..8a8fa4aef0 100644 --- a/tests/validation/CL/ReduceMean.cpp +++ b/tests/validation/CL/ReduceMean.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,7 @@ constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value constexpr AbsoluteTolerance<float> tolerance_f16(0.03f); /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */ constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for 8-bit asymmetric quantized type */ -const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(0, 1, 2, 3) }), +const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(2, 3), Coordinates(0, 1, 2, 3) }), framework::dataset::make("KeepDims", { true })); const auto axis_drop = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1), Coordinates(3), Coordinates(1, 2), Coordinates(2, 1) }), framework::dataset::make("KeepDims", { false })); } // namespace diff --git a/tests/validation/CL/Remap.cpp b/tests/validation/CL/Remap.cpp deleted file mode 100644 index bbb3cecea9..0000000000 --- a/tests/validation/CL/Remap.cpp +++ /dev/null @@ -1,77 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/core/Types.h" -#include "arm_compute/runtime/CL/functions/CLRemap.h" -#include "arm_compute/runtime/Tensor.h" -#include "arm_compute/runtime/TensorAllocator.h" -#include "tests/CL/CLAccessor.h" -#include "tests/PaddingCalculator.h" -#include "tests/datasets/BorderModeDataset.h" -#include "tests/datasets/ShapeDatasets.h" -#include "tests/framework/Asserts.h" -#include "tests/framework/Macros.h" -#include "tests/framework/datasets/Datasets.h" -#include "tests/validation/Validation.h" -#include "tests/validation/fixtures/RemapFixture.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -namespace -{ -constexpr AbsoluteTolerance<uint8_t> tolerance_value(1); -} // namespace - -TEST_SUITE(CL) -TEST_SUITE(Remap) -template <typename T> -using CLRemapFixture = RemapValidationFixture<CLTensor, CLAccessor, CLRemap, T>; -template <typename T> -using CLRemapLayoutFixture = RemapValidationMixedLayoutFixture<CLTensor, CLAccessor, CLRemap, T>; - -FIXTURE_DATA_TEST_CASE(RunSmall, CLRemapLayoutFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("InterpolationPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR, InterpolationPolicy::BILINEAR })), - framework::dataset::make("DataType", DataType::U8)), - framework::dataset::make("BorderModes", { BorderMode::UNDEFINED, BorderMode::CONSTANT })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) -{ - // Validate output - validate(CLAccessor(_target), _reference, _valid_mask, tolerance_value); -} - -FIXTURE_DATA_TEST_CASE(RunLarge, CLRemapFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("InterpolationPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR, InterpolationPolicy::BILINEAR })), - framework::dataset::make("DataType", DataType::U8)), - framework::dataset::make("BorderModes", { BorderMode::UNDEFINED, BorderMode::CONSTANT }))) -{ - // Validate output - validate(CLAccessor(_target), _reference, _valid_mask, tolerance_value); -} - -TEST_SUITE_END() -TEST_SUITE_END() -} // namespace validation -} // namespace test -} // namespace arm_compute diff --git a/tests/validation/CL/Reverse.cpp b/tests/validation/CL/Reverse.cpp index 11df0e7803..82effc2136 100644 --- a/tests/validation/CL/Reverse.cpp +++ b/tests/validation/CL/Reverse.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -41,9 +41,10 @@ namespace test { namespace validation { +using framework::dataset::make; namespace { -auto run_small_dataset = combine(datasets::SmallShapes(), datasets::Tiny1DShapes()); +auto run_small_dataset = combine(datasets::Small3DShapes(), datasets::Tiny1DShapes()); auto run_large_dataset = combine(datasets::LargeShapes(), datasets::Tiny1DShapes()); } // namespace @@ -53,33 +54,34 @@ TEST_SUITE(Reverse) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), // Invalid axis datatype + make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), // Invalid axis datatype TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid axis shape TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid axis length (> 4) TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Mismatching shapes TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(2U), 1, DataType::U8), }), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), + make("OutputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(2U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(2U), 1, DataType::U8), })), - framework::dataset::make("AxisInfo",{ TensorInfo(TensorShape(3U), 1, DataType::U8), + make("AxisInfo",{ TensorInfo(TensorShape(3U), 1, DataType::U8), TensorInfo(TensorShape(2U, 10U), 1, DataType::U32), TensorInfo(TensorShape(8U), 1, DataType::U32), TensorInfo(TensorShape(2U), 1, DataType::U32), TensorInfo(TensorShape(2U), 1, DataType::U32), TensorInfo(TensorShape(2U), 1, DataType::U32), })), - framework::dataset::make("Expected", { false, false, false, false, true, true})), + make("Expected", { false, false, false, false, true, true})), src_info, dst_info, axis_info, expected) { Status s = CLReverse::validate(&src_info.clone()->set_is_resizable(false), &dst_info.clone()->set_is_resizable(false), - &axis_info.clone()->set_is_resizable(false)); + &axis_info.clone()->set_is_resizable(false), + false); ARM_COMPUTE_EXPECT(bool(s) == expected, framework::LogLevel::ERRORS); } // clang-format on @@ -93,7 +95,11 @@ TEST_SUITE(F16) FIXTURE_DATA_TEST_CASE(RunSmall, CLReverseFixture<half>, framework::DatasetMode::PRECOMMIT, - combine(run_small_dataset, framework::dataset::make("DataType", DataType::F16))) + combine( + run_small_dataset, + make("DataType", DataType::F16), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -102,7 +108,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, FIXTURE_DATA_TEST_CASE(RunLarge, CLReverseFixture<half>, framework::DatasetMode::NIGHTLY, - combine(run_large_dataset, framework::dataset::make("DataType", DataType::F16))) + combine( + run_large_dataset, + make("DataType", DataType::F16), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -113,7 +123,11 @@ TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, CLReverseFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(run_small_dataset, framework::dataset::make("DataType", DataType::F32))) + combine( + run_small_dataset, + make("DataType", DataType::F32), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -122,7 +136,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, FIXTURE_DATA_TEST_CASE(RunLarge, CLReverseFixture<float>, framework::DatasetMode::NIGHTLY, - combine(run_large_dataset, framework::dataset::make("DataType", DataType::F32))) + combine( + run_large_dataset, + make("DataType", DataType::F32), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -135,7 +153,11 @@ TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, CLReverseFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(run_small_dataset, framework::dataset::make("DataType", DataType::QASYMM8))) + combine( + run_small_dataset, + make("DataType", DataType::QASYMM8), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference); @@ -144,7 +166,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, FIXTURE_DATA_TEST_CASE(RunLarge, CLReverseFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(run_large_dataset, framework::dataset::make("DataType", DataType::QASYMM8))) + combine( + run_large_dataset, + make("DataType", DataType::QASYMM8), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(CLAccessor(_target), _reference); diff --git a/tests/validation/CL/RsqrtLayer.cpp b/tests/validation/CL/RsqrtLayer.cpp index 936d853d34..2353bda8d3 100644 --- a/tests/validation/CL/RsqrtLayer.cpp +++ b/tests/validation/CL/RsqrtLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 Arm Limited. + * Copyright (c) 2018-2019, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,8 +42,11 @@ namespace validation { namespace { -RelativeTolerance<float> tolerance_fp32(0.000001f); -RelativeTolerance<float> tolerance_fp16(0.001f); +RelativeTolerance<float> tolerance_fp32(0.000001f); +RelativeTolerance<float> tolerance_fp16(0.001f); +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for unsigned 8-bit asymmetric type */ +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_s(1); /**< Tolerance value for comparing reference's output against implementation's output for signed 8-bit asymmetric type */ + } // namespace TEST_SUITE(CL) TEST_SUITE(RsqrtLayer) @@ -68,6 +71,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( // *INDENT-ON* template <typename T> using CLRsqrtLayerFixture = RsqrtValidationFixture<CLTensor, CLAccessor, CLRsqrtLayer, T>; +template <typename T> +using CLRsqrtLayerQuantizedFixture = RsqrtQuantizedValidationFixture<CLTensor, CLAccessor, CLRsqrtLayer, T>; TEST_SUITE(Float) TEST_SUITE(FP16) @@ -102,6 +107,30 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLRsqrtLayerFixture<float>, framework::DatasetM TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, CLRsqrtLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED)), + framework::dataset::make("SrcQInfo", { QuantizationInfo(0.4044, -128) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(0.0027, -128) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8_s); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE(QASYMM8) + +FIXTURE_DATA_TEST_CASE(RunSmall, CLRsqrtLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::QASYMM8)), + framework::dataset::make("SrcQInfo", { QuantizationInfo(0.4044, 0) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(0.0027, 0) }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // RsqrtLayer TEST_SUITE_END() // CL } // namespace validation diff --git a/tests/validation/CL/Scale.cpp b/tests/validation/CL/Scale.cpp index 2b34f1f353..10a99ae34f 100644 --- a/tests/validation/CL/Scale.cpp +++ b/tests/validation/CL/Scale.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -82,6 +82,7 @@ constexpr AbsoluteTolerance<int16_t> tolerance_s16(1); constexpr float tolerance_f32_absolute(0.001f); RelativeTolerance<float> tolerance_f32(0.05); +constexpr float abs_tolerance_f16(0.1f); RelativeTolerance<half> tolerance_f16(half(0.1)); constexpr float tolerance_num_f32(0.01f); @@ -186,16 +187,6 @@ TEST_CASE(AlignedCornerNotSupported, framework::DatasetMode::ALL) ARM_COMPUTE_EXPECT(bool(result) == false, framework::LogLevel::ERRORS); } -TEST_CASE(WindowShrink, framework::DatasetMode::ALL) -{ - const auto input = TensorInfo{ TensorShape(37U, 37U, 2U), 1, DataType::F32 }; - const auto output = TensorInfo{ TensorShape(39U, 55U, 2U), 1, DataType::F32 }; - Status result{}; - - result = CLScale::validate(&input.clone()->set_is_resizable(false), &output.clone()->set_is_resizable(false), ScaleKernelInfo{ default_interpolation_policy, default_border_mode }); - ARM_COMPUTE_EXPECT(bool(result) == false, framework::LogLevel::ERRORS); -} - TEST_CASE(IncorrectScaleFactor, framework::DatasetMode::ALL) { const auto input = TensorInfo{ TensorShape(28U, 33U, 2U), 1, DataType::F32 }; @@ -272,7 +263,7 @@ FIXTURE_DATA_TEST_CASE(Run, CLScaleFixture<half>, framework::DatasetMode::ALL, A const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); // Validate output - validate(CLAccessor(_target), _reference, valid_region, tolerance_f16); + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); } FIXTURE_DATA_TEST_CASE(RunAlignCorners, CLScaleFixture<half>, framework::DatasetMode::ALL, ASSEMBLE_DATASET(f16_shape, ScaleAlignCornersSamplingPolicySet)) { @@ -281,7 +272,7 @@ FIXTURE_DATA_TEST_CASE(RunAlignCorners, CLScaleFixture<half>, framework::Dataset const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); // Validate output - validate(CLAccessor(_target), _reference, valid_region, tolerance_f16); + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); } const auto f16_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector<half>())), framework::dataset::make("DataType", DataType::F16)); FIXTURE_DATA_TEST_CASE(RunNightly, CLScaleFixture<half>, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET(f16_nightly_shape, ScaleSamplingPolicySet)) @@ -291,7 +282,7 @@ FIXTURE_DATA_TEST_CASE(RunNightly, CLScaleFixture<half>, framework::DatasetMode: const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); // Validate output - validate(CLAccessor(_target), _reference, valid_region, tolerance_f16); + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); } FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, CLScaleFixture<half>, framework::DatasetMode::NIGHTLY, ASSEMBLE_DATASET(f16_nightly_shape, ScaleAlignCornersSamplingPolicySet)) { @@ -300,7 +291,7 @@ FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, CLScaleFixture<half>, framework:: const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); // Validate output - validate(CLAccessor(_target), _reference, valid_region, tolerance_f16); + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); } TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float diff --git a/tests/validation/CL/ScatterLayer.cpp b/tests/validation/CL/ScatterLayer.cpp new file mode 100644 index 0000000000..b1531eb64a --- /dev/null +++ b/tests/validation/CL/ScatterLayer.cpp @@ -0,0 +1,298 @@ +/* + * Copyright (c) 2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/functions/CLScatter.h" +#include "tests/validation/fixtures/ScatterLayerFixture.h" +#include "tests/datasets/ScatterDataset.h" +#include "tests/CL/CLAccessor.h" +#include "arm_compute/function_info/ScatterInfo.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for fp32 data type */ +RelativeTolerance<float> tolerance_f16(0.02f); /**< Tolerance value for comparing reference's output against implementation's output for fp16 data type */ +RelativeTolerance<int32_t> tolerance_int(0); /**< Tolerance value for comparing reference's output against implementation's output for integer data types */ +} // namespace + +template <typename T> +using CLScatterLayerFixture = ScatterValidationFixture<CLTensor, CLAccessor, CLScatter, T>; + +using framework::dataset::make; + +TEST_SUITE(CL) +TEST_SUITE(Scatter) +DATA_TEST_CASE(Validate, framework::DatasetMode::PRECOMMIT, zip( + make("InputInfo", { TensorInfo(TensorShape(9U), 1, DataType::F32), // Mismatching data types + TensorInfo(TensorShape(15U), 1, DataType::F32), // Valid + TensorInfo(TensorShape(15U), 1, DataType::U8), // Valid + TensorInfo(TensorShape(8U), 1, DataType::F32), + TensorInfo(TensorShape(217U), 1, DataType::F32), // Mismatch input/output dims. + TensorInfo(TensorShape(217U), 1, DataType::F32), // Updates dim higher than Input/Output dims. + TensorInfo(TensorShape(12U), 1, DataType::F32), // Indices wrong datatype. + TensorInfo(TensorShape(9U, 3U, 4U), 1, DataType::F32), // Number of updates != number of indices + TensorInfo(TensorShape(17U, 3U, 3U, 2U), 1, DataType::F32), // index_len != (dst_dims - upt_dims + 1) + TensorInfo(TensorShape(17U, 3U, 3U, 2U, 2U, 2U), 1, DataType::F32), // index_len > 5 + }), + make("UpdatesInfo",{TensorInfo(TensorShape(3U), 1, DataType::F16), + TensorInfo(TensorShape(15U), 1, DataType::F32), + TensorInfo(TensorShape(15U), 1, DataType::U8), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(217U), 1, DataType::F32), + TensorInfo(TensorShape(217U, 3U), 1, DataType::F32), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(9U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(17U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(1U), 1, DataType::F32), + }), + make("IndicesInfo",{TensorInfo(TensorShape(1U, 3U), 1, DataType::S32), + TensorInfo(TensorShape(1U, 15U), 1, DataType::S32), + TensorInfo(TensorShape(1U, 15U), 1, DataType::S32), + TensorInfo(TensorShape(1U, 2U), 1, DataType::S32), + TensorInfo(TensorShape(1U, 271U), 1, DataType::S32), + TensorInfo(TensorShape(1U, 271U), 1, DataType::S32), + TensorInfo(TensorShape(1U, 2U), 1 , DataType::F32), + TensorInfo(TensorShape(1U, 4U), 1, DataType::S32), + TensorInfo(TensorShape(3U, 2U), 1, DataType::S32), + TensorInfo(TensorShape(6U, 2U), 1, DataType::S32), + }), + make("OutputInfo",{TensorInfo(TensorShape(9U), 1, DataType::F16), + TensorInfo(TensorShape(15U), 1, DataType::F32), + TensorInfo(TensorShape(15U), 1, DataType::U8), + TensorInfo(TensorShape(8U), 1, DataType::F32), + TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), + TensorInfo(TensorShape(271U), 1, DataType::F32), + TensorInfo(TensorShape(12U), 1, DataType::F32), + TensorInfo(TensorShape(9U, 3U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(17U, 3U, 3U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(17U, 3U, 3U, 2U, 2U, 2U), 1, DataType::F32), + }), + make("ScatterInfo",{ ScatterInfo(ScatterFunction::Add, false), + ScatterInfo(ScatterFunction::Max, false), + ScatterInfo(ScatterFunction::Max, false), + ScatterInfo(ScatterFunction::Min, false), + ScatterInfo(ScatterFunction::Add, false), + ScatterInfo(ScatterFunction::Update, false), + ScatterInfo(ScatterFunction::Sub, false), + ScatterInfo(ScatterFunction::Sub, false), + ScatterInfo(ScatterFunction::Update, false), + ScatterInfo(ScatterFunction::Update, false), + }), + make("Expected", { false, true, true, true, false, false, false, false, false, false })), + input_info, updates_info, indices_info, output_info, scatter_info, expected) +{ + const Status status = CLScatter::validate(&input_info, &updates_info, &indices_info, &output_info, scatter_info); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); +} + +const auto allScatterFunctions = make("ScatterFunction", + {ScatterFunction::Update, ScatterFunction::Add, ScatterFunction::Sub, ScatterFunction::Min, ScatterFunction::Max }); + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::Small1DScatterDataset(), + make("DataType", {DataType::F32}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {true}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +// With this test, src should be passed as nullptr. +FIXTURE_DATA_TEST_CASE(RunSmallZeroInit, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::Small1DScatterDataset(), + make("DataType", {DataType::F32}), + make("ScatterFunction", {ScatterFunction::Add}), + make("ZeroInit", {true}), + make("Inplace", {false}), + make("Padding", {true}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +// Updates/src/dst have same no. dims. +FIXTURE_DATA_TEST_CASE(RunSmallMultiDim, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMultiDimDataset(), + make("DataType", {DataType::F32}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {true}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +// m+1-D to m+n-D cases +FIXTURE_DATA_TEST_CASE(RunSmallMultiIndices, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMultiIndicesDataset(), + make("DataType", {DataType::F32}), + make("ScatterFunction", {ScatterFunction::Update, ScatterFunction::Add }), + make("ZeroInit", {false}), + make("Inplace", {false, true}), + make("Padding", {true}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +// m+k, k-1-D m+n-D case +FIXTURE_DATA_TEST_CASE(RunSmallBatchedMultiIndices, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterBatchedDataset(), + make("DataType", {DataType::F32}), + make("ScatterFunction", {ScatterFunction::Update, ScatterFunction::Add}), + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {true}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +// m+k, k-1-D m+n-D case +FIXTURE_DATA_TEST_CASE(RunSmallScatterScalar, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterScalarDataset(), + make("DataType", {DataType::F32}), + make("ScatterFunction", {ScatterFunction::Update, ScatterFunction::Add}), + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) // NOTE: Padding not supported in this datset +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() // FP32 + + +// NOTE: Padding is disabled for the SmallScatterMixedDataset due certain shapes not supporting padding. +// Padding is well tested in F32 Datatype test cases. + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::F16}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float + +TEST_SUITE(Integer) +TEST_SUITE(S32) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<int32_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::S32}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_int); +} +TEST_SUITE_END() // S32 + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<int16_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::S16}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_int); +} +TEST_SUITE_END() // S16 + +TEST_SUITE(S8) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::S8}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_int); +} +TEST_SUITE_END() // S8 + +TEST_SUITE(U32) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<uint32_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::U32}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_int); +} +TEST_SUITE_END() // U32 + +TEST_SUITE(U16) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<uint16_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::U16}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_int); +} +TEST_SUITE_END() // U16 + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmallMixed, CLScatterLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallScatterMixedDataset(), + make("DataType", {DataType::U8}), + allScatterFunctions, + make("ZeroInit", {false}), + make("Inplace", {false}), + make("Padding", {false}))) +{ + validate(CLAccessor(_target), _reference, tolerance_int); +} +TEST_SUITE_END() // U8 +TEST_SUITE_END() // Integer + +TEST_SUITE_END() // Scatter +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/Tile.cpp b/tests/validation/CL/Tile.cpp index a06c05744f..f243780c00 100644 --- a/tests/validation/CL/Tile.cpp +++ b/tests/validation/CL/Tile.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,6 +42,7 @@ namespace validation namespace { const auto MultiplesDataset = framework::dataset::make("Multiples", { Multiples{ 3 }, + Multiples{ 7 }, Multiples{ 2, 2 }, Multiples{ 1, 1, 3, 4 }, Multiples{ 2, 1, 2, 2 }, diff --git a/tests/validation/CL/Transpose.cpp b/tests/validation/CL/Transpose.cpp index 943534058b..6cf5fe8537 100644 --- a/tests/validation/CL/Transpose.cpp +++ b/tests/validation/CL/Transpose.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,12 +50,14 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::U16), // Invalid shape TensorInfo(TensorShape(20U, 13U), 1, DataType::U8), // Wrong data type TensorInfo(TensorShape(20U, 16U), 1, DataType::U32), // Valid + TensorInfo(TensorShape(20U, 16U, 3U, 3U), 1, DataType::U16), // Transpose only first two dimensions }), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(13U, 20U), 1, DataType::U32), TensorInfo(TensorShape(31U, 20U), 1, DataType::U16), TensorInfo(TensorShape(16U, 20U), 1, DataType::U32), + TensorInfo(TensorShape(16U, 20U, 3U, 3U), 1, DataType::U16), })), - framework::dataset::make("Expected", { false, false, true })), + framework::dataset::make("Expected", { false, false, true, true })), a_info, output_info, expected) { // Lock tensors @@ -80,6 +82,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLTransposeFixture<uint8_t>, framework::Dataset // Validate output validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunLargeHighDimensional, + CLTransposeFixture<uint8_t>, + framework::DatasetMode::NIGHTLY, + combine(concat(concat(datasets::Large3DShapes(), datasets::Large4DShapes()), + datasets::Large5dShapes()), + framework::dataset::make("DataType", DataType::U8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} TEST_SUITE_END() // U8 TEST_SUITE(U16) @@ -106,6 +118,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLTransposeFixture<uint32_t>, framework::Datase // Validate output validate(CLAccessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmallHighDimensional, + CLTransposeFixture<uint32_t>, + framework::DatasetMode::PRECOMMIT, + combine(concat(datasets::Small3DShapes(), datasets::Small4DShapes()), + framework::dataset::make("DataType", DataType::U32))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} FIXTURE_DATA_TEST_CASE(RunLarge, CLTransposeFixture<uint32_t>, framework::DatasetMode::NIGHTLY, combine(concat(datasets::Large1DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", DataType::U32))) { diff --git a/tests/validation/CL/UNIT/DynamicTensor.cpp b/tests/validation/CL/UNIT/DynamicTensor.cpp index ad2d4892ba..ac433721d8 100644 --- a/tests/validation/CL/UNIT/DynamicTensor.cpp +++ b/tests/validation/CL/UNIT/DynamicTensor.cpp @@ -29,10 +29,8 @@ #include "arm_compute/runtime/MemoryManagerOnDemand.h" #include "arm_compute/runtime/PoolManager.h" #include "src/core/CL/kernels/CLFillBorderKernel.h" -#include "src/core/CL/kernels/CLIm2ColKernel.h" #include "src/core/CL/kernels/CLL2NormalizeLayerKernel.h" #include "src/core/CL/kernels/CLReductionOperationKernel.h" -#include "src/core/CL/kernels/CLWeightsReshapeKernel.h" #include "tests/AssetsLibrary.h" #include "tests/CL/CLAccessor.h" #include "tests/Globals.h" diff --git a/tests/validation/CL/UNIT/Multithreaded.cpp b/tests/validation/CL/UNIT/Multithreaded.cpp new file mode 100644 index 0000000000..5c75df709d --- /dev/null +++ b/tests/validation/CL/UNIT/Multithreaded.cpp @@ -0,0 +1,113 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/RuntimeContext.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Macros.h" +#include "tests/framework/ParametersLibrary.h" +#include "tests/validation/Validation.h" +#include "arm_compute/runtime/CL/functions/CLActivationLayer.h" +#include "arm_compute/runtime/CL/functions/CLPixelWiseMultiplication.h" +#include "tests/validation/reference/ActivationLayer.h" +#include "tests/validation/reference/PixelWiseMultiplication.h" +#include <thread> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(UNIT) +TEST_SUITE(RuntimeContext) +// This test tries scheduling work concurrently from two independent threads +TEST_CASE(MultipleThreadedScheduller, framework::DatasetMode::ALL) +{ + constexpr auto num_threads(16u); + std::array<CLActivationLayer, num_threads> func{}; + std::array<CLPixelWiseMultiplication, num_threads> pmul{}; + std::array<CLTensor, num_threads> s0{}; + std::array<CLTensor, num_threads> s1{}; + + std::array<CLTensor, num_threads> st{}; + std::array<CLTensor, num_threads> dt{}; + + const TensorShape tensor_shape(128u, 4u, 5u); + const ActivationLayerInfo ainfo(ActivationLayerInfo::ActivationFunction::LOGISTIC, 0.5f, 1.f); + std::array<std::thread, num_threads> threads; + auto ctx = parameters->get_ctx<CLTensor>(); + + for(auto i = 0u; i < num_threads; ++i) + { + s0[i] = create_tensor<CLTensor>(tensor_shape, DataType::F32, 1); + s1[i] = create_tensor<CLTensor>(tensor_shape, DataType::F32, 1); + st[i] = create_tensor<CLTensor>(tensor_shape, DataType::F32, 1); + dt[i] = create_tensor<CLTensor>(tensor_shape, DataType::F32, 1); + func[i] = CLActivationLayer(ctx); + pmul[i] = CLPixelWiseMultiplication(); + threads[i] = + std::thread([&,i] + { + auto &s = st[i]; + auto &t = dt[i]; + auto &p0 = s0[i]; + auto &p1 = s1[i]; + pmul[i].configure(&p0, &p1, &s, 1.f, ConvertPolicy::WRAP, RoundingPolicy::TO_NEAREST_UP); + func[i].configure(&s, &t, ainfo); + s.allocator()->allocate(); + t.allocator()->allocate(); + p0.allocator()->allocate(); + p1.allocator()->allocate(); + library->fill_tensor_uniform(CLAccessor(p0), 0, -1.f, 1.f); + library->fill_tensor_uniform(CLAccessor(p1), 0, -1.f, 1.f); + pmul[i].run(); + func[i].run(); + }); + } + + for(auto &t : threads) + { + t.join(); + } + + SimpleTensor<float> rs{ tensor_shape, DataType::F32, 1 }; + SimpleTensor<float> ra{ tensor_shape, DataType::F32, 1 }; + SimpleTensor<float> rb{ tensor_shape, DataType::F32, 1 }; + library->fill_tensor_uniform(ra, 0, -1.f, 1.f); + library->fill_tensor_uniform(rb, 0, -1.f, 1.f); + const auto mul = reference::pixel_wise_multiplication<float, float, float>(ra, rb, 1.f, ConvertPolicy::WRAP, RoundingPolicy::TO_NEAREST_UP, DataType::F32); + const auto golden = reference::activation_layer<float>(mul, ainfo); + for(auto &d : dt) + { + validate(CLAccessor(d), golden); + } +} + +TEST_SUITE_END() // MultipleThreadedScheduller +TEST_SUITE_END() // UNIT +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/WeightsReshape.cpp b/tests/validation/CL/WeightsReshape.cpp index d04c10cee2..4345c4b08a 100644 --- a/tests/validation/CL/WeightsReshape.cpp +++ b/tests/validation/CL/WeightsReshape.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,7 +22,7 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "src/core/CL/kernels/CLWeightsReshapeKernel.h" +#include "src/gpu/cl/kernels/ClWeightsReshapeKernel.h" #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/datasets/ShapeDatasets.h" @@ -41,7 +41,7 @@ namespace validation TEST_SUITE(CL) TEST_SUITE(WeightsReshape) -using CLWeightsReshape = CLSynthetizeFunction<CLWeightsReshapeKernel>; +using ClWeightsReshape = ClSynthetizeOperatorWithBorder<opencl::kernels::ClWeightsReshapeKernel>; /** Validate tests * @@ -87,15 +87,15 @@ framework::dataset::make("NumGroups", { 1, 1, 1, 2, 1, 2 })), framework::dataset::make("Expected", { false, false, false, false, false, false })), input_info, biases_info, output_info, num_groups, expected) { - bool status = bool(CLWeightsReshape::validate(&input_info, &biases_info, &output_info, num_groups)); + bool status = bool(opencl::kernels::ClWeightsReshapeKernel::validate(&input_info, &biases_info, &output_info, num_groups)); ARM_COMPUTE_EXPECT(status == expected, framework::LogLevel::ERRORS); } template <typename T> -using CLWeightsReshapeFixture = WeightsReshapeValidationFixture<CLTensor, CLAccessor, CLWeightsReshape, T>; +using ClWeightsReshapeFixture = WeightsReshapeOpValidationFixture<CLTensor, CLAccessor, ClWeightsReshape, T>; TEST_SUITE(Float) -FIXTURE_DATA_TEST_CASE(FP32, CLWeightsReshapeFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(3U, 3U, 48U, 120U) }), +FIXTURE_DATA_TEST_CASE(FP32, ClWeightsReshapeFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(3U, 3U, 48U, 120U) }), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("HasBias", { true, false })), framework::dataset::make("NumGroups", { 1, 2 }))) @@ -104,7 +104,7 @@ FIXTURE_DATA_TEST_CASE(FP32, CLWeightsReshapeFixture<float>, framework::DatasetM validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(FP16, CLWeightsReshapeFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(13U, 13U, 96U, 240U) }), +FIXTURE_DATA_TEST_CASE(FP16, ClWeightsReshapeFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(13U, 13U, 96U, 240U) }), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("HasBias", { true, false })), framework::dataset::make("NumGroups", { 3, 4 }))) @@ -113,7 +113,7 @@ FIXTURE_DATA_TEST_CASE(FP16, CLWeightsReshapeFixture<half>, framework::DatasetMo validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(BFloat16, CLWeightsReshapeFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(9U, 9U, 96U, 240U) }), +FIXTURE_DATA_TEST_CASE(BFloat16, ClWeightsReshapeFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(9U, 9U, 96U, 240U) }), framework::dataset::make("DataType", DataType::BFLOAT16)), framework::dataset::make("HasBias", { false })), framework::dataset::make("NumGroups", { 3, 4 }))) @@ -125,7 +125,7 @@ FIXTURE_DATA_TEST_CASE(BFloat16, CLWeightsReshapeFixture<half>, framework::Datas TEST_SUITE_END() TEST_SUITE(Quantized) -FIXTURE_DATA_TEST_CASE(QASYMM8, CLWeightsReshapeFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(5U, 5U, 48U, 120U) }), +FIXTURE_DATA_TEST_CASE(QASYMM8, ClWeightsReshapeFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(5U, 5U, 48U, 120U) }), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("HasBias", { false })), framework::dataset::make("NumGroups", { 1, 2 }))) @@ -134,7 +134,7 @@ FIXTURE_DATA_TEST_CASE(QASYMM8, CLWeightsReshapeFixture<uint8_t>, framework::Dat validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(QASYMM8_SIGNED, CLWeightsReshapeFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(5U, 5U, 48U, 120U) }), +FIXTURE_DATA_TEST_CASE(QASYMM8_SIGNED, ClWeightsReshapeFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(framework::dataset::make("InputShape", { TensorShape(5U, 5U, 48U, 120U) }), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("HasBias", { false })), framework::dataset::make("NumGroups", { 1, 2 }))) diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp index 6ac37d1475..196e7edb8c 100644 --- a/tests/validation/CL/Winograd.cpp +++ b/tests/validation/CL/Winograd.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -30,6 +30,7 @@ #include "tests/CL/CLAccessor.h" #include "tests/CL/Helper.h" #include "tests/PaddingCalculator.h" +#include "tests/datasets/ActivationFunctionsDataset.h" #include "tests/datasets/LargeConvolutionLayerDataset.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/datasets/SmallConvolutionLayerDataset.h" @@ -47,6 +48,7 @@ namespace test { namespace validation { +using framework::dataset::make; namespace { // *INDENT-OFF* @@ -57,108 +59,232 @@ const AbsoluteTolerance<half> tolerance_convolution_layer_f16(half(0.4f)); RelativeTolerance<half_float::half> rel_tolerance_f16(half(0.2)); /**< Tolerance value for comparing reference's output against implementation's output for FP16 data types */ constexpr float tolerance_num = 0.05f; /**< Tolerance number */ constexpr float abs_tolerance_convolution_layer_f16 = 2.5f; /**< Tolerance number */ -constexpr float tolerance_num_f16 = 0.15f; /**< Tolerance number */ +constexpr float tolerance_num_f16 = 0.15f; /**< Tolerance number */ -//Activation Functions -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +const auto ActivationFunctionsDataset = make("ActivationInfo", { - ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.8f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SOFT_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ABS), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SQUARE), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::HARD_SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 2.f, 1.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::GELU) }); -const auto ActivationFunctionsSmallDataset = framework::dataset::make("ActivationInfo", + +const auto ActivationFunctionsSmallDataset = make("ActivationInfo", { ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SOFT_RELU) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.8f, -0.5f) }); } // namespace using namespace arm_compute::misc::shape_calculator; +/* + Testing Strategy of CL Winograd: + - For nchw and nhwc and for each kernel size, we have a dedicated OpenCL kernel. + (except 1xN and Nx1 uses NxN under the hood). Therefore, test cases should be + stressed for each of these configurations. + - Fp32 and Fp16 kernels are the same. Only the DATA_TYPE build option changes + between these two. Because the same kernel is stressed thoroughly for both + small and large shapes for Fp32 data type, Fp16 kernels are run on a subset + of the shapes, because we get diminishing returns by exhaustively testing the + same kernel. + - Activations only affect the output stage and it's calculated on the output tile. + Exhaustively testing all activations with all the shapes does not provide much + value but increases the testing time quite significantly. Therefore, all activations + are tested in a subset of the shapes, and for all MxM kernels and data layouts as + they represent different OpenCL kernels. (1xM and Mx1 kernels use MxM under the hood). +*/ TEST_SUITE(CL) TEST_SUITE(Winograd) TEST_SUITE(ConvolutionLayer) -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { - TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F16), // Insufficient padding - TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32), // Datatype mismatch - TensorInfo(TensorShape(23U, 27U, 5U, 4U), 1, DataType::F32), // Stride y not supported - TensorInfo(TensorShape(16U, 16U, 8U), 1, DataType::F32), // Padding needed - TensorInfo(TensorShape(33U, 27U, 7U, 4U), 1, DataType::F32) // Kernel size not supported - }), - framework::dataset::make("WeightsInfo", { - TensorInfo(TensorShape(3U, 3U, 2U, 19U), 1, DataType::F16), - TensorInfo(TensorShape(3U, 3U, 2U, 19U), 1, DataType::QASYMM8), - TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32), - TensorInfo(TensorShape(3U, 3U, 8U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(5U, 5U, 7U, 16U), 1, DataType::F16) - })), - framework::dataset::make("BiasesInfo", { - TensorInfo(TensorShape(19U), 1, DataType::F16), - TensorInfo(TensorShape(19U), 1, DataType::F32), - TensorInfo(TensorShape(21U), 1, DataType::F32), - TensorInfo(TensorShape(16U), 1, DataType::F32), - TensorInfo(TensorShape(16U), 1, DataType::F32) - })), - framework::dataset::make("OutputInfo", { - TensorInfo(TensorShape(17U, 31U, 19U), 1, DataType::F16), - TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32), - TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(16U, 16U, 16U), 1, DataType::F32), - TensorInfo(TensorShape(11U, 12U, 16U, 4U), 1, DataType::F32) - })), - framework::dataset::make("ConvInfo", { - PadStrideInfo(1, 1, 1, 1), - PadStrideInfo(1, 1, 1, 1), - PadStrideInfo(1, 2, 0, 0), - PadStrideInfo(1, 1, 1, 1), - PadStrideInfo(1, 1, 1, 0) - })), - framework::dataset::make("Expected", { false, false, false, false, false })), - input_info, weights_info, bias_info, output_info, conv_info, expected) +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + make("InputInfo", { + TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F16), // Insufficient padding + TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32), // Datatype mismatch + TensorInfo(TensorShape(23U, 27U, 5U, 4U), 1, DataType::F32), // Stride y not supported + TensorInfo(TensorShape(16U, 16U, 8U), 1, DataType::F32), // Padding needed + TensorInfo(TensorShape(33U, 27U, 7U, 4U), 1, DataType::F32) // Kernel size not supported + }), + make("WeightsInfo", { + TensorInfo(TensorShape(3U, 3U, 2U, 19U), 1, DataType::F16), + TensorInfo(TensorShape(3U, 3U, 2U, 19U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 3U, 8U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 5U, 7U, 16U), 1, DataType::F16) + }), + make("BiasesInfo", { + TensorInfo(TensorShape(19U), 1, DataType::F16), + TensorInfo(TensorShape(19U), 1, DataType::F32), + TensorInfo(TensorShape(21U), 1, DataType::F32), + TensorInfo(TensorShape(16U), 1, DataType::F32), + TensorInfo(TensorShape(16U), 1, DataType::F32) + }), + make("OutputInfo", { + TensorInfo(TensorShape(17U, 31U, 19U), 1, DataType::F16), + TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32), + TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(16U, 16U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(11U, 12U, 16U, 4U), 1, DataType::F32) + }), + make("ConvInfo", { + PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 2, 0, 0), + PadStrideInfo(1, 1, 1, 1), + PadStrideInfo(1, 1, 1, 0) + }), + make("Expected", { false, false, false, false, false })), + input_info, weights_info, bias_info, output_info, conv_info, expected) { ARM_COMPUTE_EXPECT(bool(CLWinogradConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info)) == expected, framework::LogLevel::ERRORS); } +DATA_TEST_CASE(SupportedKernels, framework::DatasetMode::ALL, zip( + make("WeightsInfo", { + // Shapes are always in NCHW format. When layout is NHWC, the shape is permuted + + // Fp32/16, NCHW + // 3x1, 1x3, 3x3 --> all TRUE + TensorInfo(TensorShape(3U, 3U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 3U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(3U, 1U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + + // 5x1, 1x5, 5x5 --> all TRUE + TensorInfo(TensorShape(5U, 5U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 5U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + TensorInfo(TensorShape(5U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + + // 7x1, 1x7, 7x7 + // nchw does not support kernels with size 7 --> all FALSE + TensorInfo(TensorShape(7U, 7U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 7U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(7U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + + // unsupported kernel sizes + TensorInfo(TensorShape(2U, 2U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(5U, 2U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(3U, 6U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + + // Fp32/16, NHWC + // 7x1, 1x7, 7x7 --> all TRUE + TensorInfo(TensorShape(7U, 7U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(1U, 7U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(7U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + + // 3x1, 1x3, 3x3 --> all TRUE + TensorInfo(TensorShape(3U, 3U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(1U, 3U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + + // 5x1, 1x5, 5x5 --> all TRUE + TensorInfo(TensorShape(5U, 5U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(1U, 5U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(5U, 1U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + + // unsupported kernel sizes + TensorInfo(TensorShape(2U, 2U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(5U, 2U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 6U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + + }), + make("Expected", { + true, true, true, // nchw, 3x3, 1x3, 3x1 + true, true, true, // nchw, 5x5, 1x5, 5x1 + false, false, false, // nchw, 7x7, 1x7, 7x1 + false, false, false, // nchw, random unsupported kernels + true, true, true, // nhwc, 7x7, 1x7, 7x1 + true, true, true, // nhwc, 3x3, 1x3, 3x1 + true, true, true, // nhwc, 5x5, 1x5, 5x1 + false, false, false, // nchw, random unsupported kernels + })), + weights_info_const, expected) +{ + DataType data_type = weights_info_const.data_type(); + DataLayout data_layout = weights_info_const.data_layout(); + + TensorInfo input_info = TensorInfo(TensorShape(17U, 31U, 2U), 1, data_type); + TensorInfo bias_info = TensorInfo(TensorShape(8U), 1, data_type); + TensorInfo weights_info = weights_info_const; + + if(data_layout == DataLayout::NHWC) + { + // Convert to NHWC + PermutationVector perm = PermutationVector(2U, 0U, 1U); + + TensorShape input_shape = input_info.tensor_shape(); + TensorShape weights_shape = weights_info.tensor_shape(); + permute(input_shape, perm); + permute(weights_shape, perm); + + input_info.set_tensor_shape(input_shape); + weights_info.set_tensor_shape(weights_shape); + + input_info.set_data_layout(data_layout); + weights_info.set_data_layout(data_layout); + bias_info.set_data_layout(data_layout); + } + + PadStrideInfo conv_info(1, 1, 0, 0); + + TensorShape output_shape = compute_deep_convolution_shape(input_info, weights_info, conv_info); + TensorInfo output_info = TensorInfo(output_shape, 1, data_type, data_layout); + + Status status = CLWinogradConvolutionLayer::validate( + &input_info, + &weights_info, + &bias_info, + &output_info, + conv_info, + ActivationLayerInfo(), + true /* fast math */); + + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); +} + TEST_SUITE(FP32) using CLWinogradConvolutionLayerFastMathFixture = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, float>; using CLWinogradConvolutionLayerFastMathMixedDataLayoutFixture = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, float, float, true, true>; TEST_SUITE(Conv3x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLWinogradConvolutionLayerFastMathMixedDataLayoutFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(8U, 8U, 32U)), - framework::dataset::make("Weight", TensorShape(1U, 3U, 32U, 1U))), - framework::dataset::make("Bias", TensorShape(1U))), - framework::dataset::make("Output", TensorShape(8U, 6U, 1U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1U, 1U))), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, + combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + +FIXTURE_DATA_TEST_CASE(RunActivations, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, + combine( + make("Input", TensorShape(8U, 8U, 32U)), + make("Weight", TensorShape(3U, 3U, 32U, 4U)), + make("Bias", TensorShape(4U)), + make("Output", TensorShape(6U, 6U, 4U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); @@ -167,20 +293,20 @@ TEST_SUITE_END() // Conv3x3 TEST_SUITE(Conv3x1) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer3x1Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); @@ -189,20 +315,36 @@ TEST_SUITE_END() // Conv3x1 TEST_SUITE(Conv1x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); +} + +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, CLWinogradConvolutionLayerFastMathMixedDataLayoutFixture, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(8U, 8U, 32U)), + make("Weight", TensorShape(1U, 3U, 32U, 1U)), + make("Bias", TensorShape(1U)), + make("Output", TensorShape(8U, 6U, 1U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x3Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); @@ -211,10 +353,10 @@ TEST_SUITE_END() // Conv1x3 TEST_SUITE(Conv5x5) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset ), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -222,11 +364,27 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, fram } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset ), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); +} +FIXTURE_DATA_TEST_CASE(RunActivations, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, + combine( + make("Input", TensorShape(13U, 13U, 32U)), + make("Weight", TensorShape(5U, 5U, 32U, 4U)), + make("Bias", TensorShape(4U)), + make("Output", TensorShape(9U, 9U, 4U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); @@ -235,10 +393,10 @@ TEST_SUITE_END() // Conv5x5 TEST_SUITE(Conv5x1) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer5x1Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -246,10 +404,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, fram } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer5x1Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -259,10 +417,10 @@ TEST_SUITE_END() // Conv5x1 TEST_SUITE(Conv1x5) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x5Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -270,16 +428,63 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, fram } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x5Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); } TEST_SUITE_END() // Conv1x5 + +TEST_SUITE(Conv1x7) +FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallWinogradConvolutionLayer1x7Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NHWC }))) + +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); +} + +FIXTURE_DATA_TEST_CASE(RunActivations, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::NIGHTLY, + combine( + make("Input", TensorShape(13U, 13U, 32U)), + make("Weight", TensorShape(1U, 7U, 32U, 4U)), + make("Bias", TensorShape(4U)), + make("Output", TensorShape(13U, 11U, 4U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 2)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); +} +TEST_SUITE_END() // Conv1x7 + +TEST_SUITE(Conv7x1) +FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallWinogradConvolutionLayer7x1Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NHWC }))) + +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f32); +} +TEST_SUITE_END() // Conv7x1 + +/** @note: Although 7x7 is in the kernels, reference implementation + * does not support it. So, it remains as a "test gap". + */ + TEST_SUITE_END() // FP32 @@ -288,20 +493,36 @@ TEST_SUITE(FP16) using CLWinogradConvolutionLayerFastMathFixture16 = WinogradConvolutionLayerFastMathValidationFixture<CLTensor, CLAccessor, CLWinogradConvolutionLayer, half, float>; TEST_SUITE(Conv3x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16, tolerance_num_f16); } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer3x3DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); +} + +FIXTURE_DATA_TEST_CASE(RunActivations, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, + combine( + make("Input", TensorShape(8U, 8U, 32U)), + make("Weight", TensorShape(3U, 3U, 32U, 6U)), + make("Bias", TensorShape(6U)), + make("Output", TensorShape(6U, 6U, 6U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F16 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); @@ -310,20 +531,20 @@ TEST_SUITE_END() // Conv3x3 TEST_SUITE(Conv3x1) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16, tolerance_num_f16); } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x1Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer3x1DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); @@ -332,20 +553,20 @@ TEST_SUITE_END() // Conv3x1 TEST_SUITE(Conv1x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16, tolerance_num_f16); } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x3Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x3DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); @@ -354,10 +575,10 @@ TEST_SUITE_END() // Conv1x3 TEST_SUITE(Conv5x5) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -365,23 +586,39 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, fr } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer5x5DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } + +FIXTURE_DATA_TEST_CASE(RunActivations, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, + combine( + make("Input", TensorShape(13U, 13U, 32U)), + make("Weight", TensorShape(5U, 5U, 32U, 6U)), + make("Bias", TensorShape(6U)), + make("Output", TensorShape(9U, 9U, 6U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F16 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); +} TEST_SUITE_END() // Conv5x5 TEST_SUITE(Conv5x1) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x1Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer5x1Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -389,10 +626,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, fr } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x1Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer5x1DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -402,10 +639,10 @@ TEST_SUITE_END() // Conv5x1 TEST_SUITE(Conv1x5) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x5Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x5Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -413,10 +650,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, fr } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x5Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x5DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -426,10 +663,10 @@ TEST_SUITE_END() // Conv1x5 TEST_SUITE(Conv1x7) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x7Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsSmallDataset), - framework::dataset::make("DataLayout", { DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x7Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NHWC }))) { // Validate output @@ -437,16 +674,46 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, fr } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x7Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x7DatasetFp16Subset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) + +{ + // Validate output + validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); +} +FIXTURE_DATA_TEST_CASE(RunActivations, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, + combine( + make("Input", TensorShape(13U, 13U, 32U)), + make("Weight", TensorShape(1U, 7U, 32U, 6U)), + make("Bias", TensorShape(6U)), + make("Output", TensorShape(13U, 7U, 6U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F16 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(CLAccessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_convolution_layer_f16); } TEST_SUITE_END() // Conv1x7 + +TEST_SUITE(Conv7x1) +FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallWinogradConvolutionLayer7x1Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsSmallDataset, + make("DataLayout", { DataLayout::NHWC }))) + +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_convolution_layer_f16, tolerance_num_f16); +} +TEST_SUITE_END() // Conv7x1 + TEST_SUITE_END() // FP16 TEST_SUITE_END() // ConvolutionLayer TEST_SUITE_END() // Winograd diff --git a/tests/validation/CMakeLists.txt b/tests/validation/CMakeLists.txt new file mode 100644 index 0000000000..448e96c4f9 --- /dev/null +++ b/tests/validation/CMakeLists.txt @@ -0,0 +1,146 @@ +# Copyright (c) 2023 Arm Limited. +# +# SPDX-License-Identifier: MIT +# +# Permission is hereby granted, free of charge, to any person obtaining a copy +# of this software and associated documentation files (the "Software"), to +# deal in the Software without restriction, including without limitation the +# rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +# sell copies of the Software, and to permit persons to whom the Software is +# furnished to do so, subject to the following conditions: +# +# The above copyright notice and this permission notice shall be included in all +# copies or substantial portions of the Software. +# +# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +# SOFTWARE. + +target_sources( + arm_compute_validation + PRIVATE UNIT/SafeIntegerOps.cpp + UNIT/Version.cpp + UNIT/TensorInfo.cpp + UNIT/TensorShape.cpp + UNIT/Utils.cpp + UNIT/SubTensorInfo.cpp + UNIT/WindowIterator.cpp + UNIT/LifetimeManager.cpp + UNIT/GPUTarget.cpp + CPP/DetectionPostProcessLayer.cpp + CPP/TopKV.cpp + CPP/DFT.cpp + CPP/Permute.cpp + CPP/NonMaximumSuppression.cpp) + +if(ENABLE_NEON) + target_sources( + arm_compute_validation + PRIVATE NEON/ElementwiseNegation.cpp + NEON/BoundingBoxTransform.cpp + NEON/ChannelShuffle.cpp + NEON/Logical.cpp + NEON/DilatedConvolutionLayer.cpp + NEON/PoolingLayer.cpp + NEON/BitwiseNot.cpp + NEON/FillBorder.cpp + NEON/ElementwiseRsqrtLayer.cpp + NEON/DepthConcatenateLayer.cpp + NEON/ElementwisePower.cpp + NEON/Fill.cpp + NEON/ROIPoolingLayer.cpp + NEON/LSTMLayer.cpp + NEON/ArithmeticSubtraction.cpp + NEON/GEMMLowp.cpp + NEON/Unstack.cpp + NEON/Slice.cpp + NEON/Pooling3dLayer.cpp + NEON/BitwiseOr.cpp + NEON/HeightConcatenateLayer.cpp + NEON/ReshapeLayer.cpp + NEON/SoftmaxLayer.cpp + NEON/Gather.cpp + NEON/CropResize.cpp + NEON/ReductionOperation.cpp + NEON/PixelWiseMultiplication.cpp + NEON/LogSoftmaxLayer.cpp + NEON/DepthConvertLayer.cpp + NEON/Flatten.cpp + NEON/ElementwiseKernelSelection.cpp + NEON/DepthToSpaceLayer.cpp + NEON/ElementwiseAbsoluteValue.cpp + NEON/PadLayer.cpp + NEON/MeanStdDevNormalizationLayer.cpp + NEON/GlobalPoolingLayer.cpp + NEON/RNNLayer.cpp + NEON/DetectionPostProcessLayer.cpp + NEON/ElementwiseRound.cpp + NEON/BitwiseXor.cpp + NEON/GEMM.cpp + NEON/FuseBatchNormalization.cpp + NEON/BitwiseAnd.cpp + NEON/ElementwiseMax.cpp + NEON/ReduceMean.cpp + NEON/Reverse.cpp + NEON/L2NormalizeLayer.cpp + NEON/Convolution3D.cpp + NEON/ArithmeticAddition.cpp + NEON/ActivationLayer.cpp + NEON/SpaceToBatchLayer.cpp + NEON/ElementwiseLog.cpp + NEON/LSTMLayerQuantized.cpp + NEON/Im2Col.cpp + NEON/DequantizationLayer.cpp + NEON/DeconvolutionLayer.cpp + NEON/Select.cpp + NEON/ElementwiseSin.cpp + NEON/PReluLayer.cpp + NEON/BatchNormalizationLayer.cpp + NEON/ElementwiseMin.cpp + NEON/InstanceNormalizationLayer.cpp + NEON/ROIAlignLayer.cpp + NEON/ElementwiseDivision.cpp + NEON/ElementwiseExpLayer.cpp + NEON/ArgMinMax.cpp + NEON/QLSTMLayerNormalization.cpp + NEON/Col2Im.cpp + NEON/Split.cpp + NEON/Transpose.cpp + NEON/GenerateProposalsLayer.cpp + NEON/StackLayer.cpp + NEON/WidthConcatenateLayer.cpp + NEON/NormalizationLayer.cpp + NEON/Copy.cpp + NEON/ElementwiseSquareDiff.cpp + NEON/MaxUnpoolingLayer.cpp + NEON/Permute.cpp + NEON/Comparisons.cpp + NEON/BatchConcatenateLayer.cpp + NEON/Tile.cpp + NEON/BatchToSpaceLayer.cpp + NEON/SpaceToDepthLayer.cpp + NEON/DepthwiseConvolutionLayerNative.cpp + NEON/QuantizationLayer.cpp + NEON/ConvertFullyConnectedWeights.cpp + NEON/Floor.cpp + NEON/FFT.cpp + NEON/Cast.cpp + NEON/PriorBoxLayer.cpp + NEON/Scale.cpp + NEON/ReorgLayer.cpp + NEON/Range.cpp + NEON/DirectConvolutionLayer.cpp + NEON/DepthwiseConvolutionLayer.cpp + NEON/FullyConnectedLayer.cpp + NEON/ConvolutionLayer.cpp + NEON/StridedSlice.cpp + NEON/ReorderLayer.cpp + NEON/UNIT/DynamicTensor.cpp + NEON/UNIT/TensorAllocator.cpp + NEON/UNIT/MemoryManager.cpp + NEON/UNIT/RuntimeContext.cpp) +endif() diff --git a/tests/validation/CPP/DFT.cpp b/tests/validation/CPP/DFT.cpp index e19e850589..84431399be 100644 --- a/tests/validation/CPP/DFT.cpp +++ b/tests/validation/CPP/DFT.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -125,7 +125,7 @@ DATA_TEST_CASE(Real, framework::DatasetMode::ALL, shapes_2d_dft, auto backward = reference::ridft_2d(forward, is_odd); // Validate with input - validate(SimpleTensorAccessor<float>(src), backward, RelativeTolerance<float>(0.1f)); + validate(SimpleTensorAccessor<float>(src), backward, RelativeTolerance<float>(0.1f), 0.f, AbsoluteTolerance<float>(0.001f)); } DATA_TEST_CASE(Complex, framework::DatasetMode::ALL, shapes_2d_dft, diff --git a/tests/validation/Helpers.cpp b/tests/validation/Helpers.cpp index 0f5d5c5101..560460fd33 100644 --- a/tests/validation/Helpers.cpp +++ b/tests/validation/Helpers.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,9 +22,12 @@ * SOFTWARE. */ #include "tests/validation/Helpers.h" +#include "tests/framework/Asserts.h" #include <algorithm> #include <cmath> +#include <cstdint> +#include <tuple> namespace arm_compute { @@ -349,6 +352,225 @@ void add_padding_x(std::initializer_list<ITensor *> tensors, const DataLayout &d } } +QuantizationHint suggest_conv_dst_q_info_and_bias(const QuantizationInfo &in_q_info, + const QuantizationInfo &weight_q_info, + int32_t height, + int32_t width, + int32_t channels, + DataType data_type, + float bias_fraction) +{ + /** Quantization Setup of convolution + * + * Just like any other multiply-accummulate, convolution (2D) operation + * multiplies and accumulates the input and weight tensors. This operation + * takes place in three dimensions: height, width and channels. All of them + * belong to the weight tensor. + * + * The formula for simple convolution can be written as: + * C = sum_h sum_w sum_c(I[h_offset + h, w_offset + w, c] * W[h, w, c]) + * + * Here, h_offset and w_offset are the starting positions in the image. Effects + * of paddings are ignored. This accumulation reduces to something like + * + * C = sum_m(I_index * W_hwc) + * where m is height x width x channels. + * + * Non-unit strides and/or dilations do not change the probabilistic nature of + * this sum because we always iterate as the size of the weight tensor. + * + * Paddings may affect this summation, but it's a boundary condition and so is + * neglected for brevity. + */ + + return suggest_mac_dst_q_info_and_bias(in_q_info, weight_q_info, height * width * channels, data_type, bias_fraction); +} + +QuantizationHint suggest_matmul_dst_q_info_and_bias(const QuantizationInfo &lhs_q_info, + const QuantizationInfo &rhs_q_info, + int32_t m, int32_t n, int32_t k, DataType data_type, + float bias_fraction) +{ + ARM_COMPUTE_UNUSED(m, n); + + /** Quantization Setup of matrix multiplication + * + * We have a matrix multiplication of the form C = A * B + D + * where A is (m X k), B is (k x n) and C is therefore (m x n). + * The bias, D is (1 x n). + * + * If we have some distributional statistics of A, B and D, i.e. mean and variance, + * we can estimate the mean and variance of a single value in C matrix and pick + * good scale and offset values for the output and have non-saturated tests. + * + * Each element in the output matrix can be calculated as follows: + * C_ij = sum_k(A_ik * B_kj) + D_j + * + * Note: All possible A_ik, B_kj, D_j random variables are assumed mutually independent. + * Note: In quantized operators, bias is an integer. But, its quantization scale is + * assumed to be equal to lhs_scale * rhs_scale, and offset equal to 0. + * Note: Since, bias is an integer that should be given as input, we need to pick responsible + * values when adding it on top of the summation. This is where "bias_fraction" comes + * into play. Based on the fraction given, we also return suggested bias range (min/max) + * for not saturating the output. + * + * Because all random variables are mutually independent, any C_ij has the same statistics, + * which is why we return a single destination quantization info object; which is why we can + * resort to a more general calculation explained in suggest_mac_dst_q_info_and_bias(). + * + * From a probabilistic perspective, the above calculation reduces to + * c = sum_k (a_k * b_k) + d + */ + + return suggest_mac_dst_q_info_and_bias(lhs_q_info, rhs_q_info, k, data_type, bias_fraction); +} + +QuantizationHint suggest_mac_dst_q_info_and_bias( + const QuantizationInfo &a_q_info, const QuantizationInfo &b_q_info, int32_t K, DataType data_type, float bias_fraction, int num_sd) +{ + QuantizationInfo c_q_info; + + ARM_COMPUTE_ASSERT(data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED); + + const int32_t t_max = static_cast<int32_t>(data_type == DataType::QASYMM8 ? std::numeric_limits<uint8_t>::max() : std::numeric_limits<int8_t>::max()); + const int32_t t_min = static_cast<int32_t>(data_type == DataType::QASYMM8 ? std::numeric_limits<uint8_t>::min() : std::numeric_limits<int8_t>::min()); + + /** Quantization Setup of multiply-accummulate + * + * Expression (in float): + * C = sum_k ( A_k * B_k ) + D + * + * Lemma: An affine transformation (i.e. aX + b) to a discrete uniform random variable + * creates another discrete uniform random variable. + * + * Terminology: + * E[X]: Mean of the random variable X (sometimes referred as mu_x) + * var(X): Variance of the random variable X (someimes referred as sigma^2_x) + * std(X): sqrt(var(X)), standard deviation of X + * + * 1) Calculate the mean: + * E[C] = sum_k( E[A_k] * E[B_k] ) + D = K * mean_a * mean_b + mean_d + * + * Since elements of A and B are uniformly distributed random variables, we have + * mean_a = (max_a + min_a) / 2, mean_b = (max_b + min_b ) / 2 + * max_a and min_a can be calculated with the scale_a/b and offset_a/b + * by replacing data type minimum and maximums in the equations + * + * We don't know mean_d because we have to choose it based on bias_fraction. If we call + * the summation as M_int, similar to above, we have: + * + * E[C_int] = sum_k( E[A_k_int] * E[B_k_int] ) + E[D_int] = K * mean_a_int * mean_b_int + mean_d_int + * \___________________________/ + * E[M_int] + * + * We choose a bias mean proportional to the integer summation. This proportion is "bias_fraction". + * So, we have D_int = f * M_int (f: fraction), and + * E[D_int] = mean_d_int = f * E[M_int] + * + * This also means, for floating point value of D, the following: + * E[D] = mean_d = E[D_int] * a_scale * b_scale + * + * 2) Calculate the variance: + * var(C) = sum_k( var(A_k * B_k) ) + var(D) + * = sum_k ( E[A_k^2 * B_k^2] - E[A_k]^2E[B_k^2] ) + * = ... + * = K * (var_a * var_b + var_a * mean^2_b + var_b * mean^2_a) + var_d + * + * Similarly, due to uniform random variable properties, we have + * var_a = (max_a - min_a)^2 / 12 + * var_b = (max_b - min_b)^2 / 12 + * + * Again, we don't know var_d as we don't know the bias. As set out in the previous section, we have + * var(D_int) = var(f * M_int) = f^2 * var(M_int) + * + * Using the same expression, we can find var(M_int): + * var(C_int) = sum_k( var(A_k_int * B_k_int) ) + var(D_int) + * = sum_k ( E[A_k_int^2 * B_k_int^2] - E[A_k_int]^2E[B_k_int^2] ) + * = ... + * = K * (var_a_int * var_b_int + var_a_int * mean^2_b_int + var_b_int * mean^2_a_int) + var_d_int + * \_______________________________________________________________________________/ + * var(M_int) + * + * Now, we know mean and variance of D_int, we can return a suitable bias range with + * [mean_d_int +/- 2 * std_d_int] + * + * This also means, for floating point value of D, the following: + * var(D) = var_d = var(D_int) * a_scale^2 * b_scale^2 + * + * E[D] and var(D) calculated in steps (1) and (2) can be substituted into E[C] and var(C) calculatons. + * + * 3) Now, we have an idea of what would an average C will look like and how much deviation + * is present around it. The exact distribution of C is difficult to come up with dependent on K. + * But, as K increases, due to Central Limit Theorem, it'll look more like a bell shaped figure, + * approaching normal distribution. + * + * This is useful because, in normal distribution, we know that values +- 2 std_deviation around + * the mean constitute 95% of the values. Therefore, setting a plausible range for us: + * C_range = [C_min, C_max] = [mean_c - 2 * std_c, mean_c + 2 * std_c] + * + * 4) + * If we map this [C_min, C_max] to [0, 255] or [-128, 127] depending on the signedness of the + * data type, we can find a suitable scale and offset for the output. On average, it's expected + * that 5% of the output values will saturate and 95% will remain in the range. + * + * The equations to be solved for offset_c and scale_c are: + * C_min = scale_c * (type_min - offset_c) + * C_max = scale_c * (type_max - offset_c) + */ + + const int32_t a_offset = a_q_info.uniform().offset; + const float a_scale = a_q_info.uniform().scale; + const int32_t b_offset = b_q_info.uniform().offset; + const float b_scale = b_q_info.uniform().scale; + + // Integer value statistics. Valid for both Lhs/A and Rhs/B + const float mean_a_int = (t_max + t_min) / 2.f; + constexpr float var_a_int = (256 * 256 - 1) / 12.f; // Discrete uniform RV variance + const float mean_b_int = mean_a_int; // A_int and B_int has the same stats + constexpr float var_b_int = var_a_int; + + // Lhs/A stats + const float max_a = (t_max - a_offset) * a_scale; + const float min_a = (t_min - a_offset) * a_scale; + const float mean_a = (max_a + min_a) / 2; + const float var_a = (max_a - min_a) * (max_a - min_a) / 12; + + // Rhs/B stats + const float max_b = (t_max - b_offset) * b_scale; + const float min_b = (t_min - b_offset) * b_scale; + const float mean_b = (max_b + min_b) / 2; + const float var_b = (max_b - min_b) * (max_b - min_b) / 12; + + // Integer multiplication output/M stats + const float mean_m_int = K * mean_a_int * mean_b_int; + const float var_m_int = K * (var_a_int * var_b_int + mean_a_int * var_b_int + mean_b_int + var_a_int); + const float std_m_int = sqrt(var_m_int); + + // Bias/D both Int and Float statistics + const float mean_d_int = bias_fraction * mean_m_int; + const float std_d_int = bias_fraction * std_m_int; + const float mean_d = a_scale * b_scale * mean_d_int; + const float std_d = a_scale * b_scale * std_d_int; + const float var_d = std_d * std_d; + + // Also calculate the suggested bias range + const int32_t min_bias = mean_d_int - (num_sd * std_d_int); + const int32_t max_bias = mean_d_int + (num_sd * std_d_int); + + // Output/C stats + const float mean_out = K * mean_a * mean_b + mean_d; + const float var_out = K * (var_a * var_b + var_a * mean_b * mean_b + var_b * mean_a * mean_a) + var_d; + const float std_out = sqrt(var_out); + + // Output quantization setup + const float scale_out = (2 * num_sd) * std_out / 255; + const int32_t offset_out = static_cast<int32_t>(t_min - (mean_out - (num_sd * std_out)) / scale_out); + + c_q_info = QuantizationInfo(scale_out, offset_out); + + return { c_q_info, min_bias, max_bias }; +} + template void get_tile(const SimpleTensor<float> &in, SimpleTensor<float> &roi, const Coordinates &coord); template void get_tile(const SimpleTensor<half> &in, SimpleTensor<half> &roi, const Coordinates &coord); template void get_tile(const SimpleTensor<int> &in, SimpleTensor<int> &roi, const Coordinates &coord); @@ -361,6 +583,8 @@ template void transpose_matrix(const SimpleTensor<half> &in, SimpleTensor<half> template void transpose_matrix(const SimpleTensor<int> &in, SimpleTensor<int> &out); template void transpose_matrix(const SimpleTensor<short> &in, SimpleTensor<short> &out); template void transpose_matrix(const SimpleTensor<char> &in, SimpleTensor<char> &out); +template void transpose_matrix(const SimpleTensor<int8_t> &in, SimpleTensor<int8_t> &out); +template void transpose_matrix(const SimpleTensor<uint8_t> &in, SimpleTensor<uint8_t> &out); template void matrix_multiply(const SimpleTensor<float> &a, const SimpleTensor<float> &b, SimpleTensor<float> &out); template void matrix_multiply(const SimpleTensor<half> &a, const SimpleTensor<half> &b, SimpleTensor<half> &out); diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h index 00e588e7b7..e044620556 100644 --- a/tests/validation/Helpers.h +++ b/tests/validation/Helpers.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023,2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,16 +21,19 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_VALIDATION_HELPERS_H -#define ARM_COMPUTE_TEST_VALIDATION_HELPERS_H +#ifndef ACL_TESTS_VALIDATION_HELPERS_H +#define ACL_TESTS_VALIDATION_HELPERS_H #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" +#include "arm_compute/function_info/ActivationLayerInfo.h" + #include "support/Half.h" #include "tests/Globals.h" #include "tests/SimpleTensor.h" -#include <math.h> +#include <cmath> +#include <cstdint> #include <random> #include <type_traits> #include <utility> @@ -50,6 +53,23 @@ template <> struct is_floating_point<half> : public std::true_type { }; +template <> +struct is_floating_point<bfloat16> : public std::true_type +{ +}; + +/** Helper struct to store the hints for + * - destination quantization info + * - minimum bias value + * - maximum bias value + * in quantized test construction. + */ +struct QuantizationHint +{ + QuantizationInfo q_info; + int32_t bias_min; + int32_t bias_max; +}; /** Helper function to get the testing range for each activation layer. * @@ -63,13 +83,13 @@ std::pair<T, T> get_activation_layer_test_bounds(ActivationLayerInfo::Activation { std::pair<T, T> bounds; - switch(data_type) + switch (data_type) { case DataType::F16: { using namespace half_float::literal; - switch(activation) + switch (activation) { case ActivationLayerInfo::ActivationFunction::TANH: case ActivationLayerInfo::ActivationFunction::SQUARE: @@ -89,7 +109,7 @@ std::pair<T, T> get_activation_layer_test_bounds(ActivationLayerInfo::Activation break; } case DataType::F32: - switch(activation) + switch (activation) { case ActivationLayerInfo::ActivationFunction::SOFT_RELU: // Reduce range as exponent overflows @@ -111,23 +131,6 @@ std::pair<T, T> get_activation_layer_test_bounds(ActivationLayerInfo::Activation return bounds; } -/** Calculate output tensor shape give a vector of input tensor to concatenate - * - * @param[in] input_shapes Shapes of the tensors to concatenate across depth. - * - * @return The shape of output concatenated tensor. - */ -TensorShape calculate_depth_concatenate_shape(const std::vector<TensorShape> &input_shapes); - -/** Calculate output tensor shape for the concatenate operation along a given axis - * - * @param[in] input_shapes Shapes of the tensors to concatenate across width. - * @param[in] axis Axis to use for the concatenate operation - * - * @return The shape of output concatenated tensor. - */ -TensorShape calculate_concatenate_shape(const std::vector<TensorShape> &input_shapes, size_t axis); - /** Convert an asymmetric quantized simple tensor into float using tensor quantization information. * * @param[in] src Quantized tensor. @@ -142,6 +145,7 @@ SimpleTensor<float> convert_from_asymmetric(const SimpleTensor<T> &src); * @param[in] src Float tensor. * @param[in] quantization_info Quantification information. * + * \relates arm_compute::test::SimpleTensor * @return Quantized tensor. */ template <typename T> @@ -160,7 +164,7 @@ SimpleTensor<float> convert_from_symmetric(const SimpleTensor<T> &src); * * @param[in] src Float tensor. * @param[in] quantization_info Quantification information. - * + * \relates arm_compute::test::SimpleTensor * @return Quantized tensor. */ template <typename T> @@ -228,7 +232,8 @@ std::pair<int, int> get_quantized_qasymm8_signed_bounds(const QuantizationInfo & * @param[in] max Floating point maximum value to be quantized * @param[in] channel_id Channel id for per channel quantization info. */ -std::pair<int, int> get_symm_quantized_per_channel_bounds(const QuantizationInfo &quant_info, float min, float max, size_t channel_id = 0); +std::pair<int, int> +get_symm_quantized_per_channel_bounds(const QuantizationInfo &quant_info, float min, float max, size_t channel_id = 0); /** Add random padding along the X axis (between 1 and 16 columns per side) to all the input tensors. * This is used in our validation suite in order to simulate implicit padding addition after configuring, but before allocating. @@ -239,8 +244,71 @@ std::pair<int, int> get_symm_quantized_per_channel_bounds(const QuantizationInfo * * @note This function adds padding to the input tensors only if data_layout == DataLayout::NHWC */ -void add_padding_x(std::initializer_list<ITensor *> tensors, const DataLayout &data_layout = DataLayout::NHWC, bool only_right_pad = false); +void add_padding_x(std::initializer_list<ITensor *> tensors, + const DataLayout &data_layout = DataLayout::NHWC, + bool only_right_pad = false); + +/** For 2d convolution, given the Lhs/Rhs matrix quantization informations and the convolution dimension, + * calculate a suitable output quantization and suggested bias range for obtaining non-saturated outputs with high probability. + * + * @param[in] in_q_info Input matrix quantization info + * @param[in] weight_q_info Weights matrix quantization info + * @param[in] height Height of the weights tensor + * @param[in] width Width of the weights tensors + * @param[in] channels Number of input channels + * @param[in] data_type data type, only QASYMM8, QASYMM8_SIGNED are supported + * @param[in] bias_fraction see @ref suggest_mac_dst_q_info_and_bias() for explanation + * + * @return QuantizationHint object containing the suggested output quantization info and min/max bias range + */ +QuantizationHint suggest_conv_dst_q_info_and_bias(const QuantizationInfo &in_q_info, + const QuantizationInfo &weight_q_info, + int32_t height, + int32_t width, + int32_t channels, + DataType data_type, + float bias_fraction); + +/** For a matrix multiplication, given the Lhs/Rhs matrix quantization informations and the matrix multiplication dimensions, + * calculate a suitable output quantization and suggested bias range for obtaining non-saturated outputs with high probability. + * + * @param[in] lhs_q_info Lhs matrix quantization info + * @param[in] rhs_q_info Rhs matrix quantization info + * @param[in] m Number of rows of Lhs matrix + * @param[in] n Number of columns of Rhs Matrix + * @param[in] k Number of rows/columns of Rhs/Lhs Matrix + * @param[in] data_type data type, only QASYMM8, QASYMM8_SIGNED are supported + * @param[in] bias_fraction see @ref suggest_mac_dst_q_info_and_bias() for explanation + * + * @return QuantizationHint object containing the suggested output quantization info and min/max bias range + */ +QuantizationHint suggest_matmul_dst_q_info_and_bias(const QuantizationInfo &lhs_q_info, + const QuantizationInfo &rhs_q_info, + int32_t m, + int32_t n, + int32_t k, + DataType data_type, + float bias_fraction); + +/** For a multiply-accumulate (mac), given the Lhs/Rhs vector quantization informations and the dot product dimensions, + * calculate a suitable output quantization and suggested bias range for obtaining non-saturated outputs with high probability. + * + * @param[in] lhs_q_info Lhs matrix quantization info + * @param[in] rhs_q_info Rhs matrix quantization info + * @param[in] k number of accumulations taking place in the sum, i.e. c_k = sum_k(a_k * b_k) + * @param[in] data_type data type, only QASYMM8, QASYMM8_SIGNED are supported + * @param[in] bias_fraction the fraction of bias amplitude compared to integer accummulation. + * @param[in] num_sd (Optional) number of standard deviations we allow from the mean. Default value is 2. + * + * @return QuantizationHint object containing the suggested output quantization info and min/max bias range + */ +QuantizationHint suggest_mac_dst_q_info_and_bias(const QuantizationInfo &lhs_q_info, + const QuantizationInfo &rhs_q_info, + int32_t k, + DataType data_type, + float bias_fraction, + int num_sd = 2); } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_VALIDATION_HELPERS_H */ +#endif // ACL_TESTS_VALIDATION_HELPERS_H diff --git a/tests/validation/NEON/ActivationLayer.cpp b/tests/validation/NEON/ActivationLayer.cpp index 111e969bae..73f5de68ac 100644 --- a/tests/validation/NEON/ActivationLayer.cpp +++ b/tests/validation/NEON/ActivationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,10 +23,13 @@ */ #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/Traits.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NEActivationLayer.h" #include "arm_compute/runtime/RuntimeContext.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuActivationKernel.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/ActivationFunctionsDataset.h" @@ -37,7 +40,8 @@ #include "tests/validation/Validation.h" #include "tests/validation/fixtures/ActivationLayerFixture.h" -#include "support/Requires.h" +#include "arm_compute/Acl.hpp" +#include "support/AclRequires.h" namespace arm_compute { @@ -65,6 +69,8 @@ RelativeTolerance<float> relative_tolerance(DataType data_type, ActivationLayerI case ActivationLayerInfo::ActivationFunction::SQRT: case ActivationLayerInfo::ActivationFunction::TANH: case ActivationLayerInfo::ActivationFunction::HARD_SWISH: + case ActivationLayerInfo::ActivationFunction::SWISH: + case ActivationLayerInfo::ActivationFunction::GELU: switch(data_type) { case DataType::F16: @@ -107,6 +113,7 @@ AbsoluteTolerance<float> absolute_tolerance(DataType data_type, ActivationLayerI case ActivationLayerInfo::ActivationFunction::LOGISTIC: case ActivationLayerInfo::ActivationFunction::SQRT: case ActivationLayerInfo::ActivationFunction::TANH: + case ActivationLayerInfo::ActivationFunction::SWISH: case ActivationLayerInfo::ActivationFunction::HARD_SWISH: switch(data_type) { @@ -169,7 +176,8 @@ const auto CNNDataTypes = framework::dataset::make("DataType", DataType::F32, }); -const auto NeonActivationFunctionsDataset = concat(datasets::ActivationFunctions(), framework::dataset::make("ActivationFunction", ActivationLayerInfo::ActivationFunction::HARD_SWISH)); +const auto NeonActivationFunctionsDataset = concat(datasets::ActivationFunctions(), + framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::HARD_SWISH, ActivationLayerInfo::ActivationFunction::SWISH })); /** Input data sets. */ const auto ActivationDataset = combine(combine(framework::dataset::make("InPlace", { false, true }), NeonActivationFunctionsDataset), framework::dataset::make("AlphaBeta", { 0.5f, 1.f })); @@ -215,6 +223,48 @@ void test_float_sqrt_boundary_value() TEST_SUITE(NEON) TEST_SUITE(ActivationLayer) +/** Test case for memory injection in @ref cpu::CpuWinogradConv2d. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(ActivationAPI, framework::DatasetMode::ALL) +{ + acl::StatusCode err = acl::StatusCode::Success; + + // Create context & Queue + acl::Context ctx(acl::Target::Cpu, &err); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); + + acl::Queue queue(ctx, &err); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); + + // Create activation operator + acl::TensorDescriptor src_info({ 2, 3 }, acl::DataType::Float32); + acl::TensorDescriptor dst_info({ 2, 3 }, acl::DataType::Float32); + acl::ActivationDesc desc{ AclRelu, 6.f, 0.f, false }; + + acl::Activation act(ctx, src_info, dst_info, desc, &err); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); + + // Create tensors and feed + acl::Tensor src(ctx, src_info, &err); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); + acl::Tensor dst(ctx, dst_info, &err); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); + + acl::TensorPack pack(ctx); + err = pack.add(src, ACL_SRC); + err = pack.add(dst, ACL_DST); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); + + // Execute operator + err = act.run(queue, pack); + ARM_COMPUTE_ASSERT(err == acl::StatusCode::Success); +} + // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( @@ -236,6 +286,49 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( bool is_valid = bool(NEActivationLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), act_info)); ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); } + +DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, concat(concat( + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED, + DataType::QSYMM16 + })), + combine(framework::dataset::make("CpuExt", std::string("SVE")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + }))), + combine(framework::dataset::make("CpuExt", std::string("SVE2")), + framework::dataset::make("DataType", { DataType::QASYMM8, + DataType::QASYMM8_SIGNED, + DataType::QSYMM16 + }))), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.sve = (cpu_ext == "SVE"); + cpu_isa.sve2 = (cpu_ext == "SVE2"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuActivationKernel::get_implementation(ActivationDataTypeISASelectorData{data_type, CPUModel::GENERIC, cpu_isa,ActivationLayerInfo::ActivationFunction::BOUNDED_RELU}, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_activation"; + if( data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED) + { +#ifdef __aarch64__ + expected = "neon_q8_activation_lut"; +#else // __aarch64__ + expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_activation"; +#endif // __aarch64__ + } + std::string actual = selected_impl->name; + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} // clang-format on // *INDENT-ON* @@ -316,9 +409,12 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerQuantizedFixture<int8_t>, fram TEST_SUITE_END() // QASYMM8_SIGNED /** Input data sets. */ -const auto Int16QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::LOGISTIC, - ActivationLayerInfo::ActivationFunction::TANH - }); +const auto Int16QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationFunction", +{ + ActivationLayerInfo::ActivationFunction::LOGISTIC, + ActivationLayerInfo::ActivationFunction::TANH, + ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, +}); const auto Int16QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false }), Int16QuantizedActivationFunctionsDataset), framework::dataset::make("AlphaBeta", { 0.5f, 1.f })); diff --git a/tests/validation/NEON/AddMulAdd.cpp b/tests/validation/NEON/AddMulAdd.cpp new file mode 100644 index 0000000000..77e3d80fe6 --- /dev/null +++ b/tests/validation/NEON/AddMulAdd.cpp @@ -0,0 +1,234 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifdef __aarch64__ + +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/NEON/functions/NEAddMulAdd.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" + +#include "tests/NEON/Accessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/AddMulAddFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */ +const AbsoluteTolerance<half> tolerance_fp16(half(0.1f)); /**< Tolerance for 16-bit floating point tests */ +constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance for quantized tests */ + +const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + + // Boundaries are aligned with Quantized Data ranges -- DOUBLE check before changing + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, -2.f) +}); + +// QASYMM8 test quantizations +const auto qasymm8_input1_qinfo_set = framework::dataset::make("Input1QInfo", { QuantizationInfo(0.1, 10) }); // Representable Range: [-1, 24.5] +const auto qasymm8_input2_qinfo_set = framework::dataset::make("Input2QInfo", { QuantizationInfo(0.2, 60) }); // Representable Range: [-12, 39] +const auto qasymm8_bn_mul_qinfo_set = framework::dataset::make("BnMulInfo", { QuantizationInfo(0.001, 55) }); // Representable Range: [-0.11, 0.2] +const auto qasymm8_bn_add_qinfo_set = framework::dataset::make("BnAddInfo", { QuantizationInfo(0.02, 20) }); // Representable Range: [-0.4, 4.7] + +// Representable Range: [-9.36, 51.84], Expected F32 range: [-13, 63.5], leaving some space for saturation +const auto qasymm8_add_output_qinfo_set = framework::dataset::make("AddOutputInfo", { QuantizationInfo(0.24, 39) }); + +// Representable Range: [-4.8, 10.5], Expected FP32 range: [-6.985, 12.7], leaving some space for saturation +// This range also makes sense with the activation boundaries above, i.e. [-2, 8] for LU_BOUNDED_RELU and [0, 6] for BOUNDED_RELU +const auto qasymm8_final_output_qinfo_set = framework::dataset::make("FinalOutputInfo", { QuantizationInfo(0.06, 80) }); + +// QASYMM8_SIGNED test quantizations +const auto qasymm8_signed_input1_qinfo_set = framework::dataset::make("Input1QInfo", { QuantizationInfo(0.1, 10) }); // Representable Range: [-13.8, 11.7] +const auto qasymm8_signed_input2_qinfo_set = framework::dataset::make("Input2QInfo", { QuantizationInfo(0.2, -60) }); // Representable Range: [-13.6, 39.4] +const auto qasymm8_signed_bn_mul_qinfo_set = framework::dataset::make("BnMulInfo", { QuantizationInfo(0.001, 55) }); // Representable Range: [-0.183, 0.072] +const auto qasymm8_signed_bn_add_qinfo_set = framework::dataset::make("BnAddInfo", { QuantizationInfo(0.4, -120) }); // Representable Range: [-0.32, 9.08] + +// Representable Range: [-21.36, 39.84], Expected F32 range: [-27.4, 51.1], leaving some space for saturation +const auto qasymm8_signed_add_output_qinfo_set = framework::dataset::make("AddOutputInfo", { QuantizationInfo(0.24, -39) }); + +// Representable Range: [-4.8, 10.5], Expected FP32 range: [-9.6713, 14.0942], leaving some space for saturation +// This range also makes sense with the activation boundaries above, i.e. [-2, 8] for LU_BOUNDED_RELU and [0, 6] for BOUNDED_RELU +const auto qasymm8_signed_final_output_qinfo_set = framework::dataset::make("FinalOutputInfo", { QuantizationInfo(0.06, -48) }); + +} // namespace + +TEST_SUITE(NEON) +TEST_SUITE(AddMulAdd) + +template <typename T> +using NEAddMulAddFloatFixture = AddMulAddFloatValidationFixture<Tensor, Accessor, NEAddMulAdd, T, true>; + +template <typename T> +using NEAddMulAddFloatFixtureWoIntermOut = AddMulAddFloatValidationFixture<Tensor, Accessor, NEAddMulAdd, T, false>; + +TEST_SUITE(Float) + +TEST_SUITE(F32) +FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulAddFloatFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::F32)), + ActivationFunctionsDataset)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance + validate(Accessor(_target), _reference, tolerance_fp32); +} + +// This test is to stress the case when there is no intermediate output required (i.e. nullptr) +FIXTURE_DATA_TEST_CASE(RunSmallWithoutIntermOutput, NEAddMulAddFloatFixtureWoIntermOut<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ActivationInfo", { ActivationLayerInfo() }))) +{ + // Validate outputs + validate(Accessor(_target), _reference, tolerance_fp32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulAddFloatFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), + framework::dataset::make("DataType", DataType::F32)), + ActivationFunctionsDataset)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance + validate(Accessor(_target), _reference, tolerance_fp32); +} + +TEST_SUITE_END() // F32 + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +TEST_SUITE(F16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulAddFloatFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::F16)), + ActivationFunctionsDataset)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance + validate(Accessor(_target), _reference, tolerance_fp16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulAddFloatFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), + framework::dataset::make("DataType", DataType::F16)), + ActivationFunctionsDataset)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference); // Arithmetic Addition has more strict tolerance + validate(Accessor(_target), _reference, tolerance_fp16); +} +TEST_SUITE_END() // F16 +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +TEST_SUITE_END() // Float + +template <typename T> +using NEAddMulQuantizedFixture = AddMulAddQuantizedValidationFixture<Tensor, Accessor, NEAddMulAdd, T, true>; + +template <typename T> +using NEAddMulAddQuantizedFixtureWoIntermOut = AddMulAddQuantizedValidationFixture<Tensor, Accessor, NEAddMulAdd, T, false>; + +TEST_SUITE(Quantized) + +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + ActivationFunctionsDataset), + qasymm8_input1_qinfo_set), + qasymm8_input2_qinfo_set), + qasymm8_bn_mul_qinfo_set), + qasymm8_bn_add_qinfo_set), + qasymm8_add_output_qinfo_set), + qasymm8_final_output_qinfo_set)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference, tolerance_quant); + validate(Accessor(_target), _reference, tolerance_quant); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(datasets::LargeShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + ActivationFunctionsDataset), + qasymm8_input1_qinfo_set), + qasymm8_input2_qinfo_set), + qasymm8_bn_mul_qinfo_set), + qasymm8_bn_add_qinfo_set), + qasymm8_add_output_qinfo_set), + qasymm8_final_output_qinfo_set)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference, tolerance_quant); + validate(Accessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NEAddMulQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + ActivationFunctionsDataset), + qasymm8_signed_input1_qinfo_set), + qasymm8_signed_input2_qinfo_set), + qasymm8_signed_bn_mul_qinfo_set), + qasymm8_signed_bn_add_qinfo_set), + qasymm8_signed_add_output_qinfo_set), + qasymm8_signed_final_output_qinfo_set)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference, tolerance_quant); + validate(Accessor(_target), _reference, tolerance_quant); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEAddMulQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(combine(datasets::LargeShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + ActivationFunctionsDataset), + qasymm8_signed_input1_qinfo_set), + qasymm8_signed_input2_qinfo_set), + qasymm8_signed_bn_mul_qinfo_set), + qasymm8_signed_bn_add_qinfo_set), + qasymm8_signed_add_output_qinfo_set), + qasymm8_signed_final_output_qinfo_set)) +{ + // Validate outputs + validate(Accessor(_interm_target), _interm_reference, tolerance_quant); + validate(Accessor(_target), _reference, tolerance_quant); +} +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE_END() // Quantized + +TEST_SUITE_END() // AddMulAdd +TEST_SUITE_END() // NEON +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // __aarch64__ diff --git a/tests/validation/NEON/ArgMinMax.cpp b/tests/validation/NEON/ArgMinMax.cpp index 0a4071076a..91b8128dea 100644 --- a/tests/validation/NEON/ArgMinMax.cpp +++ b/tests/validation/NEON/ArgMinMax.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,6 +43,27 @@ namespace test { namespace validation { +namespace +{ +const auto OpsDataset = framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }); +const auto AxisDataset = framework::dataset::make("Axis", { 0, 1, 2, 3 }); +const auto QInfoDataset = framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }); + +const auto ArgMinMaxSmallDatasetAxis0 = framework::dataset::make("Shape", +{ + TensorShape{ 1U, 5U }, + TensorShape{ 2U, 3U }, + TensorShape{ 1U }, + TensorShape{ 3U }, + TensorShape{ 2U }, + TensorShape{ 5U }, + TensorShape{ 17U }, + TensorShape{ 15U, 2U }, +}); +using ArgMinMaxSmallDataset = datasets::Small4DShapes; +using ArgMinMaxLargeDataset = datasets::Large4DShapes; +} + TEST_SUITE(NEON) TEST_SUITE(ArgMinMax) @@ -70,23 +91,50 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( // clang-format on // *INDENT-ON* -template <typename T> -using NEArgMinMaxValidationFixture = ArgMinMaxValidationFixture<Tensor, Accessor, NEArgMinMaxLayer, T>; +template <typename T1, typename T2> +using NEArgMinMaxValidationFixture = ArgMinMaxValidationFixture<Tensor, Accessor, NEArgMinMaxLayer, T1, T2>; + +using NEArgMinMaxValidationFixture_S32_S32 = NEArgMinMaxValidationFixture<int32_t, int32_t>; +using NEArgMinMaxValidationFixture_F16_S32 = NEArgMinMaxValidationFixture<half, int32_t>; +using NEArgMinMaxValidationFixture_F32_S32 = NEArgMinMaxValidationFixture<float, int32_t>; +#ifdef __aarch64__ +using NEArgMinMaxValidationFixture_F32_S64 = NEArgMinMaxValidationFixture<float, int64_t>; +#endif // __aarch64__ TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxValidationFixture<int32_t>, +FIXTURE_DATA_TEST_CASE(RunSmallAxis0, + NEArgMinMaxValidationFixture_S32_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::S32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxSmallDatasetAxis0, + framework::dataset::make("DataTypeIn", DataType::S32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + framework::dataset::make("Axis", { 0 })), + OpsDataset)) { // Validate output validate(Accessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmall, + NEArgMinMaxValidationFixture_S32_S32, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(ArgMinMaxSmallDataset(), + framework::dataset::make("DataTypeIn", DataType::S32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxValidationFixture<int32_t>, + NEArgMinMaxValidationFixture_S32_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::S32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxLargeDataset(), + framework::dataset::make("DataTypeIn", DataType::S32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(Accessor(_target), _reference); @@ -97,18 +145,26 @@ TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxValidationFixture<half>, + NEArgMinMaxValidationFixture_F16_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxSmallDataset(), + framework::dataset::make("DataTypeIn", DataType::F16)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxValidationFixture<half>, + NEArgMinMaxValidationFixture_F16_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxLargeDataset(), + framework::dataset::make("DataTypeIn", DataType::F16)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(Accessor(_target), _reference); @@ -118,18 +174,41 @@ TEST_SUITE_END() // FP16 TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxValidationFixture<float>, + NEArgMinMaxValidationFixture_F32_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxSmallDataset(), + framework::dataset::make("DataTypeIn", DataType::F32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(Accessor(_target), _reference); } +#ifdef __aarch64__ +FIXTURE_DATA_TEST_CASE(RunSmall_F32_S64, + NEArgMinMaxValidationFixture_F32_S64, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(ArgMinMaxSmallDataset(), + framework::dataset::make("DataTypeIn", DataType::F32)), + framework::dataset::make("DataTypeOut", DataType::S64)), + AxisDataset), + OpsDataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +#endif // __aarch64__ + FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxValidationFixture<float>, + NEArgMinMaxValidationFixture_F32_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX }))) + combine(combine(combine(combine(ArgMinMaxLargeDataset(), + framework::dataset::make("DataTypeIn", DataType::F32)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset)) { // Validate output validate(Accessor(_target), _reference); @@ -137,27 +216,35 @@ FIXTURE_DATA_TEST_CASE(RunLarge, TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float -template <typename T> -using NEArgMinMaxQuantizedValidationFixture = ArgMinMaxValidationQuantizedFixture<Tensor, Accessor, NEArgMinMaxLayer, T>; +template <typename T1, typename T2> +using NEArgMinMaxQuantizedValidationFixture = ArgMinMaxValidationQuantizedFixture<Tensor, Accessor, NEArgMinMaxLayer, T1, T2>; + +using NEArgMinMaxQuantizedValidationFixture_U8_S32 = NEArgMinMaxQuantizedValidationFixture<uint8_t, int32_t>; +using NEArgMinMaxQuantizedValidationFixture_S8_S32 = NEArgMinMaxQuantizedValidationFixture<int8_t, int32_t>; TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxQuantizedValidationFixture<uint8_t>, + NEArgMinMaxQuantizedValidationFixture_U8_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxSmallDataset(), + framework::dataset::make("DataTypeIn", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(Accessor(_target), _reference); } - FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxQuantizedValidationFixture<uint8_t>, + NEArgMinMaxQuantizedValidationFixture_U8_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxLargeDataset(), + framework::dataset::make("DataTypeIn", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(Accessor(_target), _reference); @@ -166,22 +253,27 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, - NEArgMinMaxQuantizedValidationFixture<int8_t>, + NEArgMinMaxQuantizedValidationFixture_S8_S32, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(datasets::Small4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 127.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxSmallDataset(), + framework::dataset::make("DataTypeIn", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(Accessor(_target), _reference); } - FIXTURE_DATA_TEST_CASE(RunLarge, - NEArgMinMaxQuantizedValidationFixture<int8_t>, + NEArgMinMaxQuantizedValidationFixture_S8_S32, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(datasets::Large4DShapes(), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("Axis", { 0, 1, 2, 3 })), - framework::dataset::make("Operation", { ReductionOperation::ARG_IDX_MIN, ReductionOperation::ARG_IDX_MAX })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 127.f, 20) }))) + combine(combine(combine(combine(combine(ArgMinMaxLargeDataset(), + framework::dataset::make("DataTypeIn", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeOut", DataType::S32)), + AxisDataset), + OpsDataset), + QInfoDataset)) { // Validate output validate(Accessor(_target), _reference); diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp index ea6656eefe..535c3e634e 100644 --- a/tests/validation/NEON/ArithmeticAddition.cpp +++ b/tests/validation/NEON/ArithmeticAddition.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,9 +22,12 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NEArithmeticAddition.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuAddKernel.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/ConvertPolicyDataset.h" @@ -48,26 +51,8 @@ constexpr AbsoluteTolerance<float> tolerance_quant(1); /**< Tolerance value for #else // !defined(__aarch64__) || defined(ENABLE_SVE) constexpr AbsoluteTolerance<float> tolerance_quant(0); #endif // !defined(__aarch64__) || defined(ENABLE_SVE) - -/** Input data sets **/ -const auto ArithmeticAdditionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), framework::dataset::make("DataType", DataType::U8)), framework::dataset::make("DataType", - DataType::U8)); -const auto ArithmeticAdditionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), framework::dataset::make("DataType", DataType::S16)), - framework::dataset::make("DataType", DataType::S16)); -const auto ArithmeticAdditionS32Dataset = combine(combine(framework::dataset::make("DataType", { DataType::S32 }), framework::dataset::make("DataType", DataType::S32)), - framework::dataset::make("DataType", DataType::S32)); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -const auto ArithmeticAdditionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("DataType", DataType::F16)); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -const auto ArithmeticAdditionFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataType", DataType::F32)); -const auto ArithmeticAdditionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataType", DataType::QASYMM8)); -const auto ArithmeticAdditionQASYMM8SIGNEDDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)); -const auto ArithmeticAdditionQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), framework::dataset::make("DataType", DataType::QSYMM16)), - framework::dataset::make("DataType", DataType::QSYMM16)); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(NEON) @@ -79,25 +64,22 @@ using NEArithmeticAdditionFixture = ArithmeticAdditionValidationFixture<Tensor, // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::U8), // Unsupported broadcast TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32),// Mismatching shapes }), - framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(1U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, false, false, false})), + framework::dataset::make("Expected", { true, false, false, false})), input1_info, input2_info, output_info, expected) { Status s = NEArithmeticAddition::validate(&input1_info.clone()->set_is_resizable(false), @@ -106,6 +88,63 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( ConvertPolicy::WRAP); ARM_COMPUTE_EXPECT(bool(s) == expected, framework::LogLevel::ERRORS); } + +DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, concat(concat( + combine(combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::U8, + DataType::S16, + DataType::S32, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED, + DataType::QSYMM16 + })), + framework::dataset::make("CanUseFixedpoint", {true, false})), + combine(combine(framework::dataset::make("CpuExt", std::string("SVE")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::U8, + DataType::S16, + DataType::S32 + })), + framework::dataset::make("CanUseFixedpoint", {true, false}))), + combine(combine(framework::dataset::make("CpuExt", std::string("SVE2")), + framework::dataset::make("DataType", { DataType::QASYMM8, + DataType::QASYMM8_SIGNED, + DataType::QSYMM16 + })), + framework::dataset::make("CanUseFixedpoint", {true, false}))), + cpu_ext, data_type, can_use_fixedpoint) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.sve = (cpu_ext == "SVE"); + cpu_isa.sve2 = (cpu_ext == "SVE2"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuAddKernel::get_implementation(CpuAddKernelDataTypeISASelectorData{data_type, cpu_isa, can_use_fixedpoint}, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + bool qasymm8_any = (data_type == DataType::QASYMM8 || data_type == DataType::QASYMM8_SIGNED); + + std::string expected; + if(qasymm8_any && can_use_fixedpoint) + { + expected = "neon_" + cpu_impl_dt(data_type) + "_add_fixedpoint"; + } + else + { + expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_add"; + } + + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} // clang-format on // *INDENT-ON* @@ -127,8 +166,10 @@ TEST_CASE(NoPaddingAdded, framework::DatasetMode::PRECOMMIT) TEST_SUITE(Integer) TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticAdditionU8Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -136,15 +177,19 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<uint8_t>, framework TEST_SUITE_END() // U8 TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticAdditionS16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticAdditionS16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -152,8 +197,10 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<int16_t>, framework TEST_SUITE_END() // S16 TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<int32_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ArithmeticAdditionS32Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<int32_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S32)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -164,8 +211,9 @@ TEST_SUITE_END() // Integer TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ArithmeticAdditionFP16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -174,15 +222,19 @@ TEST_SUITE_END() // F16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ArithmeticAdditionFP32Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::F32)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ArithmeticAdditionFP32Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::F32)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -191,17 +243,19 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<float>, framework:: template <typename T> using NEArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<Tensor, Accessor, NEArithmeticAddition, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(), - ArithmeticAdditionFP32Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapesBroadcast(), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(), - ArithmeticAdditionFP32Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapesBroadcast(), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -220,11 +274,12 @@ TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQASYMM8Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), - framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), - framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_quant); @@ -235,22 +290,24 @@ TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionQuantizedFixture<int8_t>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQASYMM8SIGNEDDataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), - framework::dataset::make("Src0QInfo", { QuantizationInfo(0.5f, 20) })), - framework::dataset::make("Src1QInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(0.5f, 5) }))) + combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("Src0QInfo", { QuantizationInfo(0.5f, 20) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(0.5f, 10) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(0.5f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_quant); } -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionQuantizedBroadcastFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine( - datasets::SmallShapesBroadcast(), ArithmeticAdditionQASYMM8SIGNEDDataset), +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionQuantizedBroadcastFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine( + datasets::SmallShapesBroadcast(), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), framework::dataset::make("Src0QInfo", { QuantizationInfo(0.5f, 20) })), framework::dataset::make("Src1QInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(0.5f, 5) }))) + framework::dataset::make("OutQInfo", { QuantizationInfo(0.5f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_quant); @@ -261,11 +318,12 @@ TEST_SUITE(QSYMM16) FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionQuantizedFixture<int16_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQSYMM16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), - framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), - framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), - framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) }))) + combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QSYMM16)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + framework::dataset::make("Src0QInfo", { QuantizationInfo(1.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), + framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0), QuantizationInfo(5.f / 32768.f, 0) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 0) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_quant); diff --git a/tests/validation/NEON/ArithmeticSubtraction.cpp b/tests/validation/NEON/ArithmeticSubtraction.cpp index 7a36893445..8886ca2db5 100644 --- a/tests/validation/NEON/ArithmeticSubtraction.cpp +++ b/tests/validation/NEON/ArithmeticSubtraction.cpp @@ -50,45 +50,16 @@ constexpr AbsoluteTolerance<float> tolerance_qasymm8(1); /**< Tolerance value fo #endif //__aarch64__ constexpr AbsoluteTolerance<int16_t> tolerance_qsymm16(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ -/** Input data sets **/ -const auto ArithmeticSubtractionQASYMM8Dataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataType", DataType::QASYMM8)); - -const auto ArithmeticSubtractionQASYMM8SIGNEDDataset = combine(combine(framework::dataset::make("DataType", DataType::QASYMM8_SIGNED), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)); - -const auto ArithmeticSubtractionQSYMM16Dataset = combine(combine(framework::dataset::make("DataType", DataType::QSYMM16), - framework::dataset::make("DataType", DataType::QSYMM16)), - framework::dataset::make("DataType", DataType::QSYMM16)); - -const auto ArithmeticSubtractionU8Dataset = combine(combine(framework::dataset::make("DataType", DataType::U8), - framework::dataset::make("DataType", DataType::U8)), - framework::dataset::make("DataType", DataType::U8)); - -const auto ArithmeticSubtractionS16Dataset = combine(combine(framework::dataset::make("DataType", { DataType::U8, DataType::S16 }), - framework::dataset::make("DataType", DataType::S16)), - framework::dataset::make("DataType", DataType::S16)); - -const auto ArithmeticSubtractionS32Dataset = combine(combine(framework::dataset::make("DataType", DataType::S32), - framework::dataset::make("DataType", DataType::S32)), - framework::dataset::make("DataType", DataType::S32)); -#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -const auto ArithmeticSubtractionFP16Dataset = combine(combine(framework::dataset::make("DataType", DataType::F16), - framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("DataType", DataType::F16)); -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -const auto ArithmeticSubtractionFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataType", DataType::F32)); - +// Quantization Infomation DataSet const auto ArithmeticSubtractionQuantizationInfoDataset = combine(combine(framework::dataset::make("QuantizationInfoIn1", { QuantizationInfo(10, 120) }), framework::dataset::make("QuantizationInfoIn2", { QuantizationInfo(20, 110) })), framework::dataset::make("QuantizationInfoOut", { QuantizationInfo(15, 125) })); const auto ArithmeticSubtractionQuantizationInfoSignedDataset = combine(combine(framework::dataset::make("QuantizationInfoIn1", { QuantizationInfo(0.5f, 10) }), framework::dataset::make("QuantizationInfoIn2", { QuantizationInfo(0.5f, 20) })), framework::dataset::make("QuantizationInfoOut", { QuantizationInfo(0.5f, 50) })); +const auto ArithmeticSubtractionQuantizationInfoSignedInPlaceDataset = combine(combine(framework::dataset::make("QuantizationInfoIn1", { QuantizationInfo(0.8f, 10) }), + framework::dataset::make("QuantizationInfoIn2", { QuantizationInfo(0.8f, 10) })), + framework::dataset::make("QuantizationInfoOut", { QuantizationInfo(0.8f, 10) })); const auto ArithmeticSubtractionQuantizationInfoSymmetric = combine(combine(framework::dataset::make("QuantizationInfoIn1", { QuantizationInfo(0.3f, 0) }), framework::dataset::make("QuantizationInfoIn2", { QuantizationInfo(0.7f, 0) })), framework::dataset::make("QuantizationInfoOut", { QuantizationInfo(0.2f, 0) })); @@ -105,35 +76,31 @@ using NEArithmeticSubtractionFixture = ArithmeticSubtractionValidationFixture<Te // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( - framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input1Info", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid data type combination TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::QASYMM8), // Mismatching types TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Invalid convert policy }), - framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("Input2Info",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), - TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), })), - framework::dataset::make("ConvertPolicy",{ ConvertPolicy::WRAP, - ConvertPolicy::SATURATE, - ConvertPolicy::SATURATE, - ConvertPolicy::WRAP, - ConvertPolicy::WRAP, - ConvertPolicy::WRAP, + framework::dataset::make("ConvertPolicy",{ ConvertPolicy::SATURATE, + ConvertPolicy::SATURATE, + ConvertPolicy::WRAP, + ConvertPolicy::WRAP, + ConvertPolicy::WRAP, })), - framework::dataset::make("Expected", { true, true, false, false, false, false})), + framework::dataset::make("Expected", { true, false, false, false, false})), input1_info, input2_info, output_info, policy, expected) { ARM_COMPUTE_EXPECT(bool(NEArithmeticSubtraction::validate(&input1_info.clone()->set_is_resizable(false), &input2_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), policy)) == expected, framework::LogLevel::ERRORS); @@ -194,7 +161,8 @@ TEST_CASE(InvalidBroadcastBoth, framework::DatasetMode::ALL) TEST_SUITE_END() // InPlaceValidate TEST_SUITE(U8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionU8Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -210,10 +178,11 @@ using NEArithmeticSubtractionQSYMM16Fixture = ArithmeticSubtracti TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionQASYMM8Fixture, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionQASYMM8Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionQASYMM8Fixture, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::QASYMM8)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), ArithmeticSubtractionQuantizationInfoDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -222,19 +191,17 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionQASYMM8SignedFixture, framework::DatasetMode::ALL, combine(combine(combine(combine( - datasets::SmallShapes(), - ArithmeticSubtractionQASYMM8SIGNEDDataset), + datasets::SmallShapes(), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), ArithmeticSubtractionQuantizationInfoSignedDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } - FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticSubtractionQASYMM8SignedBroadcastFixture, framework::DatasetMode::ALL, combine(combine(combine(combine( datasets::SmallShapesBroadcast(), - ArithmeticSubtractionQASYMM8SIGNEDDataset), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), ArithmeticSubtractionQuantizationInfoSignedDataset), OutOfPlaceDataSet)) @@ -242,12 +209,22 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticSubtractionQASYMM8SignedBr // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEArithmeticSubtractionQASYMM8SignedBroadcastFixture, framework::DatasetMode::ALL, combine(combine(combine(combine( + datasets::TinyShapesBroadcastInplace(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), + ArithmeticSubtractionQuantizationInfoSignedInPlaceDataset), + InPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE(QSYMM16) FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionQSYMM16Fixture, framework::DatasetMode::ALL, combine(combine(combine(combine( datasets::SmallShapes(), - ArithmeticSubtractionQSYMM16Dataset), + framework::dataset::make("DataType", DataType::QSYMM16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE })), ArithmeticSubtractionQuantizationInfoSymmetric), OutOfPlaceDataSet)) @@ -259,7 +236,8 @@ TEST_SUITE_END() // QSYMM16 TEST_SUITE_END() // Quantized TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -267,7 +245,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int16_t>, framew validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS16Dataset), +FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -277,7 +256,8 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<int16_t>, framew TEST_SUITE_END() // S16 TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionS32Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -285,7 +265,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<int32_t>, framew validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<int32_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticSubtractionS32Dataset), +FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<int32_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -297,7 +278,8 @@ TEST_SUITE_END() // S32 TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionFP16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<half>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::F16)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -308,7 +290,8 @@ TEST_SUITE_END() // F16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), InPlaceDataSet)) { @@ -316,7 +299,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixture<float>, framewor validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticSubtractionFP32Dataset), +FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -328,7 +312,7 @@ template <typename T> using NEArithmeticSubtractionBroadcastFixture = ArithmeticSubtractionBroadcastValidationFixture<Tensor, Accessor, NEArithmeticSubtraction, T>; FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticSubtractionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapesBroadcast(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { @@ -337,7 +321,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticSubtractionBroadcastFixtur } FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticSubtractionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapesBroadcast(), - ArithmeticSubtractionFP32Dataset), + framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), OutOfPlaceDataSet)) { diff --git a/tests/validation/NEON/BatchNormalizationLayer.cpp b/tests/validation/NEON/BatchNormalizationLayer.cpp index a1ae6971f4..50eaf0c667 100644 --- a/tests/validation/NEON/BatchNormalizationLayer.cpp +++ b/tests/validation/NEON/BatchNormalizationLayer.cpp @@ -51,7 +51,7 @@ namespace RelativeTolerance<float> rel_tolerance_f32(0.05f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -constexpr AbsoluteTolerance<float> abs_tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr AbsoluteTolerance<float> abs_tolerance_f16(0.015f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC const auto act_infos = framework::dataset::make("ActivationInfo", diff --git a/tests/validation/NEON/BatchToSpaceLayer.cpp b/tests/validation/NEON/BatchToSpaceLayer.cpp index a305dcbcc4..8cf11b7b95 100644 --- a/tests/validation/NEON/BatchToSpaceLayer.cpp +++ b/tests/validation/NEON/BatchToSpaceLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -49,55 +49,38 @@ using NEBatchToSpaceLayerFixture = BatchToSpaceLayerValidationFixture<Tensor, Ac // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blockx > blocky - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blocky > blockx - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), // Mismatching data types - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), // Wrong data type block shape - TensorInfo(TensorShape(32U, 13U, 2U, 2U, 4U), 1, DataType::F32), // Wrong tensor shape - }), - framework::dataset::make("BlockShapeInfo",{ TensorInfo(TensorShape(2U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(2U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(2U, 4U), 1, DataType::S32), - TensorInfo(TensorShape(4U, 2U), 1, DataType::S32), - TensorInfo(TensorShape(2U, 2U), 1, DataType::F16), - TensorInfo(TensorShape(2U, 2U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(64U, 16U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 32U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F16), - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U, 2U, 2U), 1, DataType::F32), - })), - framework::dataset::make("Expected", { true, true, true, false, false, false})), - input_info, block_shape_info, output_info, expected) -{ - bool has_error = bool(NEBatchToSpaceLayer::validate(&input_info.clone()->set_is_resizable(false), &block_shape_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))); - ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS); -} -DATA_TEST_CASE(ValidateStatic, framework::DatasetMode::ALL, zip(zip(zip(zip( +DATA_TEST_CASE(ValidateStatic, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blockx > blocky - TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // blockx != blocky && blocky > blockx - TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Mismatching data types - TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Negative block shapes - TensorInfo(TensorShape(32U, 16U, 2U, 4U, 4U), 1, DataType::F32), // Wrong tensor shape + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Supported: blockx != blocky && blockx > blocky + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Supported: blockx != blocky && blocky > blockx + TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Invalid: Mismatching data types + TensorInfo(TensorShape(16U, 8U, 2U, 4U), 1, DataType::F32), // Invalid: Negative block shapes + TensorInfo(TensorShape(32U, 16U, 2U, 4U, 4U), 1, DataType::F32),// Unsupported tensor rank + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Invalid output tensor shape (invalid batch dimension) + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Invalid output tensor shape (invalid spatial dimension) + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Supported: correct tensor shape with cropping + TensorInfo(TensorShape(16U, 8U, 2U, 16U), 1, DataType::F32), // Invalid tensor shape with cropping }), - framework::dataset::make("BlockShapeX", { 2, 4, 2, 2, 2, 2 })), - framework::dataset::make("BlockShapeY", { 2, 2, 4, 2, -2, 2 })), + framework::dataset::make("BlockShapeX", { 2, 4, 2, 2, 2, 2, 2, 2, 2, 2 })), + framework::dataset::make("BlockShapeY", { 2, 2, 4, 2, -2, 2, 2, 2, 2, 2 })), + framework::dataset::make("CropInfo", { + CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{}, CropInfo{3, 2, 1, 3}, CropInfo{3, 2, 1, 3} + })), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(64U, 16U, 2U, 1U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 32U, 2U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(64U, 16U, 2U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 32U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F16), TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F32), TensorInfo(TensorShape(32U, 8U, 2U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 16U, 2U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(33U, 32U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(27, 12U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 16U, 2U, 4U), 1, DataType::F32), })), - framework::dataset::make("Expected", { true, true, true, false, false, false})), - input_info, block_shape_x, block_shape_y, output_info, expected) + framework::dataset::make("Expected", { true, true, true, false, false, false, false, false, true, false})), + input_info, block_shape_x, block_shape_y, crop_info, output_info, expected) { - bool has_error = bool(NEBatchToSpaceLayer::validate(&input_info.clone()->set_is_resizable(false), block_shape_x, block_shape_y, &output_info.clone()->set_is_resizable(false))); + bool has_error = bool(NEBatchToSpaceLayer::validate(&input_info.clone()->set_is_resizable(false), block_shape_x, block_shape_y, &output_info.clone()->set_is_resizable(false), crop_info)); ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS); } // clang-format on @@ -112,6 +95,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEBatchToSpaceLayerFixture<float>, framework::D // Validate output validate(Accessor(_target), _reference); } + +FIXTURE_DATA_TEST_CASE(RunSmallWithCropping, NEBatchToSpaceLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::SmallBatchToSpaceLayerWithCroppingDataset(), framework::dataset::make("DataType", + DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, NEBatchToSpaceLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeBatchToSpaceLayerDataset(), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) @@ -129,6 +122,15 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEBatchToSpaceLayerFixture<half>, framework::Da // Validate output validate(Accessor(_target), _reference); } +FIXTURE_DATA_TEST_CASE(RunSmallWithCropping, NEBatchToSpaceLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::SmallBatchToSpaceLayerWithCroppingDataset(), framework::dataset::make("DataType", + DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + FIXTURE_DATA_TEST_CASE(RunLarge, NEBatchToSpaceLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeBatchToSpaceLayerDataset(), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) diff --git a/tests/validation/NEON/Cast.cpp b/tests/validation/NEON/Cast.cpp index db73bea9cb..b56594546b 100644 --- a/tests/validation/NEON/Cast.cpp +++ b/tests/validation/NEON/Cast.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,9 +22,12 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NECast.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuCastKernel.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/ConvertPolicyDataset.h" @@ -34,7 +37,6 @@ #include "tests/framework/datasets/Datasets.h" #include "tests/validation/Validation.h" #include "tests/validation/fixtures/CastFixture.h" - namespace arm_compute { namespace test @@ -99,6 +101,11 @@ const auto CastF32toS32Dataset = combine(framework::dataset::make("Da const auto CastF32toQASYMM8Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::QASYMM8)); const auto CastF32toQASYMM8_SIGNEDDataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)); +// U64 +const auto CastU64toF32Dataset = combine(framework::dataset::make("DataType", DataType::U64), framework::dataset::make("DataType", DataType::F32)); + +// S64 +const auto CastS64toF32Dataset = combine(framework::dataset::make("DataType", DataType::S64), framework::dataset::make("DataType", DataType::F32)); } // namespace TEST_SUITE(NEON) @@ -106,6 +113,8 @@ TEST_SUITE(Cast) template <typename T> using NECastToU8Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, uint8_t>; template <typename T> +using NECastToS8Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, int8_t>; +template <typename T> using NECastToU16Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, uint16_t>; template <typename T> using NECastToS16Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, int16_t>; @@ -114,6 +123,10 @@ using NECastToU32Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, ui template <typename T> using NECastToS32Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, int32_t>; template <typename T> +using NECastToU64Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, uint64_t>; +template <typename T> +using NECastToS64Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, int64_t>; +template <typename T> using NECastToF16Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, half>; template <typename T> using NECastToF32Fixture = CastValidationFixture<Tensor, Accessor, NECast, T, float>; @@ -187,6 +200,66 @@ CAST_SUITE(F32_to_F16, DataType::F32, DataType::F16, NECastToF16Fixture<float>, CAST_SUITE(F32_to_S32, DataType::F32, DataType::S32, NECastToS32Fixture<float>, CastF32toS32Dataset, one_tolerance) CAST_SUITE(F32_to_U8, DataType::F32, DataType::S32, NECastToS32Fixture<float>, CastF32toS32Dataset, one_tolerance) +#ifdef __aarch64__ +// S64 +CAST_SUITE(S64_to_F32, DataType::S64, DataType::F32, NECastToF32Fixture<int64_t>, CastS64toF32Dataset, zero_tolerance) + +// U64 +CAST_SUITE(U64_to_F32, DataType::U64, DataType::F32, NECastToF32Fixture<uint64_t>, CastU64toF32Dataset, zero_tolerance) +#endif // __aarch64__ + +DATA_TEST_CASE(KernelSelectionDstFP16, framework::DatasetMode::ALL, + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", +{ + DataType::F16, + DataType::U8, + DataType::S32, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED, +})), +cpu_ext, data_type) +{ + using namespace cpu::kernels; + const CpuCastKernel::CastKernel *selected_impl; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.fp16 = true; + + selected_impl = CpuCastKernel::get_implementation(CastDataTypeISASelectorData{ data_type, DataType::F16, cpu_isa }, cpu::KernelSelectionType::Preferred); + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_cast"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + +DATA_TEST_CASE(KernelSelectionSrcFP32, framework::DatasetMode::ALL, + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", +{ + DataType::F16, +})), +cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuCastKernel::get_implementation(CastDataTypeISASelectorData{ DataType::F32, data_type, cpu_isa }, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_fp32_to_" + cpu_impl_dt(data_type) + "_cast"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + TEST_SUITE_END() // Cast TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/Col2Im.cpp b/tests/validation/NEON/Col2Im.cpp index 9139f0cca8..7eb8cbf0f6 100644 --- a/tests/validation/NEON/Col2Im.cpp +++ b/tests/validation/NEON/Col2Im.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,7 +22,7 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "src/core/NEON/kernels/NECol2ImKernel.h" +#include "src/cpu/kernels/CpuCol2ImKernel.h" #include "tests/NEON/Helper.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/framework/Asserts.h" @@ -39,7 +39,7 @@ namespace validation TEST_SUITE(NEON) TEST_SUITE(Col2Im) -using NECol2Im = NESynthetizeFunction<NECol2ImKernel>; +using CpuCol2Im = NESynthetizeFunction<cpu::kernels::CpuCol2ImKernel>; // *INDENT-OFF* // clang-format off @@ -59,7 +59,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( framework::dataset::make("Expected", { false, false, false, true })), input_info, output_info, convolved_width, convolved_height, expected) { - bool status = bool(NECol2Im::validate(&input_info, &output_info, Size2D(convolved_width, convolved_height))); + bool status = bool(CpuCol2Im::validate(&input_info, &output_info, Size2D(convolved_width, convolved_height))); ARM_COMPUTE_EXPECT(status == expected, framework::LogLevel::ERRORS); } // clang-format on diff --git a/tests/validation/NEON/Convolution3D.cpp b/tests/validation/NEON/Convolution3D.cpp new file mode 100644 index 0000000000..4185488742 --- /dev/null +++ b/tests/validation/NEON/Convolution3D.cpp @@ -0,0 +1,228 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/NEON/functions/NEConv3D.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" +#include "tests/NEON/Accessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/DirectConvolution3DFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +const RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2f)); /**< Relative tolerance value for FP16 types */ +const AbsoluteTolerance<float> abs_tolerance_f16(0.2f); /**< Absolute tolerance for FP16 types */ +constexpr float tolerance_num = 0.07f; /**< Tolerance number for the FP16 implementation */ +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance for quantized tests */ + +/** Activation function Dataset*/ +const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) +}); + +const auto data_precommit = combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + datasets::SmallDirectConv3DShapes(), + framework::dataset::make("StrideX", { 1, 5, 8 })), + framework::dataset::make("StrideY", { 1, 2, 3 })), + framework::dataset::make("StrideZ", { 1, 2, 1 })), + framework::dataset::make("PadX", { 0, 1, 2 })), + framework::dataset::make("PadY", { 0, 2, 1 })), + framework::dataset::make("PadZ", { 0, 3, 5 })), + framework::dataset::make("KernelWidth", { 3, 5, 9 })), + framework::dataset::make("KernelHeight", { 2, 1, 3 })), + framework::dataset::make("KernelDepth", { 1, 2, 3 })), + framework::dataset::make("NumKernels", { 2, 3, 8 })), + framework::dataset::make("HasBias", { true, false })), + ActivationFunctionsDataset); +} // namespace + +TEST_SUITE(NEON) +TEST_SUITE(Convolution3D) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Mismatching data type input/weights + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Mismatching input feature maps + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid weights dimensions + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NHWC), // Invalid data layout + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid biases size + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid biases dimensions + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::F32, DataLayout::NDHWC), // Invalid output size + TensorInfo(TensorShape(27U, 13U, 2U, 4U), 1U, DataType::U32, DataLayout::NDHWC), // Invalid data type + }), + framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F16), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 3U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U, 3U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 3U, 3U, 3U, 2U), 1U, DataType::U32), + })), + framework::dataset::make("BiasesInfo",{ TensorInfo(TensorShape(4U), 1U, DataType::F32), + TensorInfo(TensorShape(4U), 1U, DataType::F32), + TensorInfo(TensorShape(4U), 1U, DataType::F32), + TensorInfo(TensorShape(4U), 1U, DataType::F32), + TensorInfo(TensorShape(3U), 1U, DataType::F32), + TensorInfo(TensorShape(4U, 2U), 1U, DataType::F32), + TensorInfo(TensorShape(4U), 1U, DataType::F32), + TensorInfo(TensorShape(4U), 1U, DataType::F32), + })), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(26U, 11U, 4U), 1U, DataType::F32), + TensorInfo(TensorShape(25U, 11U, 4U), 1U, DataType::U32), + })), + framework::dataset::make("Expected", { false, false, false, false, false, false, false, false})), + input_info, weights_info, biases_info, output_info, expected) +{ + const Conv3dInfo conv3d_info(Size3D(1, 1, 1), Padding3D(0, 0, 0), ActivationLayerInfo(), Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false); + bool is_valid = bool(NEConv3D::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv3d_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using NEDirectConvolution3DFixture = DirectConvolution3DValidationFixture<Tensor, Accessor, NEConv3D, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolution3DFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(data_precommit, + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NDHWC }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() // FP32 + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolution3DFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(data_precommit, + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", { DataLayout::NDHWC }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_f16); +} +TEST_SUITE_END() // FP16 +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + +TEST_SUITE_END() // Float + +template <typename T> +using NEDirectConvolution3DQuantizedFixture = DirectConvolution3DValidationQuantizedFixture<Tensor, Accessor, NEConv3D, T>; + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolution3DQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(7U, 5U, 3U, 13U, 3U), + TensorShape(15U, 7U, 11U, 7U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U) + }), + framework::dataset::make("StrideX", { 1, 3, 2, 1 })), + framework::dataset::make("StrideY", { 2, 1, 3, 1 })), + framework::dataset::make("StrideZ", { 3, 2, 1, 1 })), + framework::dataset::make("PadX", { 0, 2, 1, 0 })), + framework::dataset::make("PadY", { 1, 0, 2, 0 })), + framework::dataset::make("PadZ", { 2, 1, 0, 0 })), + framework::dataset::make("KernelWidth", { 3, 7, 5, 1 })), + framework::dataset::make("KernelHeight", { 5, 3, 7, 1 })), + framework::dataset::make("KernelDepth", { 7, 5, 3, 1 })), + framework::dataset::make("NumKernels", { 5, 3, 1, 11 })), + framework::dataset::make("HasBias", { true, true, true, false })), + framework::dataset::make("Activation", ActivationLayerInfo())), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("DataLayout", DataLayout::NDHWC)), + framework::dataset::make("SrcQuantizationInfo", QuantizationInfo(0.1f, 10))), + framework::dataset::make("WeightsQuantizationInfo", QuantizationInfo(0.3f, 20))), + framework::dataset::make("DstQuantizationInfo", QuantizationInfo(0.2f, 5)))) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolution3DQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(combine(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(7U, 5U, 3U, 13U, 3U), + TensorShape(15U, 7U, 11U, 7U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U) + }), + framework::dataset::make("StrideX", { 1, 3, 2, 1 })), + framework::dataset::make("StrideY", { 2, 1, 3, 1 })), + framework::dataset::make("StrideZ", { 3, 2, 1, 1 })), + framework::dataset::make("PadX", { 0, 2, 1, 0 })), + framework::dataset::make("PadY", { 1, 0, 2, 0 })), + framework::dataset::make("PadZ", { 2, 1, 0, 0 })), + framework::dataset::make("KernelWidth", { 3, 7, 5, 1 })), + framework::dataset::make("KernelHeight", { 5, 3, 7, 1 })), + framework::dataset::make("KernelDepth", { 7, 5, 3, 1 })), + framework::dataset::make("NumKernels", { 5, 3, 1, 11 })), + framework::dataset::make("HasBias", { true, true, true, false })), + framework::dataset::make("Activation", ActivationLayerInfo())), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataLayout", DataLayout::NDHWC)), + framework::dataset::make("SrcQuantizationInfo", QuantizationInfo(0.1f, 10))), + framework::dataset::make("WeightsQuantizationInfo", QuantizationInfo(0.3f, 20))), + framework::dataset::make("DstQuantizationInfo", QuantizationInfo(0.2f, 5)))) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + +TEST_SUITE_END() // Convolution3D +TEST_SUITE_END() // Neon +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index 9e00da16ae..d739d4e1a4 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -28,11 +28,16 @@ #include "arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" + +#include "src/core/CPP/Validate.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/cpu/operators/CpuGemmConv2d.h" +#include "src/cpu/operators/CpuGemmDirectConv2d.h" +#include "src/cpu/operators/CpuWinogradConv2d.h" + #include "tests/NEON/Accessor.h" -#include "tests/PaddingCalculator.h" #include "tests/datasets/LargeConvolutionLayerDataset.h" #include "tests/datasets/SmallConvolutionLayerDataset.h" -#include "tests/datasets/TinyConvolutionLayerDataset.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -46,6 +51,8 @@ namespace test { namespace validation { +using framework::dataset::make; + namespace detail { template <> @@ -77,10 +84,17 @@ const RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2 const AbsoluteTolerance<float> abs_tolerance_f16(0.2f); /**< Absolute tolerance for FP16 types */ constexpr float tolerance_num = 0.07f; /**< Tolerance number for the FP16 implementation */ #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -constexpr AbsoluteTolerance<float> tolerance_qasymm8(0.0); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ + +#ifdef ARM_COMPUTE_ENABLE_SME +// TODO(COMPMID-6011): SME kernels and the reference model use different rounding mode. +// Temporarily increase the tolerance for quantized data. +constexpr AbsoluteTolerance<float> tolerance_qasymm8(1.0); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +#else // ARM_COMPUTE_ENABLE_SME +constexpr AbsoluteTolerance<float> tolerance_qasymm8(0.0); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +#endif // ARM_COMPUTE_ENABLE_SME /** CNN data types */ -const auto CNNDataTypes = framework::dataset::make("DataType", +const auto CNNDataTypes = make("DataType", { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC DataType::F16, @@ -88,14 +102,41 @@ const auto CNNDataTypes = framework::dataset::make("DataType", DataType::F32, DataType::QASYMM8, }); -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +const auto ActivationFunctionsDataset = make("ActivationInfo", { ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f) }); -const auto QuantizationData = framework::dataset::make("QuantizationInfo", +const auto NoActivation = make("ActivationInfo", +{ + ActivationLayerInfo(), +}); + +const auto ActivationFunctionsDatasetNightly = make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), + + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f, -0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SOFT_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ABS), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SQUARE), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::HARD_SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 2.f, 1.f), +#ifdef __aarch64__ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::GELU), +#endif // __aarch64__ +}); + +const auto QuantizationData = make("QuantizationInfo", { QuantizationInfo(0.5f, 10), QuantizationInfo(0.3f, 3), @@ -110,32 +151,32 @@ TEST_SUITE(ConvolutionLayer) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(18U, 18U, 32U), 1, DataType::F32), + make("InputInfo", { TensorInfo(TensorShape(18U, 18U, 32U), 1, DataType::F32), TensorInfo(TensorShape(23U, 27U, 32U, 4U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 1U), 1, DataType::F32), TensorInfo(TensorShape(33U, 27U, 7U, 4U), 1, DataType::F32) }), - framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 32U, 21U), 1, DataType::F32), + make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 32U, 21U), 1, DataType::F32), TensorInfo(TensorShape(5U, 5U, 32U, 21U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32), TensorInfo(TensorShape(5U, 5U, 7U, 16U), 1, DataType::F16) })), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(16U, 16U, 21U), 1, DataType::F32), + make("OutputInfo", { TensorInfo(TensorShape(16U, 16U, 21U), 1, DataType::F32), TensorInfo(TensorShape(19U, 23U, 21U, 4U), 1, DataType::F32), TensorInfo(TensorShape(11U, 25U, 21U), 1, DataType::F32), TensorInfo(TensorShape(11U, 12U, 16U, 4U), 1, DataType::F32) })), - framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(3, 2, 1, 0) })), - framework::dataset::make("FastMath", { true, + make("FastMath", { true, true, false, false })), - framework::dataset::make("Expected", { ConvolutionMethod::WINOGRAD, ConvolutionMethod::WINOGRAD, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM })), + make("Expected", { ConvolutionMethod::WINOGRAD, ConvolutionMethod::WINOGRAD, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM })), input_info, weights_info, output_info, conv_info, fast_math, expected) { ConvolutionMethod is_valid = NEConvolutionLayer::get_convolution_method(&input_info.clone()->set_is_resizable(true), @@ -147,6 +188,14 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z // *INDENT-ON* TEST_SUITE_END() // ConvolutionLayer +/* + Testing Strategy of Neon Winograd: + - There is no need to thoroughly test nchw cases because winograd kernels accept + nhwc and the tensors are permuted before and after if they're nchw. + - Except relu and bounded relu, testing activations for a single input + combination is enough because activation is not fused into winograd and called + separately. +*/ TEST_SUITE(WinogradLayer) template <typename T> using NEWinogradConvolutionLayerFixture = WinogradConvolutionLayerFastMathValidationFixture<Tensor, Accessor, NEWinogradConvolutionLayer, T>; @@ -156,38 +205,250 @@ using NEWinogradConvolutionLayerMixedDataLayoutFixture = WinogradConvolutionLaye template <typename T> using NEWinogradConvolutionLayerNoBiasFixture = WinogradConvolutionLayerFastMathValidationFixture<Tensor, Accessor, NEWinogradConvolutionLayer, T, T, false>; +/** Test case for memory injection in @ref cpu::CpuWinogradConv2d. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) +{ + auto winograd = std::make_unique<cpu::CpuWinogradConv2d>(); + const auto src_info = TensorInfo(TensorShape(8U, 8U, 32U), 1, DataType::F32); + const auto w_info = TensorInfo(TensorShape(1U), 1, DataType::F32); + const auto b_info = TensorInfo(TensorShape(1U, 3U, 32U, 1U), 1, DataType::F32); + auto dst_info = TensorInfo(TensorShape(8U, 6U, 1U), 1, DataType::F32); + const PadStrideInfo pad_info{}; + + winograd->configure(&src_info, &b_info, &w_info, &dst_info, pad_info); + + // telhs are newly created every call of this lambda function + auto a = create_tensor<Tensor>(src_info); + auto b = create_tensor<Tensor>(b_info); + auto c = create_tensor<Tensor>(w_info); + a.allocator()->allocate(); + b.allocator()->allocate(); + c.allocator()->allocate(); + + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &a }, { TensorType::ACL_SRC_1, &b }, { TensorType::ACL_SRC_2, &c } }; + ITensorPack prep_pack{ { TensorType::ACL_SRC_1, &b }, { TensorType::ACL_SRC_2, &c } }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(winograd->workspace(), mg, run_pack, prep_pack); + auto run_conv = [&]() -> Tensor + { + auto dst = create_tensor<Tensor>(dst_info); + dst.allocator()->allocate(); + + run_pack.add_tensor(TensorType::ACL_DST, &dst); + library->fill_tensor_value(Accessor(a), 1.f); + library->fill_tensor_value(Accessor(b), 2.f); + library->fill_tensor_value(Accessor(c), 3.f); + + // This operator is configured once and captured by this lambda. + winograd->prepare(prep_pack); + winograd->run(run_pack); + return dst; + }; + + auto result_0 = run_conv(); + auto result_1 = run_conv(); + + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +/** Test case for memory injection in @ref NEWinogradConvolutionLayer. + * + * Make sure @ref NEWinogradConvolutionLayer still works through injecting the memory at configure time using the old API. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) +{ + auto gemm = std::make_unique<NEWinogradConvolutionLayer>(); + const auto src_info = TensorInfo(TensorShape(8U, 8U, 32U), 1, DataType::F32); + const auto w_info = TensorInfo(TensorShape(1U), 1, DataType::F32); + const auto b_info = TensorInfo(TensorShape(1U, 3U, 32U, 1U), 1, DataType::F32); + auto dst_info = TensorInfo(TensorShape(8U, 6U, 1U), 1, DataType::F32); + const PadStrideInfo pad_info{}; + + auto run_conv = [&]() + { + auto src = create_tensor<Tensor>(src_info); + auto w = create_tensor<Tensor>(w_info); + auto b = create_tensor<Tensor>(b_info); + auto dst = create_tensor<Tensor>(dst_info); + + gemm->configure(&src, &b, &w, &dst, pad_info); + + src.allocator()->allocate(); + b.allocator()->allocate(); + w.allocator()->allocate(); + dst.allocator()->allocate(); + + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(b), 2.f); + library->fill_tensor_value(Accessor(w), 3.f); + gemm->run(); + return dst; + }; + + auto result_0 = run_conv(); + auto result_1 = run_conv(); + + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +DATA_TEST_CASE(SupportedKernels, framework::DatasetMode::ALL, zip( + make("WeightsInfo", +{ + // Shapes are always in NCHW format. When layout is NHWC, the shape is permuted + + // Fp32, NCHW/NHWC (layout does not matter as it's ) + // 3x1, 1x3, 3x3 --> all TRUE + TensorInfo(TensorShape(3U, 3U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(1U, 3U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + + // 5x1, 1x5, 5x5 --> all TRUE + TensorInfo(TensorShape(5U, 5U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 5U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(5U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + + // 7x1, 1x7, 7x7 + // --> all FALSE + TensorInfo(TensorShape(7U, 7U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 7U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(7U, 1U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + + // unsupported kernel sizes + TensorInfo(TensorShape(2U, 2U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(5U, 2U, 2U, 8U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 6U, 2U, 8U), 1, DataType::F32, DataLayout::NCHW), + + // Fp16 + TensorInfo(TensorShape(3U, 3U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(1U, 3U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 1U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + + // 5x1, 1x5, 5x5 --> all TRUE + TensorInfo(TensorShape(5U, 5U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 5U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(5U, 1U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + + // 7x1, 1x7, 7x7 + // --> all FALSE + TensorInfo(TensorShape(7U, 7U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + TensorInfo(TensorShape(1U, 7U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(7U, 1U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + + // unsupported kernel sizes + TensorInfo(TensorShape(2U, 2U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(5U, 2U, 2U, 8U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 6U, 2U, 8U), 1, DataType::F16, DataLayout::NCHW), + +}), +make("Expected", +{ + // fp32 + true, true, true, // 3x3, 1x3, 3x1 + true, true, true, // 5x5, 1x5, 5x1 + false, true, true, // 7x7, 1x7, 7x1 + false, false, false, // random unsupported kernels + + // fp16 + true, false, false, // 3x3, 1x3, 3x1 + false, false, false, // 5x5, 1x5, 5x1 + false, false, false, // 7x7, 1x7, 7x1 + false, false, false, // random unsupported kernels +})), +weights_info_const, expected_const) +{ + DataType data_type = weights_info_const.data_type(); + DataLayout data_layout = weights_info_const.data_layout(); + + TensorInfo input_info = TensorInfo(TensorShape(17U, 31U, 2U), 1, data_type); + TensorInfo bias_info = TensorInfo(TensorShape(8U), 1, data_type); + TensorInfo weights_info = weights_info_const; + + if(data_layout == DataLayout::NHWC) + { + // Convert to NHWC + PermutationVector perm = PermutationVector(2U, 0U, 1U); + + TensorShape input_shape = input_info.tensor_shape(); + TensorShape weights_shape = weights_info.tensor_shape(); + permute(input_shape, perm); + permute(weights_shape, perm); + + input_info.set_tensor_shape(input_shape); + weights_info.set_tensor_shape(weights_shape); + + input_info.set_data_layout(data_layout); + weights_info.set_data_layout(data_layout); + bias_info.set_data_layout(data_layout); + } + + PadStrideInfo conv_info(1, 1, 0, 0); + + TensorShape output_shape = compute_deep_convolution_shape(input_info, weights_info, conv_info); + TensorInfo output_info = TensorInfo(output_shape, 1, data_type, data_layout); + + Status status = NEWinogradConvolutionLayer::validate( + &input_info, + &weights_info, + &bias_info, + &output_info, + conv_info, + ActivationLayerInfo(), + true /* fast math */); + + Status fp16_supported = ::arm_compute::error_on_unsupported_cpu_fp16("N/A", "N/A", 0, &input_info); + bool expected = expected_const && static_cast<bool>(fp16_supported); + + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); +} + TEST_SUITE(FP32) TEST_SUITE(Conv1x3) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x3Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEWinogradConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(8U, 8U, 32U)), - framework::dataset::make("Weight", TensorShape(1U, 3U, 32U, 1U))), - framework::dataset::make("Bias", TensorShape(1U))), - framework::dataset::make("Output", TensorShape(8U, 6U, 1U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1U, 1U))), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine( + make("Input", TensorShape(8U, 8U, 32U)), + make("Weight", TensorShape(1U, 3U, 32U, 1U)), + make("Bias", TensorShape(1U)), + make("Output", TensorShape(8U, 6U, 1U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x3Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_1xN_f32); @@ -197,19 +458,19 @@ TEST_SUITE_END() // Conv1x3 TEST_SUITE(Conv3x1) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x1Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer3x1Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_1xN_f32); @@ -219,19 +480,19 @@ TEST_SUITE_END() // Conv3x1 TEST_SUITE(Conv1x5) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x5Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer1x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer1x5Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_1xN_f32); @@ -241,19 +502,19 @@ TEST_SUITE_END() // Conv1x5 TEST_SUITE(Conv5x1) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer5x1Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer5x1Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_1xN_f32); @@ -263,10 +524,10 @@ TEST_SUITE_END() // Conv5x1 TEST_SUITE(Conv7x1) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer7x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer7x1Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); @@ -274,9 +535,9 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, frame FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeWinogradConvolutionLayer7x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + make("DataType", { DataType::F32 })), + make("ActivationInfo", { ActivationLayerInfo() })), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_1xN_f32); @@ -285,20 +546,20 @@ TEST_SUITE_END() // Conv7x1 TEST_SUITE(Conv1x7) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer1x7Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer1x7Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer7x1Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer7x1Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_1xN_f32); @@ -307,20 +568,40 @@ TEST_SUITE_END() // Conv1x7 TEST_SUITE(Conv3x3) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + +{ + // Validate output + validate(Accessor(_target), _reference, abs_tolerance_f32); +} +/// It's enough to run the activations for a single weight/input combination and data type because +/// activation function is called on top of the winograd output as a separate operator +/// TODO: Enable after COMPMID-6573 is resolved +FIXTURE_DATA_TEST_CASE(RunActivations, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::DISABLED, + combine( + make("Input", TensorShape(3U, 3U, 32U)), + make("Weight", TensorShape(3U, 3U, 32U, 4U)), + make("Bias", TensorShape(4U)), + make("Output", TensorShape(1U, 1U, 4U)), + make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1U, 1U)), + make("DataType", { DataType::F32 }), + ActivationFunctionsDatasetNightly, + make("DataLayout", { DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } + FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output @@ -331,20 +612,20 @@ TEST_SUITE_END() // Conv3x3 TEST_SUITE(Conv5x5) FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer5x5Dataset(), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer5x5Dataset(), + make("DataType", { DataType::F32 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output @@ -354,12 +635,12 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEWinogradConvolutionLayerFixture<float>, frame TEST_SUITE_END() // Conv5x5 FIXTURE_DATA_TEST_CASE(RunSmallNoBias, NEWinogradConvolutionLayerNoBiasFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(framework::dataset::concat(datasets::SmallWinogradConvolutionLayer3x3Dataset(), - datasets::SmallWinogradConvolutionLayer5x5Dataset()), - framework::dataset::make("DataType", { DataType::F32 })), - ActivationFunctionsDataset), - - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(framework::dataset::concat( + datasets::SmallWinogradConvolutionLayer3x3Dataset(), + datasets::SmallWinogradConvolutionLayer5x5Dataset()), + make("DataType", { DataType::F32 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, abs_tolerance_f32); @@ -371,12 +652,39 @@ TEST_SUITE_END() // FP32 TEST_SUITE(FP16) using CLWinogradConvolutionLayerFastMathFixture16 = WinogradConvolutionLayerFastMathValidationFixture<Tensor, Accessor, NEWinogradConvolutionLayer, half, float>; +DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip( + make("InputInfo", { TensorInfo(TensorShape(18U, 18U, 32U), 1, DataType::F16), + TensorInfo(TensorShape(18U, 18U, 32U), 1, DataType::F16) + }), + make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 32U, 21U), 1, DataType::F16), + TensorInfo(TensorShape(3U, 3U, 32U, 21U), 1, DataType::F16) + }), + make("OutputInfo", { TensorInfo(TensorShape(16U, 16U, 21U), 1, DataType::F32), + TensorInfo(TensorShape(16U, 16U, 21U), 1, DataType::F16) + }), + make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0) + }), + make("FastMath", +{ + false, // case fp16 and fast_math False then disable Winograd + true // case fp16 and fast_math True then enable Winograd +}), +make("Expected", { ConvolutionMethod::GEMM, ConvolutionMethod::WINOGRAD })), +input_info, weights_info, output_info, conv_info, fast_math, expected) +{ + ConvolutionMethod is_valid = NEConvolutionLayer::get_convolution_method(&input_info.clone()->set_is_resizable(true), + &weights_info.clone()->set_is_resizable(true), + &output_info.clone()->set_is_resizable(true), conv_info, WeightsInfo(), Size2D(1U, 1U), ActivationLayerInfo(), fast_math); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} + TEST_SUITE(Conv3x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::SmallWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F16 }), + ActivationFunctionsDataset, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output @@ -384,10 +692,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradConvolutionLayerFastMathFixture16, fr } FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradConvolutionLayerFastMathFixture16, framework::DatasetMode::NIGHTLY, - combine(combine(combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), - framework::dataset::make("DataType", { DataType::F16 })), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + combine(datasets::LargeWinogradConvolutionLayer3x3Dataset(), + make("DataType", { DataType::F16 }), + make("ActivationInfo", { ActivationLayerInfo() }), + make("DataLayout", { DataLayout::NHWC }))) { // Validate output @@ -398,18 +706,470 @@ TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE_END() // WinogradLayer +#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +TEST_SUITE(FIXED_FORMAT_KERNELS) +TEST_SUITE(VariableWeightUtils) + +// UC2_1_* tests: the user requests a specific fixed format, but there is no kernel that supports it. + +template <typename ConvolutionClass> +using HasOptImplFixtureNoFastMath = HasOptImplFixture<ConvolutionClass, /*enable_fast_math*/ false>; + +template <typename ConvolutionClass> +using HasOptImplFixtureFastMath = HasOptImplFixture<ConvolutionClass, /*enable_fast_math*/ true>; + +// UC2_1 + +FIXTURE_DATA_TEST_CASE(UC2_1_CpuGemmConv2d, HasOptImplFixtureNoFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo2 }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} +FIXTURE_DATA_TEST_CASE(UC2_1_NEGEMMConvolutionLayer, HasOptImplFixtureNoFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo2 }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC2_1_CpuGemmConv2d_FastMath, HasOptImplFixtureFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo2 }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC2_1_NEGEMMConvolutionLayer_FastMath, HasOptImplFixtureFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo2 }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +// UC2_2_* tests: the user requests a specific fixed format, and a +// kernel that support that fixed format is found. + +FIXTURE_DATA_TEST_CASE(UC2_2_CpuGemmConv2d, HasOptImplFixtureNoFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo4 }))) +{ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format == arm_compute::WeightFormat::OHWIo4, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC2_2_NEGEMMConvolutionLayer, HasOptImplFixtureNoFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo4 }))) +{ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format == arm_compute::WeightFormat::OHWIo4, framework::LogLevel::ERRORS); +} + +#if defined(ARM_COMPUTE_ENABLE_BF16) +// These tests currently only works with SVE length 256 +// If other SVE length is used a kernel will fail to be found +// This needs to be addressed in order to ensure it doesn't revert to FP32 kernels for systems with SVE length other than 256 +FIXTURE_DATA_TEST_CASE(UC2_2_CpuGemmConv2d_FastMath, HasOptImplFixtureFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo8i4_bf16 }))) +{ + if(Scheduler::get().cpu_info().has_bf16() && (arm_gemm::utils::get_vector_length<float>() == 8)){ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT_EQUAL(_computed_weight_format, arm_compute::WeightFormat::OHWIo8i4_bf16, framework::LogLevel::ERRORS); + } + else{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); + } +} + +FIXTURE_DATA_TEST_CASE(UC2_2_NEGEMMConvolutionLayer_FastMath, HasOptImplFixtureFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::OHWIo8i4_bf16 }))) +{ + if(Scheduler::get().cpu_info().has_bf16() && (arm_gemm::utils::get_vector_length<float>() == 8)){ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format == arm_compute::WeightFormat::OHWIo8i4_bf16, framework::LogLevel::ERRORS); + } + else{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); + } +} + +#endif // ARM_COMPUTE_ENABLE_BF16 + +// UC3_1_* tests: the user queries for ANY fixed format, but there is +// no kernel that support the use case specified by the user (for +// example, there is no fixed format kernel for the datatype of the +// problem). + +FIXTURE_DATA_TEST_CASE(UC3_1_CpuGemmConv2d, HasOptImplFixtureNoFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::S32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC3_1_NEGEMMConvolutionLayer, HasOptImplFixtureNoFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::S32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC3_1_CpuGemmConv2d_FastMath, HasOptImplFixtureFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::S32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC3_1_NEGEMMConvolutionLayer_FastMath, HasOptImplFixtureFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::S32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + ARM_COMPUTE_EXPECT(!_kernel_found, framework::LogLevel::ERRORS); +} + +// UC3_2_* tests: the user queries for ANY fixed format. The search +// succeeded and the fixed format found is prompted back for +// consumption by the user. Note that we just test the +// _computed_weight_format to be anything but not the formats that are +// not fixed formats (ANY and UNSPECIFIED). This is because the weight +// format that the runtime produces depends on the size of the vector +// units of the hardware where the tests is executed. For example, a +// format like OHWIo4 for FP32 data returned for 128-bit NEON hardware +// is replaced by OHWIo8 when running on 256-bit SVE. + +FIXTURE_DATA_TEST_CASE(UC3_2_CpuGemmConv2d, HasOptImplFixtureNoFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::ANY, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::UNSPECIFIED, framework::LogLevel::ERRORS); +} + +FIXTURE_DATA_TEST_CASE(UC3_2_NEGEMMConvolutionLayer, HasOptImplFixtureNoFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::ANY, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::UNSPECIFIED, framework::LogLevel::ERRORS); +} + +#if defined(ARM_COMPUTE_ENABLE_BF16) + +FIXTURE_DATA_TEST_CASE(UC3_2_CpuGemmConv2d_FastMath, HasOptImplFixtureFastMath<cpu::CpuGemmConv2d>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + if(Scheduler::get().cpu_info().has_bf16()){ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::ANY, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::UNSPECIFIED, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(arm_compute::is_fixed_format_fast_math(_computed_weight_format), framework::LogLevel::ERRORS); + } + else{ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::ANY, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::UNSPECIFIED, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!arm_compute::is_fixed_format_fast_math(_computed_weight_format), framework::LogLevel::ERRORS); + } +} + +FIXTURE_DATA_TEST_CASE(UC3_2_NEGEMMConvolutionLayer_FastMath, HasOptImplFixtureFastMath<NEGEMMConvolutionLayer>, framework::DatasetMode::ALL, + combine(framework::dataset::make("DataType", { DataType::F32 }), + framework::dataset::make("QueryWeightFormat", { arm_compute::WeightFormat::ANY }))) +{ + if(Scheduler::get().cpu_info().has_bf16()){ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::ANY, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::UNSPECIFIED, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(arm_compute::is_fixed_format_fast_math(_computed_weight_format), framework::LogLevel::ERRORS); + } + else{ + ARM_COMPUTE_EXPECT(_kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::ANY, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(_computed_weight_format != arm_compute::WeightFormat::UNSPECIFIED, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!arm_compute::is_fixed_format_fast_math(_computed_weight_format), framework::LogLevel::ERRORS); + } +} + +#endif // ARM_COMPUTE_ENABLE_BF16 + +namespace +{ +using TestCaseType = std::tuple<TensorShape, TensorShape, arm_compute::WeightFormat>; +auto prepare_weights_shapes = framework::dataset::make("TensorShape", +{ + // OHWIo<interleave_by>i<block_by> + // + // OHWI --> O'HWI', where: + // + // O'= smallest multiple of <interleave_by> such that O<=O' + // I'= smallest multiple of <block_by> such that I<=I' + // + + // Change N for OHWIo4 + TestCaseType({ { 1U, 1U, 1U, 1U }, { 1U, 1U, 1U, 4U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 2U }, { 1U, 1U, 1U, 4U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 3U }, { 1U, 1U, 1U, 4U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 4U }, { 1U, 1U, 1U, 4U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 5U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 6U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 7U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 8U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1U, 1U, 1U, 9U }, { 1U, 1U, 1U, 12U }, arm_compute::WeightFormat::OHWIo4 }), + // // Change N for OHWIo8 + TestCaseType({ { 1U, 1U, 1U, 1U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 2U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 3U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 4U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 5U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 6U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 7U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 8U }, { 1U, 1U, 1U, 8U }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1U, 1U, 1U, 9U }, { 1U, 1U, 1U, 16U }, arm_compute::WeightFormat::OHWIo8 }), + // // Change N for OHWIo4 when H, W and C are not 1 + TestCaseType({ { 3U, 4U, 2U, 1U }, { 3, 4, 2, 4 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 2U }, { 3, 4, 2, 4 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 3U }, { 3, 4, 2, 4 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 4U }, { 3, 4, 2, 4 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 5U }, { 3, 4, 2, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 6U }, { 3, 4, 2, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 7U }, { 3, 4, 2, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 8U }, { 3, 4, 2, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 9U }, { 3, 4, 2, 12 }, arm_compute::WeightFormat::OHWIo4 }), + + // // Fix N and move HWI around, with different data layouts and formats + TestCaseType({ { 2U, 4U, 3U, 5U }, { 2, 4, 3, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 3U, 4U, 2U, 5U }, { 3, 4, 2, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 2U, 4U, 3U, 9U }, { 2, 4, 3, 16 }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 3U, 4U, 2U, 9U }, { 3, 4, 2, 16 }, arm_compute::WeightFormat::OHWIo8 }), + TestCaseType({ { 1024U, 1U, 1U, 1001U }, { 1024, 1, 1, 1008 }, arm_compute::WeightFormat::OHWIo8 }), + + // // Adding <block_by> on I (=C) + TestCaseType({ { 1U, 4U, 3U, 5U }, { 2, 4, 3, 8 }, arm_compute::WeightFormat::OHWIo4i2 }), + TestCaseType({ { 2U, 4U, 3U, 5U }, { 2, 4, 3, 8 }, arm_compute::WeightFormat::OHWIo4i2 }), + TestCaseType({ { 3U, 4U, 3U, 5U }, { 4, 4, 3, 8 }, arm_compute::WeightFormat::OHWIo4i2 }), + + // --------- + TestCaseType({ { 2, 2, 1, 5 }, { 2, 2, 1, 8 }, arm_compute::WeightFormat::OHWIo4 }), + TestCaseType({ { 1, 2, 2, 5 }, { 1, 2, 2, 8 }, arm_compute::WeightFormat::OHWIo4 }), + +}); +} // unnamed namespace + +DATA_TEST_CASE(PrepareWeightShape, framework::DatasetMode::ALL, + prepare_weights_shapes, shapes) +{ + const TensorShape input_shape = std::get<0>(shapes); + const TensorShape expected_shape = std::get<1>(shapes); + const arm_compute::WeightFormat wf = std::get<2>(shapes); + const DataType DT = DataType::F32; + const DataLayout DL = DataLayout::NHWC; + const auto TI = TensorInfo(input_shape, 1 /*num_channels, deprecated*/, DT, DL); + const TensorInfo computed_info = ::arm_compute::test::validation::prepare_weights(TI, wf); + ARM_COMPUTE_EXPECT_EQUAL(computed_info.tensor_shape(), expected_shape, framework::LogLevel::ERRORS); +} + +TEST_SUITE_END() // VariableWeightUtils + +TEST_SUITE(ExperimentalCpuAPIVariableWeightWithFixtures) + +template <typename ScalarType> +using VarWidth = VariableWeightsFixture<cpu::CpuGemmConv2d, Tensor, Accessor, ScalarType, /*enable_fast_math*/ false>; + +FIXTURE_DATA_TEST_CASE(RunSmallFloat, VarWidth<float>, framework::DatasetMode::ALL, + combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ACL Scalar type", { DataType::F32 }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); +} + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) +FIXTURE_DATA_TEST_CASE(RunSmallHalf, VarWidth<half>, framework::DatasetMode::ALL, + combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ACL Scalar type", { DataType::F16 }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f16, 0.f, half(abs_tolerance_f16)); +} +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#if defined(ARM_COMPUTE_ENABLE_BF16) +template <typename ScalarType> +using VarWidthFastMath = VariableWeightsFixture<cpu::CpuGemmConv2d, Tensor, Accessor, ScalarType, /*enable_fast_math*/ true>; + +FIXTURE_DATA_TEST_CASE(RunSmallFloatFastMath, VarWidthFastMath<float>, framework::DatasetMode::ALL, + combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ACL Scalar type", { DataType::F32 }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); +} +#endif // ARM_COMPUTE_ENABLE_BF16 + +TEST_SUITE_END() // ExperimentalCpuAPIVariableWeightWithFixtures + +TEST_SUITE(ExperimentalNEAPIVariableWeightWithFixtures) + +template <typename ScalarType> +using NEGEMMVarWidth = VariableWeightsFixtureNEInterface<NEGEMMConvolutionLayer, Tensor, Accessor, ScalarType, /*enable_fast_math*/ false>; + +FIXTURE_DATA_TEST_CASE(NEGEMMRunSmallFloat, NEGEMMVarWidth<float>, framework::DatasetMode::ALL, + combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ACL Scalar type", { DataType::F32 }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); +} + +#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) +FIXTURE_DATA_TEST_CASE(NEGEMMRunSmallHalf, NEGEMMVarWidth<half>, framework::DatasetMode::ALL, + combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ACL Scalar type", { DataType::F16 }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f16, 0.f, half(abs_tolerance_f16)); +} +#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#if defined(ARM_COMPUTE_ENABLE_BF16) +template <typename ScalarType> +using NEGEMMVarWidthFastMath = VariableWeightsFixtureNEInterface<NEGEMMConvolutionLayer, Tensor, Accessor, ScalarType, /*enable_fast_math*/ true>; + +FIXTURE_DATA_TEST_CASE(NEGEMMRunSmallFloatFastMath, NEGEMMVarWidthFastMath<float>, framework::DatasetMode::ALL, + combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ACL Scalar type", { DataType::F32 }))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); +} +#endif // ARM_COMPUTE_ENABLE_BF16 + +TEST_SUITE_END() // ExperimentalNEAPIVariableWeightWithFixtures +TEST_SUITE_END() // FIXED_FORMAT_KERNELS + +#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS + TEST_SUITE(GEMMConvolutionLayer) template <typename T> using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture<Tensor, Accessor, NEConvolutionLayer, T>; template <typename T> +using NEGEMMConvolutionLayerPaddedWeightsFixture = ConvolutionValidationPaddedWeightsFixture<Tensor, Accessor, NEConvolutionLayer, T>; +template <typename T> using NEGEMMConvolutionLayerMixedDataLayoutFixture = ConvolutionValidationFixture<Tensor, Accessor, NEConvolutionLayer, T, true>; +/** Test case for memory injection in @ref cpu::CpuGemmConv2d. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) +{ + auto conv = std::make_unique<cpu::CpuGemmConv2d>(); + const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NCHW); + const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NCHW); + const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NCHW); + auto dst_info = TensorInfo(TensorShape(1U, 7U, 3U), 1, DataType::F32, DataLayout::NCHW); + const auto conv_info = PadStrideInfo(1, 1, 0, 0, 2, 2, DimensionRoundingType::FLOOR); + WeightsInfo weights_info(false, 3U, 3U, 1U); + conv->configure(&src_info, &weight_info, &bias_info, &dst_info, conv_info, weights_info); + + // tensors are newly created every call of this lambda function + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &src }, { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + ITensorPack prep_pack{ { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(conv->workspace(), mg, run_pack, prep_pack); + + auto run_conv = [&]() -> Tensor + { + auto dst = create_tensor<Tensor>(dst_info); + dst.allocator()->allocate(); + run_pack.add_tensor(TensorType::ACL_DST, &dst); + + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(weight), 2.f); + library->fill_tensor_value(Accessor(bias), 3.f); + // This operator is configured once and captured by this lambda. + conv->prepare(prep_pack); + conv->run(run_pack); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +/** Test case for memory injection in @ref NEGEMMConvolutionLayer. + * + * Make sure @ref NEGEMMConvolutionLayer still works through injecting the memory at configure time using the old API. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) +{ + auto conv = std::make_unique<NEGEMMConvolutionLayer>(); + const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NCHW); + const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NCHW); + const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NCHW); + auto dst_info = TensorInfo(TensorShape(1U, 7U, 3U), 1, DataType::F32, DataLayout::NCHW); + const auto conv_info = PadStrideInfo(1, 1, 0, 0, 2, 2, DimensionRoundingType::FLOOR); + WeightsInfo weights_info(false, 3U, 3U, 1U); + auto run_conv = [&]() + { + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + auto dst = create_tensor<Tensor>(dst_info); + conv->configure(&src, &weight, &bias, &dst, conv_info, weights_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(weight), 2.f); + library->fill_tensor_value(Accessor(bias), 3.f); + conv->run(); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + TEST_SUITE(Float) -#if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) +#if defined(ARM_COMPUTE_ENABLE_BF16) TEST_SUITE(BFLOAT16) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::BFLOAT16)), + framework::dataset::make("DataType", Scheduler::get().cpu_info().has_bf16() ? DataType::BFLOAT16 : DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { @@ -417,7 +1177,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<float>, framework validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); } TEST_SUITE_END() // BFLOAT16 -#endif /* defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) */ +#endif /* defined(ARM_COMPUTE_ENABLE_BF16) */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -459,9 +1219,39 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEGEMMConvolutionLayerMixedDataLayout // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); } +/** Padded weights + * CpuGemmConv2d uses two different paths for reshaping the weights based on if the weight tensor has holes (a common + * way to have "holes" in tensor is via extended paddings) + * + * We only need to test the padded weight path here on a single floating data type and a single layout, because the fallback path is agnostic of them + */ +FIXTURE_DATA_TEST_CASE(RunPaddedWeights, NEGEMMConvolutionLayerPaddedWeightsFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true }), + framework::dataset::make("DataType", DataType::F32), + framework::dataset::make("DataLayout", { DataLayout::NHWC }) + )) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); +} + +// This very large shape test is required to test heuristic paths where the tensor size is > 1e7 bytes +// and weight dimensions larger than 7 +FIXTURE_DATA_TEST_CASE(RunVeryLarge, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, + combine(datasets::VeryLargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true }), + framework::dataset::make("DataType", DataType::F32), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); +} + TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float +// TODO: COMPMID-6596 Extend quantized tests with at least one suite where the weight is padded (the legacy case, see floating point's RunPaddedWeights) template <typename T> using NEGEMMConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<Tensor, Accessor, NEConvolutionLayer, T>; template <typename T> @@ -477,12 +1267,17 @@ const auto QuantizedActivationFunctionsDataset = framework::dataset::make("Activ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) }); TEST_SUITE(Quantized) +/// @note: Every asymmetric quantized test where there's no fused activation will have its quantization info ignored +/// This is because instead of using the same quantization information for all the tensors, the fixture generates +/// separate quantization info for each input and the output tensor. +/// When we can also support dynamic quantization with the presence of activation, these two versions should be merged +/// again, with the explicitly specified quantization info removed TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("QuantizationInfoIfActivationEnabled", { QuantizationInfo(2.f / 255.f, 10) })), QuantizedActivationFunctionsDataset)) { // Validate output @@ -499,7 +1294,7 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEGEMMConvolutionLayerQuantizedFixtur framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("QuantizationInfoIfActivationEnabled", { QuantizationInfo(2.f / 255.f, 10) })), QuantizedActivationFunctionsDataset)) { // Validate output @@ -512,7 +1307,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerQuantizedFixture<int8_t>, framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.01f, -10) })), + framework::dataset::make("QuantizationInfoIfActivationEnabled", { QuantizationInfo(0.01f, -10) })), QuantizedActivationFunctionsDataset)) { // Validate output @@ -529,7 +1324,7 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEGEMMConvolutionLayerQuantizedFixtur framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("QuantizationInfoIfActivationEnabled", { QuantizationInfo(2.f / 255.f, 10) })), QuantizedActivationFunctionsDataset)) { // Validate output @@ -562,6 +1357,27 @@ FIXTURE_DATA_TEST_CASE(RunSmallSigned, NEGEMMConvolutionLayerQuantizedPerChannel // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } + +FIXTURE_DATA_TEST_CASE(MemoryStressLargeChannels, NEGEMMConvolutionLayerQuantizedPerChannelFixture<int8_t>, + framework::DatasetMode::ALL, + combine( + make("In", TensorShape(1U)), + make("Weights", TensorShape(1U, 1U, 1U, 17000U)), + make("Biases", TensorShape(17000U)), + make("Out", TensorShape(1U, 1U, 17000U)), + make("Info", PadStrideInfo(1, 1, 0, 0)), + make("Dilation", Size2D(1, 1)), + make("ReshapeWeights", { true }), + make("DataType", { DataType::QASYMM8_SIGNED }), + make("DataLayout", { DataLayout::NHWC }), + make("QuantizationInfo", QuantizationInfo(0.5f, 10)), + make("ActivationInfo", ActivationLayerInfo()), + make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + TEST_SUITE_END() // QSYMM8_PER_CHANNEL TEST_SUITE_END() // Quantized @@ -571,6 +1387,99 @@ TEST_SUITE(DirectGEMMConv2d) template <typename T> using NEDirectGEMMConv2dLayerFixture = ConvolutionValidationFixture<Tensor, Accessor, NEGEMMConv2d, T>; +/** Test case for memory injection in @ref cpu::CpuGemmDirectConv2d. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) +{ + auto conv = std::make_unique<cpu::CpuGemmDirectConv2d>(); + const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(1U, 7U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto conv_info = Conv2dInfo{}; + conv->configure(&src_info, &weight_info, &bias_info, &dst_info, conv_info); + + // tensors are newly created every call of this lambda function + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &src }, { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + ITensorPack prep_pack{ { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(conv->workspace(), mg, run_pack, prep_pack); + + auto run_conv = [&]() -> Tensor + { + auto dst = create_tensor<Tensor>(dst_info); + dst.allocator()->allocate(); + run_pack.add_tensor(TensorType::ACL_DST, &dst); + + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(weight), 2.f); + library->fill_tensor_value(Accessor(bias), 3.f); + // This operator is configured once and captured by this lambda. + conv->prepare(prep_pack); + conv->run(run_pack); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +/** Test case for memory injection in @ref NEGEMMConv2d. + * + * Make sure @ref NEGEMMConv2d still works through injecting the memory at configure time using the old API. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) +{ + auto conv = std::make_unique<NEGEMMConv2d>(); + const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(1U, 7U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto conv_info = Conv2dInfo{}; + auto run_conv = [&]() + { + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + auto dst = create_tensor<Tensor>(dst_info); + conv->configure(&src, &weight, &bias, &dst, conv_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(weight), 2.f); + library->fill_tensor_value(Accessor(bias), 3.f); + conv->run(); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + TEST_SUITE(Float) TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectGEMMConv2dLayerFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp index 19bd742a61..b4c049f6f9 100644 --- a/tests/validation/NEON/DeconvolutionLayer.cpp +++ b/tests/validation/NEON/DeconvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,55 +47,86 @@ constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for constexpr AbsoluteTolerance<float> tolerance_quantized(1.0f); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC const RelativeTolerance<half_float::half> tolerance_fp16(half_float::half(0.2f)); /**< Relative tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr float tolerance_num_fp16 = 0.02f; /**< Tolerance number for FP16 tests -- follows a slightly stricter approach compared to ConvolutionLayer tests */ #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC*/ -constexpr float tolerance_num = 0.07f; /**< Tolerance number */ +constexpr float tolerance_num_quant = 0.07f; /**< Tolerance number for quantized types */ const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 3) - * framework::dataset::make("PadY", 0, 3) * framework::dataset::make("NumKernels", { 3 }); + * framework::dataset::make("PadY", 0, 3) * framework::dataset::make("NumKernels", +{ + 3 +}); const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2) - * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 }); + * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", +{ + 3 +}); const auto data3x3_asymm = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadLeft", 0, 1) - * framework::dataset::make("PadRight", 0, 1) * framework::dataset::make("PadTop", 0, 1) * framework::dataset::make("PadBottom", 0, 1) * framework::dataset::make("NumKernels", { 3 }); + * framework::dataset::make("PadRight", 0, 1) * framework::dataset::make("PadTop", 0, 1) * framework::dataset::make("PadBottom", 0, 1) * framework::dataset::make("NumKernels", +{ + 3 +}); -const auto data9x9_small_asymm = framework::dataset::make("InputShape", TensorShape{ 10U, 10U, 1U, 1U }) *framework::dataset::make("StrideX", 2) *framework::dataset::make("StrideY", - 2) - *framework::dataset::make("PadLeft", 3) - *framework::dataset::make("PadRight", 4) *framework::dataset::make("PadTop", 3) *framework::dataset::make("PadBottom", 4) *framework::dataset::make("NumKernels", { 1 }); +const auto data9x9_small_asymm = framework::dataset::make("InputShape", TensorShape +{ + 10U, 10U, 1U, 1U +}) +*framework::dataset::make("StrideX", 2) *framework::dataset::make("StrideY", 2) *framework::dataset::make("PadLeft", 3) *framework::dataset::make("PadRight", 4) *framework::dataset::make("PadTop", + 3) *framework::dataset::make("PadBottom", 4) *framework::dataset::make("NumKernels", { 1 }); -const auto data9x9_large_asymm = framework::dataset::make("InputShape", TensorShape{ 640U, 360U, 56U, 1U }) *framework::dataset::make("StrideX", 2) *framework::dataset::make("StrideY", - 2) - *framework::dataset::make("PadLeft", 3) - *framework::dataset::make("PadRight", 4) *framework::dataset::make("PadTop", 3) *framework::dataset::make("PadBottom", 4) *framework::dataset::make("NumKernels", { 1 }); +const auto data9x9_large_asymm = framework::dataset::make("InputShape", TensorShape +{ + 640U, 360U, 56U, 1U +}) +*framework::dataset::make("StrideX", 2) *framework::dataset::make("StrideY", 2) *framework::dataset::make("PadLeft", 3) *framework::dataset::make("PadRight", 4) *framework::dataset::make("PadTop", + 3) *framework::dataset::make("PadBottom", 4) *framework::dataset::make("NumKernels", { 1 }); const auto data3x3_precommit = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 2) * framework::dataset::make("StrideY", 1, 2) * framework::dataset::make("PadX", 0, 2) - * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", { 3 }); + * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("NumKernels", +{ + 3 +}); const auto data1x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 1) - * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("NumKernels", { 3 }); + * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("NumKernels", +{ + 3 +}); -const auto data_layouts_dataset = framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }); +const auto data5x1 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 1) + * framework::dataset::make("PadY", 0, 1) * framework::dataset::make("NumKernels", +{ + 3 +}); + +const auto data_layouts_dataset = framework::dataset::make("DataLayout", +{ + DataLayout::NCHW, DataLayout::NHWC +}); -const auto add_bias_dataset = framework::dataset::make("AddBias", { true, false }); +const auto add_bias_dataset = framework::dataset::make("AddBias", +{ + true, false +}); const auto input_qinfo_dataset = framework::dataset::make("InputQInfo", { QuantizationInfo(1.f / 255.f, 0), - QuantizationInfo(2.f, 0), + QuantizationInfo(2.f, 0), }); const auto output_qinfo_dataset = framework::dataset::make("OutputQInfo", { QuantizationInfo(3.f / 255.f, 0), - QuantizationInfo(4.f, 0), + QuantizationInfo(4.f, 0), }); } // namespace TEST_SUITE(NEON) TEST_SUITE(DeconvolutionLayer) - // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( @@ -105,6 +136,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid bias shape TensorInfo(TensorShape(13U, 11U, 4U, 3U), 1, DataType::F32), // Window shrink TensorInfo(TensorShape(32U, 16U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(2U,2U,1U,1U), 1, DataType::F32), // Small shape no padding + TensorInfo(TensorShape(3U,26U,26U,1U), 1, DataType::F32), // Negative padding }), framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F16), TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32), @@ -112,6 +145,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(3U, 2U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 4U), 1, DataType::F32), TensorInfo(TensorShape(1U, 1U, 2U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(3U,3U,1U,1U), 1, DataType::F32), + TensorInfo(TensorShape(1U,1U,26U,88U), 1, DataType::F32), })), framework::dataset::make("BiasInfo", { TensorInfo(TensorShape(1U), 1, DataType::F16), TensorInfo(TensorShape(1U), 1, DataType::F32), @@ -119,6 +154,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(25U, 11U), 1, DataType::F32), TensorInfo(TensorShape(1U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), + TensorInfo(TensorShape(1U), 1, DataType::F32), + TensorInfo(TensorShape(88U), 1, DataType::F32), })), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F16), TensorInfo(TensorShape(25U, 10U, 2U), 1, DataType::F32), @@ -126,6 +163,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( TensorInfo(TensorShape(13U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(11U, 9U, 1U, 3U), 1, DataType::F32), TensorInfo(TensorShape(32U, 16U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(4U,4U,1U,1U), 1, DataType::F32), + TensorInfo(TensorShape(1U,78U,88U,1U), 1, DataType::F32), })), framework::dataset::make("PadStrideInfo", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), @@ -133,8 +172,10 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 1, 1), PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 3, 3, 1), })), - framework::dataset::make("Expected", { false, false, false, false, false, true })), + framework::dataset::make("Expected", { false, false, false, false, false, true,true, false })), input_info, weights_info, bias_info, output_info, pad_info, expected) { bool is_valid = bool(NEDeconvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pad_info)); @@ -158,6 +199,9 @@ using NEDeconvolutionLayerAsymmFixture9x9 = DeconvolutionValidationAsymmFixture< template <typename T> using NEDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 1, 1>; +template <typename T> +using NEDeconvolutionLayerFixture5x1 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 5, 1>; + TEST_SUITE(Float) TEST_SUITE(FP32) TEST_SUITE(W4x4) @@ -221,6 +265,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerAsymmFixture9x9<float>, fra validate(Accessor(_target), _reference, tolerance_fp32); } TEST_SUITE_END() // W9x9 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture5x1<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data5x1, framework::dataset::make("DataType", DataType::F32)), + data_layouts_dataset), + add_bias_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() // W5x1 TEST_SUITE_END() // FP32 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC @@ -231,7 +284,7 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture4x4<half>, framework::Dat add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_fp16); + validate(Accessor(_target), _reference, tolerance_fp16, tolerance_num_fp16); } TEST_SUITE_END() // W4x4 TEST_SUITE(W3x3) @@ -241,14 +294,14 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEDeconvolutionLayerFixture3x3<half>, framework add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_fp16); + validate(Accessor(_target), _reference, tolerance_fp16, tolerance_num_fp16); } FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerFixture3x3<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::F16)), data_layouts_dataset), add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_fp16); + validate(Accessor(_target), _reference, tolerance_fp16, tolerance_num_fp16); } TEST_SUITE_END() // W3x3 TEST_SUITE(W1x1) @@ -257,9 +310,18 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture1x1<half>, framework::Dat add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_fp16); + validate(Accessor(_target), _reference, tolerance_fp16, tolerance_num_fp16); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture5x1<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(data5x1, framework::dataset::make("DataType", DataType::F16)), + data_layouts_dataset), + add_bias_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp16, tolerance_num_fp16); +} +TEST_SUITE_END() // W5x1 TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ @@ -275,6 +337,9 @@ template <typename T> using NEDeconvolutionLayerQuantizedFixture1x1 = DeconvolutionValidationQuantizedFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 1, 1>; template <typename T> +using NEDeconvolutionLayerQuantizedFixture5x1 = DeconvolutionValidationQuantizedFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 5, 1>; + +template <typename T> using NEDeconvolutionLayerQuantizedPerChannelFixture4x4 = DeconvolutionValidationQuantizedPerChannelFixture<Tensor, Accessor, NEDeconvolutionLayer, T, int8_t, 4, 4>; template <typename T> @@ -283,6 +348,9 @@ using NEDeconvolutionLayerQuantizedPerChannelFixture3x3 = DeconvolutionValidatio template <typename T> using NEDeconvolutionLayerQuantizedPerChannelFixture1x1 = DeconvolutionValidationQuantizedPerChannelFixture<Tensor, Accessor, NEDeconvolutionLayer, T, int8_t, 1, 1>; +template <typename T> +using NEDeconvolutionLayerQuantizedPerChannelFixture5x1 = DeconvolutionValidationQuantizedPerChannelFixture<Tensor, Accessor, NEDeconvolutionLayer, T, int8_t, 5, 1>; + TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) @@ -295,7 +363,7 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture4x4<uint8_t>, fr add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W4x4 @@ -309,7 +377,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEDeconvolutionLayerQuantizedFixture3x3<uint8_t add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data3x3, framework::dataset::make("DataType", @@ -320,7 +388,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerQuantizedFixture3x3<uint8_t add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W3x3 @@ -333,10 +401,23 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture1x1<uint8_t>, fr add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture5x1<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data5x1, framework::dataset::make("DataType", + DataType::QASYMM8)), + data_layouts_dataset), + input_qinfo_dataset), + output_qinfo_dataset), + add_bias_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) @@ -350,7 +431,7 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture4x4<int8_t>, fra add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W4x4 @@ -364,7 +445,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEDeconvolutionLayerQuantizedFixture3x3<int8_t> add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerQuantizedFixture3x3<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data3x3, framework::dataset::make("DataType", @@ -375,7 +456,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDeconvolutionLayerQuantizedFixture3x3<int8_t> add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W3x3 @@ -389,16 +470,41 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture1x1<int8_t>, fra add_bias_dataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedFixture5x1<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(data5x1, framework::dataset::make("DataType", + DataType::QASYMM8_SIGNED)), + data_layouts_dataset), + input_qinfo_dataset), + output_qinfo_dataset), + add_bias_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // QASYMM8_SIGNED -const auto input_qinfo_per_channel_dataset = framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, 10) }); -const auto output_qinfo_per_channel_dataset = framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 0) }); -const auto input_signed_qinfo_per_channel_dataset = framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(1.f / 255.f, -10) }); -const auto output_signed_qinfo_per_channel_dataset = framework::dataset::make("OutputQuantizationInfo", { QuantizationInfo(3.f / 255.f, 10) }); +const auto input_qinfo_per_channel_dataset = framework::dataset::make("InputQuantizationInfo", +{ + QuantizationInfo(1.f / 255.f, 10) +}); +const auto output_qinfo_per_channel_dataset = framework::dataset::make("OutputQuantizationInfo", +{ + QuantizationInfo(3.f / 255.f, 0) +}); +const auto input_signed_qinfo_per_channel_dataset = framework::dataset::make("InputQuantizationInfo", +{ + QuantizationInfo(1.f / 255.f, -10) +}); +const auto output_signed_qinfo_per_channel_dataset = framework::dataset::make("OutputQuantizationInfo", +{ + QuantizationInfo(3.f / 255.f, 10) +}); TEST_SUITE(QSYMM8_PER_CHANNEL) @@ -412,7 +518,7 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedPerChannelFixture4x4<ui framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture4x4<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data4x4, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), @@ -423,7 +529,7 @@ FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W4x4 @@ -437,7 +543,7 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedPerChannelFixture3x3<ui framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture3x3<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data3x3, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), @@ -448,7 +554,7 @@ FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W3x3 @@ -462,7 +568,7 @@ FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedPerChannelFixture1x1<ui framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture1x1<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data1x1, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), @@ -473,10 +579,35 @@ FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) { // Validate output - validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num); + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); } TEST_SUITE_END() // W1x1 +TEST_SUITE(W5x1) +FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerQuantizedPerChannelFixture5x1<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data5x1, + framework::dataset::make("DataType", DataType::QASYMM8)), + data_layouts_dataset), + input_qinfo_per_channel_dataset), + output_qinfo_per_channel_dataset), + add_bias_dataset), + framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); +} +FIXTURE_DATA_TEST_CASE(RunSigned, NEDeconvolutionLayerQuantizedPerChannelFixture5x1<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(data5x1, + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + data_layouts_dataset), + input_signed_qinfo_per_channel_dataset), + output_signed_qinfo_per_channel_dataset), + add_bias_dataset), + framework::dataset::make("WeightsDataType", { DataType::QSYMM8_PER_CHANNEL }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_quantized, tolerance_num_quant); +} +TEST_SUITE_END() // W5x1 + TEST_SUITE_END() // QSYMM8_PER_CHANNEL TEST_SUITE_END() // Quantized diff --git a/tests/validation/NEON/DepthConvertLayer.cpp b/tests/validation/NEON/DepthConvertLayer.cpp index 5649e5a556..4972708144 100644 --- a/tests/validation/NEON/DepthConvertLayer.cpp +++ b/tests/validation/NEON/DepthConvertLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -56,14 +56,12 @@ const auto DepthConvertLayerU16toU8Dataset = combine(framework::dataset::ma const auto DepthConvertLayerU16toU32Dataset = combine(framework::dataset::make("DataType", DataType::U16), framework::dataset::make("DataType", DataType::U32)); const auto DepthConvertLayerS16toU8Dataset = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::U8)); const auto DepthConvertLayerS16toS32Dataset = combine(framework::dataset::make("DataType", DataType::S16), framework::dataset::make("DataType", DataType::S32)); -const auto DepthConvertLayerBF16toF32Dataset = combine(framework::dataset::make("DataType", DataType::BFLOAT16), framework::dataset::make("DataType", DataType::F32)); const auto DepthConvertLayerF16toU8Dataset = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::U8)); const auto DepthConvertLayerF16toF32Dataset = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F32)); const auto DepthConvertLayerF16toS32Dataset = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::S32)); const auto DepthConvertLayerF32toF16Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F16)); const auto DepthConvertLayerF32toS32Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::S32)); const auto DepthConvertLayerF32toU8Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::U8)); -const auto DepthConvertLayerF32toBF16Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::BFLOAT16)); const auto DepthConvertLayerS32toF32Dataset = combine(framework::dataset::make("DataType", DataType::S32), framework::dataset::make("DataType", DataType::F32)); const auto DepthConvertLayerS32toQASYMM8Dataset = combine(framework::dataset::make("DataType", DataType::S32), framework::dataset::make("DataType", DataType::QASYMM8)); @@ -127,8 +125,6 @@ using NEDepthConvertLayerToU8Fixture = DepthConvertLayerValidationFixture<Tensor template <typename T> using NEDepthConvertLayerToU32Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, uint32_t>; template <typename T> -using NEDepthConvertLayerToBF16Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, bfloat16>; -template <typename T> using NEDepthConvertLayerToF16Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, half>; template <typename T> using NEDepthConvertLayerToF32Fixture = DepthConvertLayerValidationFixture<Tensor, Accessor, NEDepthConvertLayer, T, float>; @@ -342,28 +338,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConvertLayerToS32Fixture<int16_t>, frame } TEST_SUITE_END() // S16_to_S32 -#if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) -TEST_SUITE(BFLOAT16_to_F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConvertLayerToF32Fixture<bfloat16>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerBF16toF32Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - DepthConvertLayerZeroShiftDataset)) -{ - // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // BFLOAT16_to_F32 - -TEST_SUITE(F32_to_BFLOAT16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConvertLayerToBF16Fixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerF32toBF16Dataset), - framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), - DepthConvertLayerZeroShiftDataset)) -{ - // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // F32_to_BFLOAT16 -#endif /* defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) */ - #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16_to_QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConvertLayerToQASYMM8Fixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), diff --git a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp index 7260eec42d..e9609b7b72 100644 --- a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp +++ b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,32 +42,70 @@ namespace test { namespace validation { +using framework::dataset::make; using namespace arm_compute::misc::shape_calculator; namespace { -constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ -constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */ +constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */ +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8_SIGNED */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.02)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ constexpr float tolerance_num = 0.05f; /**< Tolerance number */ #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -const auto depth_multipliers = framework::dataset::make("DepthMultiplier", { 1, 2, 8 }); -const auto large_depth_multipliers = framework::dataset::make("DepthMultiplier", { 1, 2, 5, 32 }); +const auto depth_multipliers = make("DepthMultiplier", { 1, 2, 8 }); +const auto large_depth_multipliers = make("DepthMultiplier", { 5, 32 }); -//Activation Functions -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +// Activation Functions +const auto NoActivation = make("ActivationInfo", ActivationLayerInfo()); + +const auto ActivationFunctionsDataset = make("ActivationInfo", { ActivationLayerInfo(), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) }); -const auto input_qinfo_dataset = framework::dataset::make("InputQInfo", +const auto ActivationFunctionsDatasetNightly = make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f, -0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LEAKY_RELU, 0.1f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SOFT_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::ABS), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SQUARE), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::HARD_SWISH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LINEAR, 2.f, 1.f), +#ifdef __aarch64__ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::GELU), +#endif // __aarch64__ +}); + +const auto ActivationFunctionsQuantizedSmallDataset = make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}); + +const auto ActivationFunctionsQuantizedDataset = make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f, -0.5f), +}); + +// This is only used when there is fused activation +const auto input_qinfo_dataset = make("InputQInfo", { QuantizationInfo(0.3f, 10), QuantizationInfo(2.2f, 10), }); + +const auto IgnoredQuantizationInfo = make("IgnoredQuantizationInfo", QuantizationInfo()); + } // namespace TEST_SUITE(NEON) @@ -76,7 +114,7 @@ TEST_SUITE(DepthwiseConvolutionLayer) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32), // Mismatching data type input/weights + make("InputInfo", { TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32), // Mismatching data type input/weights TensorInfo(TensorShape(32U, 18U, 3U), 1, DataType::F32), // Mismatching input feature maps TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32), // Unsupported weights dimensions TensorInfo(TensorShape(32U, 18U, 2U), 1, DataType::F32), // Mismatching depth multiplier @@ -88,7 +126,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // dilation < 1 TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), }), - framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F16), + make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F16), TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(5U, 5U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F32), @@ -100,7 +138,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U, 2U), 1, DataType::F32), })), - framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32), + make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F32), @@ -112,7 +150,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32), + make("OutputInfo", { TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32), TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32), TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32), TensorInfo(TensorShape(30U, 16U, 2U), 1, DataType::F32), @@ -124,7 +162,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), })), - framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), @@ -136,7 +174,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), })), - framework::dataset::make("DepthMultiplier", { 1, + make("DepthMultiplier", { 1, 1, 1, 3, @@ -148,7 +186,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip 1, 1, })), - framework::dataset::make("Dilation", { Size2D(1U, 1U), + make("Dilation", { Size2D(1U, 1U), Size2D(1U, 1U), Size2D(1U, 1U), Size2D(1U, 1U), @@ -160,7 +198,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip Size2D(0U, 1U), Size2D(1U, 1U), })), - framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, false, true })), + make("Expected", { false, false, false, false, false, false, false, false, false, false, true })), input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier,dilation, expected) { bool is_valid = bool(NEDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), @@ -169,7 +207,7 @@ DATA_TEST_CASE(Validate3x3, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip } DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights + make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights TensorInfo(TensorShape(27U, 13U, 3U), 1, DataType::F32), // Mismatching input feature maps TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching depth multiplier TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size @@ -178,7 +216,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // Patch size bigger than input width TensorInfo(TensorShape(27U, 13U, 8U), 1, DataType::F32), // Dilation < 1 }), - framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16), + make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F16), TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 2U), 1, DataType::F32), @@ -187,7 +225,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), TensorInfo(TensorShape(3U, 3U, 16U), 1, DataType::F32), })), - framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32), + make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(4U), 1, DataType::F32), @@ -196,7 +234,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip TensorInfo(TensorShape(16U), 1, DataType::F32), TensorInfo(TensorShape(16U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), + make("OutputInfo", { TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), @@ -205,7 +243,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), TensorInfo(TensorShape(25U, 11U, 16U), 1, DataType::F32), })), - framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), @@ -214,7 +252,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip PadStrideInfo(1, 1, 0, 0), PadStrideInfo(1, 1, 0, 0), })), - framework::dataset::make("DepthMultiplier", { 1, + make("DepthMultiplier", { 1, 1, 3, 1, @@ -223,7 +261,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip 2, 2, })), - framework::dataset::make("Dilation", { Size2D(1U, 1U), + make("Dilation", { Size2D(1U, 1U), Size2D(1U, 1U), Size2D(1U, 1U), Size2D(1U, 1U), @@ -232,7 +270,7 @@ DATA_TEST_CASE(ValidateGeneric, framework::DatasetMode::ALL, zip(zip(zip(zip(zip Size2D(25U, 1U), Size2D(0U, 1U), })), - framework::dataset::make("Expected", { false, false, false, false, false, false, false, false})), + make("Expected", { false, false, false, false, false, false, false, false})), input_info, weights_info, biases_info, output_info, conv_info, depth_multiplier,dilation, expected) { bool is_valid = bool(NEDepthwiseConvolutionLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &biases_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, depth_multiplier, ActivationLayerInfo(), dilation)); @@ -244,43 +282,58 @@ template <typename T> using NEDepthwiseConvolutionLayerFixture = DepthwiseConvolutionLayerValidationFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, T>; template <typename T> using NEDepthwiseConvolutionLayerMixedDataLayoutFixture = DepthwiseConvolutionLayerValidationFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, T, true>; +template <typename T> +using NEDepthwiseConvolutionLayerVariableWeightsFixture = DepthwiseConvolutionLayerValidationFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, T, false, false, true>; TEST_SUITE(Float) TEST_SUITE(F32) + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 5 }), + make("DataType", DataType::F32), + make("DataLayout", { DataLayout::NHWC, DataLayout::NCHW }), + ActivationFunctionsDatasetNightly)) +{ + validate(Accessor(_target), _reference, tolerance_f32); +} + TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout, NEDepthwiseConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - framework::dataset::make("DepthMultiplier", { 2 })), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + make("DepthMultiplier", { 2 })), + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(Accessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("DataType", - DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f32); } - TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f32); @@ -288,10 +341,10 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f32); } @@ -301,9 +354,9 @@ TEST_SUITE_END() // Generic TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f32); @@ -311,10 +364,10 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f32); } @@ -322,9 +375,9 @@ TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f32); @@ -332,10 +385,10 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<float>, FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f32); } @@ -346,40 +399,70 @@ TEST_SUITE_END() // W3x3 TEST_SUITE(Optimized) FIXTURE_DATA_TEST_CASE_NEW(RunSmall3x3, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", + make("DepthMultiplier", 1)), + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsDataset)) +{ + validate(Accessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE_NEW(RunVariableWeightsSmall3x3, NEDepthwiseConvolutionLayerVariableWeightsFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1)), + make("DataType", + DataType::F32)), + make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout3x3, NEDepthwiseConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + make("DepthMultiplier", 1)), + make("DataType", DataType::F32)), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("ActivationInfo", ActivationLayerInfo()))) { validate(Accessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunSmall5x5, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", + make("DepthMultiplier", 1)), + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsDataset)) +{ + validate(Accessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE_NEW(RunVariableWeightsSmall5x5, NEDepthwiseConvolutionLayerVariableWeightsFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), + make("DepthMultiplier", 1)), + make("DataType", + DataType::F32)), + make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f32); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge3x3, NEDepthwiseConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", + make("DepthMultiplier", 1)), + make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) +{ + validate(Accessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE_NEW(RunVariableWeightsLarge3x3, NEDepthwiseConvolutionLayerVariableWeightsFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1)), + make("DataType", + DataType::F32)), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f32); } @@ -388,22 +471,37 @@ TEST_SUITE_END() // F32 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 5 }), + make("DataType", DataType::F16), + make("DataLayout", { DataLayout::NHWC, DataLayout::NCHW }), + ActivationFunctionsDatasetNightly)) +{ + validate(Accessor(_target), _reference, tolerance_f16, tolerance_num); +} + TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f16, tolerance_num); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f16, tolerance_num); } @@ -412,9 +510,8 @@ TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataType", DataType::F16)), + make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f16, tolerance_num); @@ -422,10 +519,9 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, f FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), large_depth_multipliers), - framework::dataset::make("DataType", - DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataType", DataType::F16)), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f16, tolerance_num); } @@ -437,9 +533,9 @@ using NEDepthwiseConvolutionLayerFixture = DepthwiseConvolutionLayerValidationFi TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f16); @@ -447,10 +543,10 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, f FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f16); } @@ -460,9 +556,9 @@ TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f16); @@ -470,10 +566,10 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerFixture<half>, f FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), large_depth_multipliers), - framework::dataset::make("DataType", + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f16); } @@ -484,31 +580,31 @@ TEST_SUITE_END() // W3x3 TEST_SUITE(Optimized) FIXTURE_DATA_TEST_CASE_NEW(RunSmallW3x3, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", + make("DepthMultiplier", 1)), + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f16); } FIXTURE_DATA_TEST_CASE_NEW(RunSmallW5x5, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", + make("DepthMultiplier", 1)), + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_f16); } FIXTURE_DATA_TEST_CASE_NEW(RunLargeW3x3, NEDepthwiseConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", + make("DepthMultiplier", 1)), + make("DataType", DataType::F16)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_f16); } @@ -526,49 +622,88 @@ using NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture = Depthwise TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 5 }), + make("DataType", DataType::QASYMM8), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.05f, 4) }), + make("DataLayout", { DataLayout::NHWC, DataLayout::NCHW }), + ActivationFunctionsQuantizedDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout, NEDepthwiseConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - framework::dataset::make("DepthMultiplier", { 2 })), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + make("DepthMultiplier", { 2 }), + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.8f, 1) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.8f, 1) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.9f, 11) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } @@ -576,47 +711,66 @@ TEST_SUITE_END() // Dilation TEST_SUITE_END() // Generic TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers, + make("DataType", DataType::QASYMM8), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } TEST_SUITE(Dilation) - FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.7f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers, + make("DataType", DataType::QASYMM8), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.7f, 10) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } @@ -625,48 +779,68 @@ TEST_SUITE_END() // W3x3 TEST_SUITE(Optimized) FIXTURE_DATA_TEST_CASE_NEW(RunSmall3x3, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmall3x3WithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunMixedDataLayout3x3, NEDepthwiseConvolutionLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - framework::dataset::make("ActivationInfo", ActivationLayerInfo()))) + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunSmall5x5, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", - DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmall5x5WithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge3x3, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", - DataType::QASYMM8)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NHWC }), + NoActivation)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } @@ -674,143 +848,242 @@ TEST_SUITE_END() // Optimized TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -TEST_SUITE(Generic) -FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 5 }), + make("DataType", DataType::QASYMM8_SIGNED), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.05f, 4) }), + make("DataLayout", { DataLayout::NHWC, DataLayout::NCHW }), + ActivationFunctionsQuantizedDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } +TEST_SUITE(Generic) +FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), - depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.8f, 1) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.8f, 1) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.9f, 11) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } TEST_SUITE_END() // Dilation TEST_SUITE_END() // Generic TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.7f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmallWithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.7f, 10) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), - large_depth_multipliers), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + large_depth_multipliers, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } TEST_SUITE_END() // Dilation TEST_SUITE_END() // W3x3 TEST_SUITE(Optimized) FIXTURE_DATA_TEST_CASE_NEW(RunSmall3x3, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmall3x3WithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8_SIGNED), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } FIXTURE_DATA_TEST_CASE_NEW(RunSmall5x5, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE_NEW(RunSmall5x5WithActivation, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset5x5(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8_SIGNED), + input_qinfo_dataset, + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) }), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }), + ActivationFunctionsQuantizedSmallDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge3x3, NEDepthwiseConvolutionLayerQuantizedFixture<int8_t>, framework::DatasetMode::NIGHTLY, - combine(combine(combine(combine(combine(combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("DataType", - DataType::QASYMM8_SIGNED)), - input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW })), - ActivationFunctionsDataset)) + combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), + make("DepthMultiplier", 1), + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationInfo, + IgnoredQuantizationInfo, + make("DataLayout", { DataLayout::NCHW }), + NoActivation)) { - validate(Accessor(_target), _reference, tolerance_qasymm8); + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } TEST_SUITE_END() // Optimized TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE(QSYMM8_PER_CHANNEL) + +FIXTURE_DATA_TEST_CASE_NEW(RunActivations, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::NIGHTLY, + combine( + make("In", TensorShape(33U, 27U, 11U, 3U)), + make("Weights", Size2D(3U, 4U)), + make("Info", PadStrideInfo(1, 2, 0, 1)), + make("Dilation", Size2D(2U, 2U)), + make("DepthMultiplier", { 5 }), + make("InputDataType", DataType::QASYMM8), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL), + make("SrcQuantizationInfo", { QuantizationInfo(0.3f, 10) }), + make("DstQuantizationInfo", { QuantizationInfo(0.05f, 4) }), + make("DataLayout", { DataLayout::NHWC, DataLayout::NCHW }), + ActivationFunctionsQuantizedDataset)) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + TEST_SUITE(Generic) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("InputDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("InputDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -820,11 +1093,11 @@ TEST_SUITE(Dilation) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("InputDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("InputDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -832,12 +1105,12 @@ FIXTURE_DATA_TEST_CASE_NEW(RunSmall, NEDepthwiseConvolutionLayerQuantizedSymmetr FIXTURE_DATA_TEST_CASE_NEW(RunLarge, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), depth_multipliers), - framework::dataset::make("InputDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("InputDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_qasymm8); } @@ -847,25 +1120,25 @@ TEST_SUITE_END() // Generic TEST_SUITE(Optimized) FIXTURE_DATA_TEST_CASE_NEW(RunSmall3x3, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine(combine(datasets::SmallOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("InputDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("DepthMultiplier", 1)), + make("InputDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE_NEW(RunLarge3x3, NEDepthwiseConvolutionLayerQuantizedSymmetricPerChannelFixture, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(combine(combine(combine(datasets::LargeOptimizedDepthwiseConvolutionLayerDataset3x3(), - framework::dataset::make("DepthMultiplier", 1)), - framework::dataset::make("InputDataType", DataType::QASYMM8)), - framework::dataset::make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), + make("DepthMultiplier", 1)), + make("InputDataType", DataType::QASYMM8)), + make("WeightsDataType", DataType::QSYMM8_PER_CHANNEL)), input_qinfo_dataset), - framework::dataset::make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - ActivationFunctionsDataset)) + make("DstQuantizationInfo", { QuantizationInfo(0.5f, 4) })), + make("DataLayout", { DataLayout::NHWC })), + make("ActivationInfo", { ActivationLayerInfo() }))) { validate(Accessor(_target), _reference, tolerance_qasymm8); } diff --git a/tests/validation/NEON/DepthwiseConvolutionLayerNative.cpp b/tests/validation/NEON/DepthwiseConvolutionLayerNative.cpp index ddf3faacb6..221fc5d249 100644 --- a/tests/validation/NEON/DepthwiseConvolutionLayerNative.cpp +++ b/tests/validation/NEON/DepthwiseConvolutionLayerNative.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,7 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "src/core/cpu/kernels/CpuDepthwiseConv2dNativeKernel.h" +#include "arm_compute/core/utils/StringUtils.h" +#include "src/cpu/kernels/CpuDepthwiseConv2dNativeKernel.h" #include "tests/NEON/Accessor.h" #include "tests/NEON/Helper.h" #include "tests/framework/Macros.h" @@ -134,6 +135,44 @@ TEST_CASE(ValidateNoPadding, framework::DatasetMode::ALL) ARM_COMPUTE_EXPECT(dst.info()->padding().empty(), framework::LogLevel::ERRORS); } +TEST_SUITE(KERNEL_SELECTION) +DATA_TEST_CASE(KernelSelection_mul_and_add, framework::DatasetMode::ALL, + combine(combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::QASYMM8_SIGNED, + DataType::QASYMM8, + DataType::QSYMM8_PER_CHANNEL + })), + framework::dataset::make("DataType_per_channel", { DataType::QASYMM8, + DataType::QASYMM8_SIGNED + })), + cpu_ext, data_type, data_type_per_channel) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuDepthwiseConv2dNativeKernel::get_implementation( + DepthwiseConv2dNativeDataTypeISASelectorData{ data_type, data_type_per_channel,cpu_isa }, + cpu::KernelSelectionType::Preferred ); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string per_channel_str = "_"; + if (data_type == DataType::QSYMM8_PER_CHANNEL) + { + per_channel_str = "_" + cpu_impl_dt(data_type_per_channel) + "_" ; + } + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + per_channel_str + "deptwiseconv2dnative"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() // KERNEL_SELECTION + TEST_SUITE(Float) TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE_NEW(RunSmall, CpuDepthwiseConvolutionNativeFixture<float>, framework::DatasetMode::ALL, diff --git a/tests/validation/NEON/DilatedConvolutionLayer.cpp b/tests/validation/NEON/DilatedConvolutionLayer.cpp index 2f0fce2ce0..fbfe8b8a7a 100644 --- a/tests/validation/NEON/DilatedConvolutionLayer.cpp +++ b/tests/validation/NEON/DilatedConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,6 +26,7 @@ #include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/cpu/operators/CpuConv2d.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/DilatedConvolutionLayerDataset.h" @@ -49,7 +50,7 @@ const AbsoluteTolerance<float> abs_tolerance_f16(0.3f); const RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2f)); /**< Relative tolerance value for comparing reference's output against implementation's output for DataType::F16 */ constexpr float tolerance_num_f16 = 0.07f; /**< Tolerance number for FP16 */ #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -constexpr AbsoluteTolerance<float> tolerance_qasymm8(0.0); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ +constexpr AbsoluteTolerance<int32_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ /** CNN data types */ const auto CNNDataTypes = framework::dataset::make("DataType", @@ -96,7 +97,7 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z framework::dataset::make("Expected", { ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM })), input_info, weights_info, output_info, conv_info, dilation, expected) { - ConvolutionMethod is_valid = NEConvolutionLayer::get_convolution_method(&input_info.clone()->set_is_resizable(false), + ConvolutionMethod is_valid = cpu::CpuConv2d::get_convolution_method(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), conv_info, WeightsInfo(), dilation); @@ -161,13 +162,18 @@ template <typename T> using NEGEMMDilatedConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<Tensor, Accessor, NEGEMMConvolutionLayer, T>; TEST_SUITE(Quantized) +/// @note: Every asymmetric quantized test where there's no fused activation will have its quantization info ignored +/// This is because instead of using the same quantization information for all the tensors, the fixture generates +/// separate quantization info for each input and the output tensor. +/// When we can also support dynamic quantization with the presence of activation, we can remove the explicit +/// quantization info. TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMDilatedConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallDilatedConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("IgnoredQuantizationInfo", { QuantizationInfo() })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output @@ -178,7 +184,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMDilatedConvolutionLayerQuantizedFixture<u framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("IgnoredQuantizationInfo", { QuantizationInfo() })), framework::dataset::make("ActivationLayerInfo", ActivationLayerInfo()))) { // Validate output diff --git a/tests/validation/NEON/DirectConvolutionLayer.cpp b/tests/validation/NEON/DirectConvolutionLayer.cpp index 368aef216a..0779c9d388 100644 --- a/tests/validation/NEON/DirectConvolutionLayer.cpp +++ b/tests/validation/NEON/DirectConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,9 +23,12 @@ */ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NEDirectConvolutionLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuDirectConv2dKernel.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/ShapeDatasets.h" @@ -70,8 +73,8 @@ const auto data_pad_f16 = concat(combine(framework::dataset::make("PadX", { 0, 1 framework::dataset::make("KernelSize", 1)))); const auto data_f32 = combine(datasets::SmallDirectConvolutionShapes(), - combine(framework::dataset::make("StrideX", { 1, 2, 3 }), - combine(framework::dataset::make("StrideY", { 1, 2, 3 }), + combine(framework::dataset::make("StrideX", { 1, 2, 3, 4 }), + combine(framework::dataset::make("StrideY", { 1, 2, 3, 4 }), data_pad_f32))); const auto data_f16 = combine(datasets::SmallDirectConvolutionShapes(), @@ -87,17 +90,25 @@ const auto data_prec = combine(datasets::SmallDirectConvolutionShapes(), framework::dataset::make("KernelSize", 3)))))); const auto data9x9 = combine(datasets::SmallDirectConvolutionShapes(), - combine(framework::dataset::make("StrideX", { 1 }), - combine(framework::dataset::make("StrideY", { 1 }), + combine(framework::dataset::make("StrideX", { 1, 2, 3 }), + combine(framework::dataset::make("StrideY", { 1, 2, 3 }), combine(framework::dataset::make("PadX", { 0, 2 }), combine(framework::dataset::make("PadY", { 0, 3 }), framework::dataset::make("KernelSize", 9)))))); -const auto data_f32_nightly = combine(data_f32, framework::dataset::make("NumKernels", { 1, 4 })); -const auto data_f16_nightly = combine(data_f16, framework::dataset::make("NumKernels", { 1, 4 })); +const auto data8x8 = combine(datasets::SmallDirectConvolutionShapes(), + combine(framework::dataset::make("StrideX", { 1, 2, 3 }), + combine(framework::dataset::make("StrideY", { 1, 2, 3 }), + combine(framework::dataset::make("PadX", { 0 }), + combine(framework::dataset::make("PadY", { 0 }), + framework::dataset::make("KernelSize", 8)))))); + +const auto data_f32_nightly = combine(data_f32, framework::dataset::make("NumKernels", { 1, 4, 5 })); +const auto data_f16_nightly = combine(data_f16, framework::dataset::make("NumKernels", { 1, 4, 5 })); const auto data_precommit = combine(data_prec, framework::dataset::make("NumKernels", { 1 })); const auto data_precommit9x9 = combine(data9x9, framework::dataset::make("NumKernels", { 4 })); +const auto data_precommit8x8 = combine(data8x8, framework::dataset::make("NumKernels", { 4 })); /* The following tests is from real use-case that made DirectConvolution * overflows in terms of its tensor indexing. This test case is using @@ -172,17 +183,52 @@ TEST_CASE(NoBias, framework::DatasetMode::PRECOMMIT) validate(Accessor(dst), ref_dst); } +DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, + concat(combine(combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32 })), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + combine(combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F16 })), + framework::dataset::make("DataLayout", { DataLayout::NCHW }))), + cpu_ext, data_type, data_layout) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuDirectConv2dKernel::get_implementation(DataTypeDataLayoutISASelectorData{ data_type, data_layout, cpu_isa }, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string data_layout_str; + if(data_layout == DataLayout::NCHW) + { + data_layout_str = "nchw"; + } + else + { + data_layout_str = "nhwc"; + } + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_" + data_layout_str + "_directconv2d"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching data type input/weights - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Mismatching input feature maps + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid: Mismatching data type input/weights + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid: Mismatching input feature maps TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported kernel width - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Non-rectangular weights dimensions + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported non-rectangular weights dimensions TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid weights dimensions - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid stride - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases size - TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid biases dimensions + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported stride + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported biases size + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Unsupported biases dimensions TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Invalid output size }), framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F16), @@ -319,13 +365,23 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectConvolutionLayerFixture<float>, framewo validate(Accessor(_target), _reference, tolerance_fp32); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEDirectConvolutionLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit, - framework::dataset::make("DataType", DataType::F32)), - ActivationFunctionsDataset), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) + framework::dataset::make("DataType", DataType::F32)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32); } + +FIXTURE_DATA_TEST_CASE(RunSmall8x8, NEDirectConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit8x8, framework::dataset::make("DataType", + DataType::F32)), + ActivationFunctionsDataset), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} + FIXTURE_DATA_TEST_CASE(RunSmall9x9, NEDirectConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(data_precommit9x9, framework::dataset::make("DataType", DataType::F32)), ActivationFunctionsDataset), diff --git a/tests/validation/NEON/ElementwiseAbsoluteValue.cpp b/tests/validation/NEON/ElementwiseAbsoluteValue.cpp index ccde670034..0667ac73f9 100644 --- a/tests/validation/NEON/ElementwiseAbsoluteValue.cpp +++ b/tests/validation/NEON/ElementwiseAbsoluteValue.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,13 @@ RelativeTolerance<float> tolerance_fp32(0.000001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<float> tolerance_fp16(0.01f); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#if defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); +#else // #if !defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO" +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); +#endif // #if !defined(__aarch64__) } // namespace TEST_SUITE(NEON) @@ -53,6 +60,9 @@ TEST_SUITE(AbsLayer) template <typename T> using NEAbsLayerFixture = AbsValidationFixture<Tensor, Accessor, NEAbsLayer, T>; +template <typename T> +using NEAbsLayerQuantizedFixture = AbsQuantizedValidationFixture<Tensor, Accessor, NEAbsLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -107,6 +117,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEAbsLayerFixture<int32_t>, framework::DatasetM TEST_SUITE_END() // S32 TEST_SUITE_END() // Integer +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEAbsLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.2, -3) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.5, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NEAbsLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.075, 6) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.1, -7) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // AbsLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwiseDivision.cpp b/tests/validation/NEON/ElementwiseDivision.cpp index 8abccb2ed6..95db4ad5fd 100644 --- a/tests/validation/NEON/ElementwiseDivision.cpp +++ b/tests/validation/NEON/ElementwiseDivision.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,7 @@ namespace validation namespace { RelativeTolerance<float> tolerance_fp32(0.000001f); -AbsoluteTolerance<int> tolerance_zero_s32(1); // Tolerance for S32 division +AbsoluteTolerance<int> tolerance_zero_s32(0); // Tolerance for S32 division /** Input data sets **/ const auto ElementwiseDivisionS32Dataset = combine(combine(framework::dataset::make("DataType", DataType::S32), @@ -56,6 +56,8 @@ const auto ElementwiseDivisionFP16Dataset = combine(combine(framewo #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const auto ElementwiseDivisionFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataType", DataType::F32)); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(NEON) @@ -105,14 +107,16 @@ using CpuElementwiseDivisionBroadcastDynamicShapeFixture = ArithmeticDivisionBro TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, CpuElementwiseDivisionDynamicShapeFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseDivisionFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, CpuElementwiseDivisionDynamicShapeFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseDivisionFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); } -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CpuElementwiseDivisionBroadcastDynamicShapeFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapesBroadcast(), - ElementwiseDivisionFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CpuElementwiseDivisionBroadcastDynamicShapeFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseDivisionFP32Dataset), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -124,7 +128,8 @@ TEST_SUITE_END() // DynamicShape TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseDivisionFP16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseDivisionFP16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp16, 0.01); @@ -133,7 +138,8 @@ TEST_SUITE_END() // F16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseDivisionFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseDivisionFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -142,8 +148,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<float>, framework: template <typename T> using NEElementwiseDivisionBroadcastFixture = ArithmeticDivisionBroadcastValidationFixture<Tensor, Accessor, NEElementwiseDivision, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseDivisionBroadcastFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapesBroadcast(), - ElementwiseDivisionFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseDivisionBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseDivisionFP32Dataset), + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32, 0.01); +} +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEElementwiseDivisionBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcastInplace(), + ElementwiseDivisionFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -153,7 +167,8 @@ TEST_SUITE_END() // Float TEST_SUITE(Integer) TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<int32_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseDivisionS32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseDivisionFixture<int32_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseDivisionS32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_zero_s32); @@ -162,7 +177,7 @@ TEST_SUITE_END() // S32 TEST_SUITE_END() // Integer TEST_SUITE_END() // ElementwiseDivision -TEST_SUITE_END() // Neon +TEST_SUITE_END() // NEON } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/NEON/ElementwiseExpLayer.cpp b/tests/validation/NEON/ElementwiseExpLayer.cpp index f9e5f39989..31cd78626f 100644 --- a/tests/validation/NEON/ElementwiseExpLayer.cpp +++ b/tests/validation/NEON/ElementwiseExpLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,15 @@ RelativeTolerance<float> tolerance_fp32(0.000001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<float> tolerance_fp16(0.01f); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#if defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); +#else // #if !defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO" +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); +#endif // #if !defined(__aarch64__) + } // namespace TEST_SUITE(NEON) TEST_SUITE(ExpLayer) @@ -53,6 +62,9 @@ TEST_SUITE(ExpLayer) template <typename T> using NEExpLayerFixture = ExpValidationFixture<Tensor, Accessor, NEExpLayer, T>; +template <typename T> +using NEExpLayerQuantizedFixture = ExpQuantizedValidationFixture<Tensor, Accessor, NEExpLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -82,6 +94,32 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEExpLayerFixture<float>, framework::DatasetMod TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEExpLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.01, 0) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.003, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NEExpLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.02, -1) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.002, -2) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // ExpLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwiseKernelSelection.cpp b/tests/validation/NEON/ElementwiseKernelSelection.cpp new file mode 100644 index 0000000000..7990a51936 --- /dev/null +++ b/tests/validation/NEON/ElementwiseKernelSelection.cpp @@ -0,0 +1,159 @@ +/* + * Copyright (c) 2022-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuElementwiseKernel.h" +#include "src/cpu/kernels/CpuElementwiseUnaryKernel.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(NEON) +TEST_SUITE(KernelSelection) + +DATA_TEST_CASE(KernelSelection_elementwise_unary, framework::DatasetMode::ALL, concat( + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::S32 + })), + combine(framework::dataset::make("CpuExt", std::string("SVE")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::S32 + }))), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.sve = (cpu_ext == "SVE"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuElementwiseUnaryKernel::get_implementation(DataTypeISASelectorData{ data_type, cpu_isa }, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_elementwise_unary"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + +DATA_TEST_CASE(KernelSelection_elementwise_arithmetic, framework::DatasetMode::ALL, concat(concat( + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::S32, + DataType::S16, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED + })), + combine(framework::dataset::make("CpuExt", std::string("SVE")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::S32, + DataType::S16 + }))), + combine(framework::dataset::make("CpuExt", std::string("SVE2")), + framework::dataset::make("DataType", { DataType::QASYMM8, + DataType::QASYMM8_SIGNED + }))), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.sve = (cpu_ext == "SVE"); + cpu_isa.sve2 = (cpu_ext == "SVE2"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuArithmeticKernel::get_implementation( + ElementwiseDataTypeISASelectorData{ data_type, cpu_isa, static_cast<int>(ArithmeticOperation::ADD) }, + cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_arithmetic"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + +DATA_TEST_CASE(KernelSelection_elementwise_comparison, framework::DatasetMode::ALL, concat(concat( + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::S32, + DataType::S16, + DataType::U8, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED + })), + combine(framework::dataset::make("CpuExt", std::string("SVE")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::S32, + DataType::S16, + DataType::U8 + }))), + combine(framework::dataset::make("CpuExt", std::string("SVE2")), + framework::dataset::make("DataType", { DataType::QASYMM8, + DataType::QASYMM8_SIGNED + }))), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.sve = (cpu_ext == "SVE"); + cpu_isa.sve2 = (cpu_ext == "SVE2"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuComparisonKernel::get_implementation( + ElementwiseDataTypeISASelectorData{ data_type, cpu_isa, static_cast<int>(ComparisonOperation::Equal) }, + cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_comparison"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + +TEST_SUITE_END() +TEST_SUITE_END() // Neon +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/ElementwiseLog.cpp b/tests/validation/NEON/ElementwiseLog.cpp index 3aa7fb3665..1175903dac 100644 --- a/tests/validation/NEON/ElementwiseLog.cpp +++ b/tests/validation/NEON/ElementwiseLog.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,15 @@ RelativeTolerance<float> tolerance_fp32(0.000001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<float> tolerance_fp16(0.01f); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC + +#if defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); +#else // #if !defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO" +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); +#endif // #if !defined(__aarch64__) + } // namespace TEST_SUITE(NEON) TEST_SUITE(LogLayer) @@ -53,6 +62,9 @@ TEST_SUITE(LogLayer) template <typename T> using NELogLayerFixture = LogValidationFixture<Tensor, Accessor, NELogLayer, T>; +template <typename T> +using NELogLayerQuantizedFixture = LogQuantizedValidationFixture<Tensor, Accessor, NELogLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -88,6 +100,33 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NELogLayerFixture<float>, framework::DatasetMod } TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NELogLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(10.5, 0), QuantizationInfo(0.5, -10) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(5, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NELogLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.75, -128) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(12.5, -2) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // LogLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwiseMax.cpp b/tests/validation/NEON/ElementwiseMax.cpp index 4bc263184e..61421ab3e5 100644 --- a/tests/validation/NEON/ElementwiseMax.cpp +++ b/tests/validation/NEON/ElementwiseMax.cpp @@ -62,6 +62,8 @@ const auto ElementwiseMaxFP16Dataset = combine(combine(framework::dataset::make( #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const auto ElementwiseMaxFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataType", DataType::F32)); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(NEON) @@ -111,7 +113,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( // *INDENT-ON* TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), ElementwiseMaxS32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ElementwiseMaxS32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -119,7 +122,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<int32_t>, framework::Da TEST_SUITE_END() // S32 TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMaxS16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMaxS16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -131,11 +135,12 @@ using NEElementwiseMaxQuantizedFixture = ElementwiseMaxValidationQuantizedFixtur TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxQASYMM8Dataset), framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -144,11 +149,13 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxQuantizedFixture<uint8_t>, fram template <typename T> using NEElementwiseMaxQuantizedBroadcastFixture = ElementwiseMaxQuantizedBroadcastValidationFixture<Tensor, Accessor, NEElementwiseMax, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMaxQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapesBroadcast(), - ElementwiseMaxQASYMM8Dataset), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) }))) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMaxQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseMaxQASYMM8Dataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -156,16 +163,26 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMaxQuantizedBroadcastFixt TEST_SUITE_END() TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxQASYMM8SignedDataset), framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, 20) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f, 0) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f, -27) }))) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f, -27) })), + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE(RunSmallInPlace, NEElementwiseMaxQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), + ElementwiseMaxQASYMM8SignedDataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, -20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, -20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, -20) })), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } - TEST_SUITE_END() TEST_SUITE_END() @@ -173,7 +190,8 @@ TEST_SUITE_END() TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMaxFP16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMaxFP16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -182,7 +200,8 @@ TEST_SUITE_END() // F16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMaxFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMaxFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -190,8 +209,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMaxFixture<float>, framework::Data template <typename T> using NEElementwiseMaxBroadcastFixture = ElementwiseMaxBroadcastValidationFixture<Tensor, Accessor, NEElementwiseMax, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMaxBroadcastFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapesBroadcast(), - ElementwiseMaxFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMaxBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseMaxFP32Dataset), + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEElementwiseMaxBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcastInplace(), + ElementwiseMaxFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); diff --git a/tests/validation/NEON/ElementwiseMin.cpp b/tests/validation/NEON/ElementwiseMin.cpp index 3836b90308..a134eb354d 100644 --- a/tests/validation/NEON/ElementwiseMin.cpp +++ b/tests/validation/NEON/ElementwiseMin.cpp @@ -62,6 +62,8 @@ const auto ElementwiseMinFP16Dataset = combine(combine(framework::dataset::make( #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const auto ElementwiseMinFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataType", DataType::F32)); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(NEON) @@ -110,7 +112,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( // *INDENT-ON* TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), ElementwiseMinS32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ElementwiseMinS32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -118,7 +121,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<int32_t>, framework::Da TEST_SUITE_END() // S32 TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMinS16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMinS16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -133,23 +137,34 @@ TEST_SUITE(QASYMM8) template <typename T> using NEElementwiseMinQuantizedBroadcastFixture = ElementwiseMinQuantizedBroadcastValidationFixture<Tensor, Accessor, NEElementwiseMin, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMinQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapesBroadcast(), - ElementwiseMinQASYMM8Dataset), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) }))) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMinQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseMinQASYMM8Dataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); } - -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEElementwiseMinQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(combine(datasets::TinyShapesBroadcastInplace(), + ElementwiseMinQASYMM8Dataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 20) })), + InPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMinQASYMM8Dataset), framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })) - - ) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -157,11 +172,12 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinQuantizedFixture<uint8_t>, fram TEST_SUITE_END() TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseMaxQASYMM8SignedDataset), framework::dataset::make("QuantizationInfo", { QuantizationInfo(10.f, 20) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f, 0) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f, -27) }))) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f, -27) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_signed); @@ -174,7 +190,8 @@ TEST_SUITE_END() TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMinFP16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMinFP16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -183,7 +200,8 @@ TEST_SUITE_END() // F16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseMinFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseMinFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -192,8 +210,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseMinFixture<float>, framework::Data template <typename T> using NEElementwiseMinBroadcastFixture = ElementwiseMinBroadcastValidationFixture<Tensor, Accessor, NEElementwiseMin, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMinBroadcastFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapesBroadcast(), - ElementwiseMinFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseMinBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseMinFP32Dataset), + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEElementwiseMinBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcastInplace(), + ElementwiseMinFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); diff --git a/tests/validation/NEON/ElementwiseNegation.cpp b/tests/validation/NEON/ElementwiseNegation.cpp index 0b63588d8a..5b8ae8fc64 100644 --- a/tests/validation/NEON/ElementwiseNegation.cpp +++ b/tests/validation/NEON/ElementwiseNegation.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,13 @@ RelativeTolerance<float> tolerance_fp32(0.000001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<float> tolerance_fp16(0.01f); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#if defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); +#else // #if !defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO" +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); +#endif // #if !defined(__aarch64__) } // namespace TEST_SUITE(NEON) TEST_SUITE(NegLayer) @@ -53,6 +60,9 @@ TEST_SUITE(NegLayer) template <typename T> using NENegLayerFixture = NegValidationInPlaceFixture<Tensor, Accessor, NENegLayer, T>; +template <typename T> +using NENegLayerQuantizedFixture = NegQuantizedValidationFixture<Tensor, Accessor, NENegLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -113,6 +123,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NENegLayerFixture<int32_t>, framework::DatasetM TEST_SUITE_END() // S32 TEST_SUITE_END() // Integer +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NENegLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.2, -3) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.5, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NENegLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.075, 6) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.1, -7) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // NegLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwisePower.cpp b/tests/validation/NEON/ElementwisePower.cpp index 4305387c5f..9ac9eec280 100644 --- a/tests/validation/NEON/ElementwisePower.cpp +++ b/tests/validation/NEON/ElementwisePower.cpp @@ -51,6 +51,8 @@ const auto ElementwisePowerFP16Dataset = combine(combine(framework: #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const auto ElementwisePowerFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataType", DataType::F32)); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(NEON) @@ -91,7 +93,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwisePowerFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwisePowerFP16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwisePowerFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwisePowerFP16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp16, 0.01); @@ -101,13 +104,15 @@ TEST_SUITE_END() // F16 TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwisePowerFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwisePowerFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwisePowerFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwisePowerFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEElementwisePowerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), ElementwisePowerFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEElementwisePowerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), ElementwisePowerFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -116,15 +121,23 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEElementwisePowerFixture<float>, framework::Da template <typename T> using NEElementwisePowerBroadcastFixture = ElementwisePowerBroadcastValidationFixture<Tensor, Accessor, NEElementwisePower, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapesBroadcast(), - ElementwisePowerFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapesBroadcast(), + ElementwisePowerFP32Dataset), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); } - -FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapesBroadcast(), - ElementwisePowerFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapesBroadcastInplace(), + ElementwisePowerFP32Dataset), + InPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32, 0.01); +} +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEElementwisePowerBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(), + ElementwisePowerFP32Dataset), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); diff --git a/tests/validation/NEON/ElementwiseRound.cpp b/tests/validation/NEON/ElementwiseRound.cpp index d2f0b456a0..620618cb0b 100644 --- a/tests/validation/NEON/ElementwiseRound.cpp +++ b/tests/validation/NEON/ElementwiseRound.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -40,12 +40,20 @@ namespace test { namespace validation { +namespace +{ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); +} // namespace TEST_SUITE(NEON) TEST_SUITE(RoundLayer) template <typename T> using NERoundLayerFixture = RoundValidationFixture<Tensor, Accessor, NERoundLayer, T>; +template <typename T> +using NERoundLayerQuantizedFixture = RoundQuantizedValidationFixture<Tensor, Accessor, NERoundLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -81,6 +89,33 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NERoundLayerFixture<float>, framework::DatasetM } TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NERoundLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.2, -3) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.5, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NERoundLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.075, 6) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.1, -7) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // RoundLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwiseRsqrtLayer.cpp b/tests/validation/NEON/ElementwiseRsqrtLayer.cpp index 2d52183b15..80788c893f 100644 --- a/tests/validation/NEON/ElementwiseRsqrtLayer.cpp +++ b/tests/validation/NEON/ElementwiseRsqrtLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,13 @@ RelativeTolerance<float> tolerance_fp32(0.000001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<float> tolerance_fp16(0.01f); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#if defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); +#else // #if !defined(__aarch64__) +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); // There is difference of 1, because quantizing in reference uses round policy "TO_NEAREST_UP", where the armv7a neon kernel uses "TO_ZERO" +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); +#endif // #if !defined(__aarch64__) } // namespace TEST_SUITE(NEON) TEST_SUITE(RsqrtLayer) @@ -72,6 +79,9 @@ TEST_SUITE_END() // DynamicShape template <typename T> using NERsqrtLayerFixture = RsqrtValidationFixture<Tensor, Accessor, NERsqrtLayer, T>; +template <typename T> +using NERsqrtLayerQuantizedFixture = RsqrtQuantizedValidationFixture<Tensor, Accessor, NERsqrtLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -102,6 +112,32 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NERsqrtLayerFixture<float>, framework::DatasetM TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NERsqrtLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(20, 0) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.5, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NERsqrtLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(25, -128) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(0.1, -7) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized + TEST_SUITE_END() // RsqrtLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwiseSin.cpp b/tests/validation/NEON/ElementwiseSin.cpp index 06775c0690..9c2d7ae268 100644 --- a/tests/validation/NEON/ElementwiseSin.cpp +++ b/tests/validation/NEON/ElementwiseSin.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,6 +46,8 @@ AbsoluteTolerance<float> tolerance_fp32(0.00001f); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC AbsoluteTolerance<float> tolerance_fp16(0.0005f); #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(0); +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(0); } // namespace TEST_SUITE(NEON) TEST_SUITE(SinLayer) @@ -53,6 +55,9 @@ TEST_SUITE(SinLayer) template <typename T> using NESinLayerFixture = SinValidationFixture<Tensor, Accessor, NESinLayer, T>; +template <typename T> +using NESinLayerQuantizedFixture = SinQuantizedValidationFixture<Tensor, Accessor, NESinLayer, T>; + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) @@ -89,6 +94,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NESinLayerFixture<float>, framework::DatasetMod TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NESinLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.2, -3) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(200, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunSmall, NESinLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine( + datasets::SmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("InputQInfo", { QuantizationInfo(0.07, 6) })), + framework::dataset::make("OutputQInfo", { QuantizationInfo(123, -7) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE_END() // Quantized TEST_SUITE_END() // SinLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/ElementwiseSquareDiff.cpp b/tests/validation/NEON/ElementwiseSquareDiff.cpp index 069cbbd7fa..9a86b541de 100644 --- a/tests/validation/NEON/ElementwiseSquareDiff.cpp +++ b/tests/validation/NEON/ElementwiseSquareDiff.cpp @@ -68,6 +68,8 @@ const auto ElementwiseSquaredDiffFP16Dataset = combine(combine(framework::datase #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ const auto ElementwiseSquaredDiffFP32Dataset = combine(combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataType", DataType::F32)); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); } // namespace TEST_SUITE(NEON) @@ -109,7 +111,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( // *INDENT-ON* TEST_SUITE(S32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapes(), ElementwiseSquaredDiffS32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<int32_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffS32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -117,7 +120,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<int32_t>, frame TEST_SUITE_END() // S32 TEST_SUITE(S16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseSquaredDiffS16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffS16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -129,13 +133,12 @@ using NEElementwiseSquaredDiffQuantizedFixture = ElementwiseSquaredDiffValidatio TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffQASYMM8Dataset), framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })) - - ) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp32, 0.01); @@ -144,11 +147,23 @@ template <typename T> using NEElementwiseSquaredDiffQuantizedBroadcastFixture = ElementwiseSquaredDiffQuantizedBroadcastValidationFixture<Tensor, Accessor, NEElementwiseSquaredDiff, T>; FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseSquaredDiffQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(datasets::SmallShapesBroadcast(), - ElementwiseSquaredDiffQASYMM8Dataset), + combine(combine(combine(combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseSquaredDiffQASYMM8Dataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) })), + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunTinyBroadcastInPlace, NEElementwiseSquaredDiffQuantizedBroadcastFixture<uint8_t>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(datasets::TinyShapesBroadcastInplace(), + ElementwiseSquaredDiffQASYMM8Dataset), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f / 255.f, 5) }))) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(5.f / 255.f, 20) })), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -156,11 +171,12 @@ FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseSquaredDiffQuantizedBroad TEST_SUITE_END() TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffQASYMM8SignedDataset), framework::dataset::make("QuantizationInfo", { QuantizationInfo(1.f, 5) })), framework::dataset::make("QuantizationInfo", { QuantizationInfo(.5f, 5) })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(.2f, 5) }))) + framework::dataset::make("QuantizationInfo", { QuantizationInfo(.2f, 5) })), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -171,7 +187,8 @@ TEST_SUITE_END() TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(F16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP16Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<half>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP16Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_fp16, 0.01); @@ -180,7 +197,8 @@ TEST_SUITE_END() // F16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(F32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<float>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), ElementwiseSquaredDiffFP32Dataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -188,15 +206,17 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEElementwiseSquaredDiffFixture<float>, framewo template <typename T> using NEElementwiseSquaredDiffBroadcastFixture = ElementwiseSquaredDiffBroadcastValidationFixture<Tensor, Accessor, NEElementwiseSquaredDiff, T>; -FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseSquaredDiffBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallShapesBroadcast(), - ElementwiseSquaredDiffFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEElementwiseSquaredDiffBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(), + ElementwiseSquaredDiffFP32Dataset), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEElementwiseSquaredDiffBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapesBroadcast(), - ElementwiseSquaredDiffFP32Dataset)) +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEElementwiseSquaredDiffBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(), + ElementwiseSquaredDiffFP32Dataset), + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); diff --git a/tests/validation/NEON/FillBorder.cpp b/tests/validation/NEON/FillBorder.cpp index 343ad831e4..928990b2b4 100644 --- a/tests/validation/NEON/FillBorder.cpp +++ b/tests/validation/NEON/FillBorder.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -60,10 +60,10 @@ DATA_TEST_CASE(FillBorder, framework::DatasetMode::ALL, combine(combine(combine( { BorderSize border_size{ static_cast<unsigned int>(size) }; - std::mt19937 generator(library->seed()); - std::uniform_int_distribution<uint8_t> distribution_u8(0, 255); - const uint8_t border_value = distribution_u8(generator); - const uint8_t tensor_value = distribution_u8(generator); + std::mt19937 generator(library->seed()); + std::uniform_int_distribution<uint32_t> distribution_u8(0, 255); + const uint8_t border_value = distribution_u8(generator); + const uint8_t tensor_value = distribution_u8(generator); // Create tensors Tensor src = create_tensor<Tensor>(shape, data_type); @@ -77,7 +77,7 @@ DATA_TEST_CASE(FillBorder, framework::DatasetMode::ALL, combine(combine(combine( validate(src.info()->padding(), padding); // Fill tensor with constant value - std::uniform_int_distribution<uint8_t> distribution{ tensor_value, tensor_value }; + std::uniform_int_distribution<uint32_t> distribution{ tensor_value, tensor_value }; library->fill(Accessor(src), distribution, 0); // Create and configure kernel diff --git a/tests/validation/NEON/Floor.cpp b/tests/validation/NEON/Floor.cpp index 419ce56e44..3cd1033ef9 100644 --- a/tests/validation/NEON/Floor.cpp +++ b/tests/validation/NEON/Floor.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,9 +22,12 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NEFloor.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuFloorKernel.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/ShapeDatasets.h" @@ -62,6 +65,30 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( const Status status = NEFloor::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false)); ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); } + + +DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + })), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuFloorKernel::get_implementation(DataTypeISASelectorData{data_type, cpu_isa}, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_floor"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} // clang-format on // *INDENT-ON* diff --git a/tests/validation/NEON/FullyConnectedLayer.cpp b/tests/validation/NEON/FullyConnectedLayer.cpp index 4bb48bf42c..ee7e56227d 100644 --- a/tests/validation/NEON/FullyConnectedLayer.cpp +++ b/tests/validation/NEON/FullyConnectedLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,6 +25,8 @@ #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/cpu/operators/CpuFullyConnected.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/FullyConnectedLayerDataset.h" @@ -40,6 +42,7 @@ namespace test { namespace validation { +using framework::dataset::make; namespace { /** Tolerance for float operations */ @@ -56,7 +59,7 @@ constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); /** CNN data types */ -const auto CNNDataTypes = framework::dataset::make("DataType", +const auto CNNDataTypes = make("DataType", { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC DataType::F16, @@ -64,18 +67,25 @@ const auto CNNDataTypes = framework::dataset::make("DataType", DataType::F32, }); -const auto FullyConnectedParameters = combine(framework::dataset::make("TransposeWeights", { false, true }), framework::dataset::make("ReshapeWeights", { false, true })); +const auto FullyConnectedParameters = combine(make("TransposeWeights", { false, true }), make("ReshapeWeights", { false, true })); -const auto QuantizationData = framework::dataset::make("QuantizationInfo", +const auto QuantizationData = make("QuantizationInfo", { QuantizationInfo(1.f / 256.f, 10), QuantizationInfo(1.1f, 10), }); -const auto EmptyActivationFunctionDataset = framework::dataset::make("ActivationInfo", + +const auto IgnoredQuantizationData = make("IgnoredQuantizationInfo", +{ + QuantizationInfo(), +}); + +const auto NoActivationFunctionDataset = make("ActivationInfo", { ActivationLayerInfo(), }); -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", + +const auto ActivationFunctionsDataset = make("ActivationInfo", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), @@ -83,7 +93,7 @@ const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), }); -const auto ActivationFunctionsQuantizedDataset = framework::dataset::make("ActivationInfo", +const auto ActivationFunctionsQuantizedDataset = make("ActivationInfo", { ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 0.5f), @@ -94,40 +104,183 @@ const auto ActivationFunctionsQuantizedDataset = framework::dataset::make("Activ TEST_SUITE(NEON) TEST_SUITE(FullyConnectedLayer) +/** Test case for memory injection in @ref cpu::CpuFullyConnected. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) +{ + auto fc = std::make_unique<cpu::CpuFullyConnected>(); + const auto src_info = TensorInfo(TensorShape(8U), 1, DataType::F32, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(8U, 4U), 1, DataType::F32, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC); + const auto fc_info = FullyConnectedLayerInfo{}; + fc->configure(&src_info, &weight_info, &bias_info, &dst_info, fc_info); + + // telhs are newly created every call of this lambda function + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &src }, { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + ITensorPack prep_pack{ { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(fc->workspace(), mg, run_pack, prep_pack); + + auto run_conv = [&]() -> Tensor + { + auto dst = create_tensor<Tensor>(dst_info); + dst.allocator()->allocate(); + run_pack.add_tensor(TensorType::ACL_DST, &dst); + + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(weight), 2.f); + library->fill_tensor_value(Accessor(bias), 3.f); + // This operator is configured once and captured by this lambda. + fc->prepare(prep_pack); + fc->run(run_pack); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +/** Test case for memory injection in @ref NEFullyConnectedLayer. + * + * Make sure @ref NEFullyConnectedLayer still works through injecting the memory at configure time using the old API. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) +{ + auto fc = std::make_unique<NEFullyConnectedLayer>(); + const auto src_info = TensorInfo(TensorShape(8U), 1, DataType::F32, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(8U, 4U), 1, DataType::F32, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC); + const auto fc_info = FullyConnectedLayerInfo{}; + auto run_conv = [&]() + { + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + auto dst = create_tensor<Tensor>(dst_info); + fc->configure(&src, &weight, &bias, &dst, fc_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + library->fill_tensor_value(Accessor(src), 1.f); + library->fill_tensor_value(Accessor(weight), 2.f); + library->fill_tensor_value(Accessor(bias), 3.f); + fc->run(); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +/** Unit test for @ref cpu::CpuFullyConnected with quantized multipler > 1 + * + * Tests output correctness. + */ +TEST_CASE(Quant8_Signed_Mult_gt_1, framework::DatasetMode::ALL) +{ + auto fc = std::make_unique<cpu::CpuFullyConnected>(); + const auto src_info = TensorInfo(TensorShape(1U, 3U), 1, DataType::QASYMM8_SIGNED, QuantizationInfo(0.5f, -1)); + const auto weight_info = TensorInfo(TensorShape(1U), 1, DataType::QASYMM8_SIGNED, QuantizationInfo(0.5, -8)); + const auto bias_info = TensorInfo(TensorShape(1U), 1, DataType::S32); + auto dst_info = TensorInfo(TensorShape(1U, 3U), 1, DataType::QASYMM8_SIGNED, QuantizationInfo(0.1f, 0)); + const auto fc_info = FullyConnectedLayerInfo{}; + fc->configure(&src_info, &weight_info, &bias_info, &dst_info, fc_info); + + // telhs are newly created every call of this lambda function + auto src = create_tensor<Tensor>(src_info); + auto weight = create_tensor<Tensor>(weight_info); + auto bias = create_tensor<Tensor>(bias_info); + auto dst = create_tensor<Tensor>(dst_info); + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &src }, { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias }, { TensorType::ACL_DST, &dst } }; + ITensorPack prep_pack{ { TensorType::ACL_SRC_1, &weight }, { TensorType::ACL_SRC_2, &bias } }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(fc->workspace(), mg, run_pack, prep_pack); + + // Initialize input values + const std::vector<int8_t> src_values = { 3, 63, 31 }; + const std::vector<int8_t> weight_values = { -4 }; + const std::vector<int32_t> bias_values = { 16 }; + const std::vector<int32_t> expected = { 80, 127, 127 }; + library->fill_static_values(Accessor(src), src_values); + library->fill_static_values(Accessor(weight), weight_values); + library->fill_static_values(Accessor(bias), bias_values); + + // Run FC layer + fc->prepare(prep_pack); + fc->run(run_pack); + + auto dst_ptr = reinterpret_cast<int8_t *>(dst.buffer()); + for(size_t i = 0; i < dst.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(dst_ptr[i] == expected[i], framework::LogLevel::ERRORS); + } +} + // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Mismatching data types + make("InputInfo", { TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Mismatching data types TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Invalid weights dimensions TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Wrongly reshaped weights TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), }), - framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(315U, 271U), 1, DataType::F16), + make("WeightsInfo",{ TensorInfo(TensorShape(315U, 271U), 1, DataType::F16), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), })), - framework::dataset::make("BiasInfo",{ TensorInfo(TensorShape(271U), 1, DataType::F32), + make("BiasInfo",{ TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(192U), 1, DataType::F32), TensorInfo(TensorShape(192U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(192U), 1, DataType::F32), })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), + make("OutputInfo",{ TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), })), - framework::dataset::make("TransposeWeights",{ true, true, false, true, true, true })), - framework::dataset::make("ReshapedWeights",{ false, false, false, false, false , false})), - framework::dataset::make("Expected", { false, true, true, false, false, true })), + make("TransposeWeights",{ true, true, false, true, true, true })), + make("ReshapedWeights",{ false, false, false, false, false , false})), + make("Expected", { false, true, true, false, false, true })), input_info, weights_info, bias_info, output_info, transpose_weights, reshaped_weights, expected) { // Create Fully Connected layer info @@ -145,74 +298,89 @@ template <typename T> using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T>; template <typename T> using NEFullyConnectedLayerMixedDataLayoutFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T, true>; +template <typename T> +using NEFullyConnectedLayerDynamicWeightsFixture = FullyConnectedWithDynamicWeightsFixture<Tensor, Accessor, NEFullyConnectedLayer, T>; +template <typename T> +using NEFullyConnectedLayerDynamicBiasFixture = FullyConnectedWithDynamicBiasFixture<Tensor, Accessor, NEFullyConnectedLayer, T>; TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F16)), - EmptyActivationFunctionDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + FullyConnectedParameters, + make("DataType", DataType::F16), + NoActivationFunctionDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16, abs_tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine( +FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::FullyConnectedLayerWithActivationDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F16)), + FullyConnectedParameters, + make("DataType", DataType::F16), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16, abs_tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F16)), - EmptyActivationFunctionDataset)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeFullyConnectedLayerDataset(), + FullyConnectedParameters, + make("DataType", DataType::F16), + NoActivationFunctionDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num_f16, abs_tolerance_f16); } +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, NEFullyConnectedLayerDynamicWeightsFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::F16), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)), + make("WeightsReshaped", { false, true }))) +{ +} TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), - EmptyActivationFunctionDataset)) +FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters, + make("DataType", DataType::F32), + NoActivationFunctionDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0, abs_tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEFullyConnectedLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(9U, 5U, 7U)), - framework::dataset::make("Weights", TensorShape(315U, 271U))), - framework::dataset::make("Biases", TensorShape(271U))), - framework::dataset::make("Output", TensorShape(271U))), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEFullyConnectedLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::F32), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0, abs_tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine( - combine(datasets::FullyConnectedLayerWithActivationDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), +FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::FullyConnectedLayerWithActivationDataset(), + FullyConnectedParameters, + make("DataType", DataType::F32), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0, abs_tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters), - framework::dataset::make("DataType", DataType::F32)), - EmptyActivationFunctionDataset)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeFullyConnectedLayerDataset(), FullyConnectedParameters, + make("DataType", DataType::F32), + NoActivationFunctionDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0, abs_tolerance_f32); } +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, NEFullyConnectedLayerDynamicWeightsFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::F32), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)), + make("WeightsReshaped", { false, true }))) +{ +} TEST_SUITE_END() TEST_SUITE_END() @@ -223,87 +391,152 @@ using NEFullyConnectedLayerQuantizedMixedDataLayoutFixture = FullyConnectedLayer TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine( - combine(datasets::SmallFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8)), - QuantizationData), - EmptyActivationFunctionDataset)) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, NEFullyConnectedLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8), + QuantizationData, + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEFullyConnectedLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(9U, 5U, 7U)), - framework::dataset::make("Weights", TensorShape(315U, 271U))), - framework::dataset::make("Biases", TensorShape(271U))), - framework::dataset::make("Output", TensorShape(271U))), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8)), - QuantizationData), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) -{ - // Validate output - validate(Accessor(_target), _reference, tolerance_qasymm8); -} -FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine( +FIXTURE_DATA_TEST_CASE(RunSmallWithActivation, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::FullyConnectedLayerWithActivationDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8)), - QuantizationData), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8), + QuantizationData, ActivationFunctionsQuantizedDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } +FIXTURE_DATA_TEST_CASE(RunDynamicWeightsWithActivation, NEFullyConnectedLayerDynamicWeightsFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)), + make("WeightsReshaped", { false }))) +{ +} +FIXTURE_DATA_TEST_CASE(RunDynamicBiasWithActivation, NEFullyConnectedLayerDynamicBiasFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +{ +} -FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine( - combine(datasets::LargeFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8)), - QuantizationData), - EmptyActivationFunctionDataset)) +// Dynamic Quantization Tests here +FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallFullyConnectedLayerDataset(), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationData, + NoActivationFunctionDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -TEST_SUITE_END() -TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine( - combine(datasets::SmallFullyConnectedLayerDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - QuantizationData), - EmptyActivationFunctionDataset)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine( + datasets::LargeFullyConnectedLayerDataset(), + FullyConnectedParameters, + framework::dataset::make("DataType", DataType::QASYMM8), + QuantizationData, + NoActivationFunctionDataset)) { // Validate output - validate(Accessor(_target), _reference, tolerance_qasymm8_signed); + validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEFullyConnectedLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(9U, 5U, 7U)), - framework::dataset::make("Weights", TensorShape(315U, 271U))), - framework::dataset::make("Biases", TensorShape(271U))), - framework::dataset::make("Output", TensorShape(271U))), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - QuantizationData), - framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +FIXTURE_DATA_TEST_CASE(RunDynamicBias, NEFullyConnectedLayerDynamicBiasFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8), + NoActivationFunctionDataset)) +{ +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEFullyConnectedLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8), + IgnoredQuantizationData, + NoActivationFunctionDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine( +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, NEFullyConnectedLayerDynamicWeightsFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8), + NoActivationFunctionDataset, + make("WeightsReshaped", { false }))) +{ +} +TEST_SUITE_END() // QASYMM8 +TEST_SUITE(QASYMM8_SIGNED) +FIXTURE_DATA_TEST_CASE(RunMixedDataLayoutWithActivation, NEFullyConnectedLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8_SIGNED), + QuantizationData, + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE(RunWithActivation, NEFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::FullyConnectedLayerWithActivationDataset(), - FullyConnectedParameters), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - QuantizationData), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8_SIGNED), + QuantizationData, ActivationFunctionsQuantizedDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } +FIXTURE_DATA_TEST_CASE(RunDynamicWeightsWithActivation, NEFullyConnectedLayerDynamicWeightsFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8_SIGNED), + make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU)), + make("WeightsReshaped", { false }))) +{ +} + +// Dynamic Quantization tests +FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine( + datasets::SmallFullyConnectedLayerDataset(), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8_SIGNED), + IgnoredQuantizationData, + NoActivationFunctionDataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEFullyConnectedLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, + combine( + make("Input", TensorShape(9U, 5U, 7U)), + make("Weights", TensorShape(315U, 271U)), + make("Biases", TensorShape(271U)), + make("Output", TensorShape(271U)), + FullyConnectedParameters, + make("DataType", DataType::QASYMM8_SIGNED), + QuantizationData, + NoActivationFunctionDataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} +FIXTURE_DATA_TEST_CASE(RunDynamicWeights, NEFullyConnectedLayerDynamicWeightsFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallFullyConnectedLayerDataset(), + make("DataType", DataType::QASYMM8_SIGNED), + NoActivationFunctionDataset, + make("WeightsReshaped", { false }))) +{ +} TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE_END() // Quantized TEST_SUITE_END() // FullyConnectedLayer diff --git a/tests/validation/NEON/GEMM.cpp b/tests/validation/NEON/GEMM.cpp index 500c6029d5..5f6a402204 100644 --- a/tests/validation/NEON/GEMM.cpp +++ b/tests/validation/NEON/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,12 +22,15 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NEGEMM.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" -#include "src/core/NEON/kernels/NEGEMMInterleave4x4Kernel.h" -#include "src/core/NEON/kernels/NEGEMMMatrixMultiplyKernel.h" -#include "src/core/NEON/kernels/NEGEMMTranspose1xWKernel.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/cpu/kernels/CpuGemmInterleave4x4Kernel.h" +#include "src/cpu/kernels/CpuGemmMatrixMultiplyKernel.h" +#include "src/cpu/kernels/CpuGemmTranspose1xWKernel.h" +#include "src/cpu/operators/CpuGemm.h" #include "tests/NEON/Accessor.h" #include "tests/NEON/Helper.h" #include "tests/PaddingCalculator.h" @@ -48,6 +51,8 @@ namespace test { namespace validation { +using framework::dataset::make; + namespace { constexpr AbsoluteTolerance<float> tolerance_f(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for FP32 data types */ @@ -57,7 +62,7 @@ const AbsoluteTolerance<float> abs_tolerance_f16(0.2f); /**< Absolute constexpr float tolerance_num = 0.07f; /**< Tolerance number for FP16 data types */ #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ /** CNN data types */ -const auto CNNDataTypes = framework::dataset::make("DataType", +const auto CNNDataTypes = make("DataType", { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC DataType::F16, @@ -65,62 +70,210 @@ const auto CNNDataTypes = framework::dataset::make("DataType", DataType::F32, }); -const auto data_interleave = framework::dataset::make("M", 8, 12) * framework::dataset::make("N", 8, 12); -const auto data_transpose = framework::dataset::make("M", 8, 14) * framework::dataset::make("N", 7, 14); +const auto data_interleave = make("M", 8, 12) * make("N", 8, 12); +const auto data_transpose = make("M", 8, 14) * make("N", 7, 14); /** Zero padding test */ template <typename FunctionType> bool validate_zero_padding(unsigned int dim0_value, unsigned int dim1_value) { const TensorShape in_shape(dim0_value, dim1_value); + TensorInfo in(in_shape, 1, DataType::U32); + TensorInfo dst; - // Create tensors - Tensor in = create_tensor<Tensor>(in_shape, DataType::U32); - Tensor dst; - - ARM_COMPUTE_EXPECT(in.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(in.is_resizable(), framework::LogLevel::ERRORS); // Validate zero-padding FunctionType func; func.configure(&in, &dst); - return in.info()->padding().empty(); + return in.padding().empty(); } /* Zero padding test for GEMM kernels */ bool validate_gemm_zero_padding(const TensorShape shape0, const TensorShape shape1) { // Create tensors - Tensor in0 = create_tensor<Tensor>(shape0, DataType::F32); - Tensor in1 = create_tensor<Tensor>(shape1, DataType::F32); - Tensor dst; + TensorInfo in0(shape0, 1, DataType::F32); + TensorInfo in1(shape1, 1, DataType::F32); + TensorInfo dst; // Validate zero-padding - NEGEMMMatrixMultiplyKernel gemm; + cpu::kernels::CpuGemmMatrixMultiplyKernel gemm; gemm.configure(&in0, &in1, &dst, 1.0, false); - return in0.info()->padding().empty() && in1.info()->padding().empty() && dst.info()->padding().empty(); + return in0.padding().empty() && in1.padding().empty() && dst.padding().empty(); } } // namespace TEST_SUITE(NEON) TEST_SUITE(GEMM) +/** Test case for memory injection in @ref cpu::CpuGemm. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) +{ + auto gemm = std::make_unique<cpu::CpuGemm>(); + const auto lhs_info = TensorInfo(TensorShape(3U, 3U), 1, DataType::F32); + const auto rhs_info = TensorInfo(TensorShape(4U, 3U), 1, DataType::F32); + const auto c_info = TensorInfo(TensorShape(4U, 3U), 1, DataType::F32); + auto dst_info = TensorInfo(TensorShape(4U, 3U), 1, DataType::F32); + const auto gemm_info = GEMMInfo{}; + gemm->configure(&lhs_info, &rhs_info, &c_info, &dst_info, 1.f, 1.f, gemm_info); + + // telhs are newly created every call of this lambda function + auto lhs = create_tensor<Tensor>(lhs_info); + auto rhs = create_tensor<Tensor>(rhs_info); + auto c = create_tensor<Tensor>(c_info); + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + c.allocator()->allocate(); + + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &lhs }, { TensorType::ACL_SRC_1, &rhs }, { TensorType::ACL_SRC_2, &c } }; + ITensorPack prep_pack{ { TensorType::ACL_SRC_1, &rhs }, { TensorType::ACL_SRC_2, &c } }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(gemm->workspace(), mg, run_pack, prep_pack); + + auto run_conv = [&]() -> Tensor + { + auto dst = create_tensor<Tensor>(dst_info); + dst.allocator()->allocate(); + run_pack.add_tensor(TensorType::ACL_DST, &dst); + + library->fill_tensor_value(Accessor(lhs), 1.f); + library->fill_tensor_value(Accessor(rhs), 2.f); + library->fill_tensor_value(Accessor(c), 3.f); + // This operator is configured once and captured by this lambda. + gemm->prepare(prep_pack); + gemm->run(run_pack); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +/** Test case for memory injection in @ref NEGEMM. + * + * Make sure @ref NEGEMM still works through injecting the memory at configure time using the old API. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) +{ + auto gemm = std::make_unique<NEGEMM>(); + const auto lhs_info = TensorInfo(TensorShape(3U, 3U), 1, DataType::F32); + const auto rhs_info = TensorInfo(TensorShape(4U, 3U), 1, DataType::F32); + const auto c_info = TensorInfo(TensorShape(4U, 3U), 1, DataType::F32); + auto dst_info = TensorInfo(TensorShape(4U, 3U), 1, DataType::F32); + const auto gemm_info = GEMMInfo{}; + auto run_conv = [&]() + { + auto lhs = create_tensor<Tensor>(lhs_info); + auto rhs = create_tensor<Tensor>(rhs_info); + auto c = create_tensor<Tensor>(c_info); + auto dst = create_tensor<Tensor>(dst_info); + gemm->configure(&lhs, &rhs, &c, &dst, 1.f, 1.f, gemm_info); + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + c.allocator()->allocate(); + dst.allocator()->allocate(); + library->fill_tensor_value(Accessor(lhs), 1.f); + library->fill_tensor_value(Accessor(rhs), 2.f); + library->fill_tensor_value(Accessor(c), 3.f); + gemm->run(); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + make("LhsInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::S32), // Unsupported data type + TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), + }), + make("RhsInfo",{ TensorInfo(TensorShape(8U, 27U), 1, DataType::S32), + TensorInfo(TensorShape(8U, 27U), 1, DataType::F32), + })), + make("OutputInfo",{ TensorInfo(TensorShape(8U, 13U), 1, DataType::S32), + TensorInfo(TensorShape(8U, 13U), 1, DataType::F32), + })), + make("Expected", { false, true })), + lhs_info, rhs_info, output_info, expected) +{ + constexpr float alpha = 1.0; + constexpr float beta = 0.0; + const auto gemm_info = GEMMInfo(); + bool is_valid = bool(NEGEMM::validate(&lhs_info.clone()->set_is_resizable(true), &rhs_info.clone()->set_is_resizable(true), nullptr, &output_info.clone()->set_is_resizable(true), alpha, beta, gemm_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* +TEST_SUITE(KERNEL_SELECTION) +DATA_TEST_CASE(KernelSelection_mul_and_add, framework::DatasetMode::ALL, + combine(make("CpuExt", std::string("NEON")), + make("DataType", { DataType::F32, + DataType::F16 + })), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl_mul = CpuGemmMatrixMultiplyKernel::get_implementation(DataTypeISASelectorData{ data_type, cpu_isa }, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl_mul); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_gemm_matrix_mul"; + std::string actual = selected_impl_mul->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); + + const auto *selected_impl_add = CpuGemmMatrixAdditionKernel::get_implementation(DataTypeISASelectorData{ data_type, cpu_isa }, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl_add); + + expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_gemm_matrix_add"; + actual = selected_impl_add->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() // KERNEL_SELECTION + TEST_SUITE(TRANSPOSE_1XW) -using NEGEMMTranspose1xW = NESynthetizeFunctionWithZeroConstantBorder<NEGEMMTranspose1xWKernel, 4>; +using CpuGemmTranspose1xW = NESynthetizeFunctionWithZeroConstantKernelBorder<cpu::kernels::CpuGemmTranspose1xWKernel>; DATA_TEST_CASE(ValidateZeroPadding, framework::DatasetMode::ALL, zip( - framework::dataset::make("N", { 1, 23, 63, 101 }), - framework::dataset::make("K", { 1, 47, 29, 27 })), + make("N", { 1, 23, 63, 101 }), + make("K", { 1, 47, 29, 27 })), n_value, k_value) { - bool status = validate_zero_padding<NEGEMMTranspose1xWKernel>(n_value, k_value); + bool status = validate_zero_padding<CpuGemmTranspose1xW>(n_value, k_value); ARM_COMPUTE_EXPECT(status, framework::LogLevel::ERRORS); } TEST_SUITE(U32) -using NEGEMMTranspose1xWFixture = GEMMTranspose1xWValidationFixture<Tensor, Accessor, NEGEMMTranspose1xW, uint32_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * framework::dataset::make("DataType", DataType::U32)) +using CpuGemmTranspose1xWFixture = GEMMTranspose1xWValidationFixture<Tensor, Accessor, CpuGemmTranspose1xW, uint32_t>; +FIXTURE_DATA_TEST_CASE(RunSmall, CpuGemmTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * make("DataType", DataType::U32)) { // Validate output validate(Accessor(_target), _reference); @@ -128,8 +281,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMTranspose1xWFixture, framework::DatasetMo TEST_SUITE_END() // U32 TEST_SUITE(U16) -using NEGEMMTranspose1xWFixture = GEMMTranspose1xWValidationFixture<Tensor, Accessor, NEGEMMTranspose1xW, uint16_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * framework::dataset::make("DataType", DataType::U16)) +using CpuGemmTranspose1xWFixture = GEMMTranspose1xWValidationFixture<Tensor, Accessor, CpuGemmTranspose1xW, uint16_t>; +FIXTURE_DATA_TEST_CASE(RunSmall, CpuGemmTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * make("DataType", DataType::U16)) { // Validate output validate(Accessor(_target), _reference); @@ -137,8 +290,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMTranspose1xWFixture, framework::DatasetMo TEST_SUITE_END() // U16 TEST_SUITE(U8) -using NEGEMMTranspose1xWFixture = GEMMTranspose1xWValidationFixture<Tensor, Accessor, NEGEMMTranspose1xW, uint8_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * framework::dataset::make("DataType", DataType::U8)) +using CpuGemmTranspose1xWFixture = GEMMTranspose1xWValidationFixture<Tensor, Accessor, CpuGemmTranspose1xW, uint8_t>; +FIXTURE_DATA_TEST_CASE(RunSmall, CpuGemmTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * make("DataType", DataType::U8)) { // Validate output validate(Accessor(_target), _reference); @@ -148,20 +301,20 @@ TEST_SUITE_END() // U8 TEST_SUITE_END() // TRANSPOSE_1XW TEST_SUITE(INTERLEAVE_4X4) -using NEGEMMInterleave4x4 = NESynthetizeFunctionWithZeroConstantBorder<NEGEMMInterleave4x4Kernel, 4>; +using CpuGemmInterleave4x4 = NESynthetizeFunctionWithZeroConstantKernelBorder<cpu::kernels::CpuGemmInterleave4x4Kernel>; DATA_TEST_CASE(ValidateZeroPadding, framework::DatasetMode::ALL, zip( - framework::dataset::make("M", { 1, 23, 63, 101 }), - framework::dataset::make("K", { 1, 47, 29, 27 })), + make("M", { 1, 23, 63, 101 }), + make("K", { 1, 47, 29, 27 })), m_value, k_value) { - bool status = validate_zero_padding<NEGEMMInterleave4x4Kernel>(m_value, k_value); + bool status = validate_zero_padding<cpu::kernels::CpuGemmInterleave4x4Kernel>(m_value, k_value); ARM_COMPUTE_EXPECT(status, framework::LogLevel::ERRORS); } TEST_SUITE(U32) -using NEGEMMInterleave4x4Fixture = GEMMInterleave4x4ValidationFixture<Tensor, Accessor, NEGEMMInterleave4x4, uint32_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * framework::dataset::make("DataType", DataType::U32)) +using CpuGemmInterleave4x4Fixture = GEMMInterleave4x4ValidationFixture<Tensor, Accessor, CpuGemmInterleave4x4, uint32_t>; +FIXTURE_DATA_TEST_CASE(RunSmall, CpuGemmInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * make("DataType", DataType::U32)) { // Validate output validate(Accessor(_target), _reference); @@ -169,8 +322,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMInterleave4x4Fixture, framework::DatasetM TEST_SUITE_END() // U32 TEST_SUITE(U16) -using NEGEMMInterleave4x4Fixture = GEMMInterleave4x4ValidationFixture<Tensor, Accessor, NEGEMMInterleave4x4, uint16_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * framework::dataset::make("DataType", DataType::U16)) +using CpuGemmInterleave4x4Fixture = GEMMInterleave4x4ValidationFixture<Tensor, Accessor, CpuGemmInterleave4x4, uint16_t>; +FIXTURE_DATA_TEST_CASE(RunSmall, CpuGemmInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * make("DataType", DataType::U16)) { // Validate output validate(Accessor(_target), _reference); @@ -178,8 +331,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMInterleave4x4Fixture, framework::DatasetM TEST_SUITE_END() // U16 TEST_SUITE(U8) -using NEGEMMInterleave4x4Fixture = GEMMInterleave4x4ValidationFixture<Tensor, Accessor, NEGEMMInterleave4x4, uint8_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * framework::dataset::make("DataType", DataType::QASYMM8)) +using CpuGemmInterleave4x4Fixture = GEMMInterleave4x4ValidationFixture<Tensor, Accessor, CpuGemmInterleave4x4, uint8_t>; +FIXTURE_DATA_TEST_CASE(RunSmall, CpuGemmInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * make("DataType", DataType::QASYMM8)) { // Validate output validate(Accessor(_target), _reference); @@ -192,17 +345,20 @@ template <typename T> using NEGEMMFixture = GEMMValidationFixture<Tensor, Accessor, NEGEMM, T>; template <typename T> -using NEGEMMFixtureDisabledC = GEMMValidationFixture<Tensor, Accessor, NEGEMM, T, true>; +using NEBatchedMatMulFixture = GEMMValidationFixture<Tensor, Accessor, NEGEMM, T, true, false, false, false, false, true>; + +template <typename T> +using NEGEMMAccumulateFixture = GEMMAccumulateValidationFixture<Tensor, Accessor, NEGEMM, T>; TEST_SUITE(Float) -DATA_TEST_CASE(ValidateZeroPadding, framework::DatasetMode::ALL, zip(framework::dataset::make("In0", { TensorShape(21U, 13U), +DATA_TEST_CASE(ValidateZeroPadding, framework::DatasetMode::ALL, zip(make("In0", { TensorShape(21U, 13U), TensorShape(31U, 1U), TensorShape(31U, 1U), TensorShape(8U, 2U), TensorShape(38U, 12U), TensorShape(32U, 1U) }), - framework::dataset::make("In1", { TensorShape(33U, 21U), + make("In1", { TensorShape(33U, 21U), TensorShape(23U, 31U), TensorShape(23U, 31U), TensorShape(16U, 8U), @@ -215,59 +371,111 @@ DATA_TEST_CASE(ValidateZeroPadding, framework::DatasetMode::ALL, zip(framework:: ARM_COMPUTE_EXPECT(status, framework::LogLevel::ERRORS); } +DATA_TEST_CASE(ValidateAccumulate, framework::DatasetMode::ALL, combine( + zip(make("In0",{ TensorShape(21U, 13U) }), + make("In1", { TensorShape(33U, 21U) }), + make("Dst", { TensorShape(33U, 13U) })), + zip( + make("alpha", { 1.0, 100.0, 1.0, 1.0 }), + make("beta", { 0.0, 0.0, 1.0, 1.0 }), + make("is_c_null", { false, false, false, true }), + make("Expected", { true, false, false, true }))), + shape_a, shape_b, shape_dst, alpha, beta, is_c_null, expected) +{ + /* Accumulation test for GEMM kernels */ + // Create tensors + TensorInfo in_a(shape_a, 1, DataType::F32); + TensorInfo in_b(shape_b, 1, DataType::F32); + TensorInfo in_c(shape_dst, 1, DataType::F32); + TensorInfo dst(shape_dst, 1, DataType::F32); + + GEMMInfo gemm_info = GEMMInfo(); + gemm_info.set_accumulate(true); + + // Validate accumulation + cpu::CpuGemm gemm; + Status status = gemm.validate(&in_a, &in_b, (is_c_null ? nullptr : &in_c), &dst, alpha, beta, gemm_info); + ARM_COMPUTE_EXPECT((expected == bool(status)), framework::LogLevel::ERRORS); +} + #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F16))) + make("ReshapeWeights", { true, false })), + make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_f16); } FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + make("ReshapeWeights", { true, false })), + make("DataType", DataType::F16))) +{ + // Validate output + validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_f16); +} - framework::dataset::make("DataType", DataType::F16))) +TEST_SUITE(BATCHED_MATMUL) +FIXTURE_DATA_TEST_CASE(RunSmall, NEBatchedMatMulFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallBatchedMatMulDataset(), + make("ReshapeWeights", { false })), + make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_f16); } -TEST_SUITE_END() +TEST_SUITE_END() // BATCHED_MATMUL + +TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - - framework::dataset::make("DataType", DataType::F32))) + make("ReshapeWeights", { true, false })), + make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f); } FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), + make("ReshapeWeights", { true, false })), + make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f); +} - framework::dataset::make("DataType", DataType::F32))) +TEST_SUITE(BATCHED_MATMUL) +FIXTURE_DATA_TEST_CASE(RunSmall, NEBatchedMatMulFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallBatchedMatMulDataset(), + make("ReshapeWeights", { false })), + make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f); } -TEST_SUITE(DisabledC) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixtureDisabledC<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), +TEST_SUITE_END() // BATCHED_MATMUL - framework::dataset::make("DataType", DataType::F32))) +TEST_SUITE(ACCUMULATE) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMAccumulateFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallAccumulateGEMMDataset(), + make("ReshapeWeights", { false }), + make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMAccumulateFixture<float>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeAccumulateGEMMDataset(), + make("ReshapeWeights", { false }), + make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f); } -TEST_SUITE_END() +TEST_SUITE_END() // ACCUMULATE -TEST_SUITE_END() -TEST_SUITE_END() +TEST_SUITE_END() // FP32 -TEST_SUITE_END() -TEST_SUITE_END() +TEST_SUITE_END() // Float +TEST_SUITE_END() // GEMM +TEST_SUITE_END() // NEON } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp index 9d075e12c1..d25f43a330 100644 --- a/tests/validation/NEON/GEMMLowp.cpp +++ b/tests/validation/NEON/GEMMLowp.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -26,6 +26,8 @@ #include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/cpu/operators/CpuGemmLowpMatrixMultiplyCore.h" #include "tests/NEON/Accessor.h" #include "tests/NEON/Helper.h" #include "tests/PaddingCalculator.h" @@ -45,10 +47,26 @@ namespace test { namespace validation { +using framework::dataset::make; + +namespace +{ + constexpr AbsoluteTolerance<float> tolerance_batched(1); + constexpr AbsoluteTolerance<float> tolerance_quant(1); +} // namespace + + TEST_SUITE(NEON) TEST_SUITE(GEMMLowp) TEST_SUITE(MatrixMultiplyCore) + using NEGEMMLowpMatrixMultiplyCoreFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>; +using NEGEMMLowpMatrixMultiplyCoreAccumulateFixture = GEMMLowpMatrixMultiplyAccumulateValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>; +using NEGEMMLowpBatchedMatMulFixture = GEMMLowpMatrixMultiplyCoreValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore, false, false, true>; +using NEGEMMLowpMatrixMultiplyCoreDynamicQuantizationFixture = GEMMLowpMatrixMultiplyCoreDynamicQuantizationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>; +using NEGEMMLowpDequantizedMatrixMultiplyValidationFixture = GEMMLowpDequantizedMatrixMultiplyValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>; + +using framework::dataset::make; DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::concat(datasets::SmallGEMMLowpDataset(), datasets::LargeGEMMLowpDataset()), shape_a, shape_b, shape_c, a_offset, b_offset) @@ -74,29 +92,69 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, framework::dataset::c validate(b.info()->padding(), PaddingSize()); validate(c.info()->padding(), PaddingSize()); } +// accumulation is not supported for Int8/UInt8 in aarch32 +#ifdef __aarch64__ +DATA_TEST_CASE(ValidateAccumulate, framework::DatasetMode::ALL, combine( + zip( + make("In0",{ TensorShape(21U, 1U) }), + make("In1", { TensorShape(1U, 21U) }), + make("Dst", { TensorShape(1U, 1U) }), + make("a_offset", { -2 }), + make("a_offset", { 13 }) + ), + zip( + make("OutputDataType", { DataType::S32, DataType::QASYMM8, DataType::QASYMM8_SIGNED}), + make("Expected", { true, false, false }) + )), + shape_a, shape_b, shape_dst, a_offset, b_offset, output_data_type, expected) +{ + DataType input_data_type = (output_data_type == DataType::S32 ? DataType::QASYMM8 : output_data_type); + // Accumulation test for GEMM kernels + TensorInfo a(shape_a, 1, input_data_type, QuantizationInfo(1.0f / 255, a_offset)); + TensorInfo b(shape_b, 1, input_data_type, QuantizationInfo(1.0f / 255, b_offset)); + TensorInfo dst(shape_dst, 1, output_data_type, QuantizationInfo()); + + // Create and configure function + GEMMInfo gemm_info = GEMMInfo(); + gemm_info.set_accumulate(true); + + if (is_data_type_quantized(output_data_type)) + { + GEMMLowpOutputStageInfo gemmLowpOutputStageInfo = GEMMLowpOutputStageInfo(); + gemmLowpOutputStageInfo.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + + gemm_info.set_gemmlowp_output_stage(gemmLowpOutputStageInfo); + } + + cpu::CpuGemmLowpMatrixMultiplyCore gemmlowp_mm; + Status status = gemmlowp_mm.validate(&a, &b, nullptr, &dst, gemm_info); + + ARM_COMPUTE_EXPECT((expected == bool(status)), framework::LogLevel::ERRORS); +} +#endif // __arch64__ // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Input not a multiple of 4 +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Input not a multiple of 4 TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Mismatching data type TensorInfo(TensorShape(20U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Invalid dimensions TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), // Invalid dimensions TensorInfo(TensorShape(16U, 32U), 1, DataType::QASYMM8, QuantizationInfo(1.f/255, 10)), }), - framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(33U, 21U), 1, DataType::QASYMM8, QuantizationInfo(1.f/256, 10)), + make("InputBInfo",{ TensorInfo(TensorShape(33U, 21U), 1, DataType::QASYMM8, QuantizationInfo(1.f/256, 10)), TensorInfo(TensorShape(33U, 21U), 1, DataType::QASYMM8, QuantizationInfo(1.f/256, 10)), TensorInfo(TensorShape(33U, 21U), 1, DataType::QASYMM8, QuantizationInfo(1.f/256, 10)), TensorInfo(TensorShape(33U, 21U), 1, DataType::QASYMM8, QuantizationInfo(1.f/256, 10)), TensorInfo(TensorShape(64U, 16U), 1, DataType::QASYMM8, QuantizationInfo(1.f/256, 10)), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(33U, 13U), 1, DataType::S32), + }), + make("OutputInfo",{ TensorInfo(TensorShape(33U, 13U), 1, DataType::S32), TensorInfo(TensorShape(33U, 13U), 1, DataType::S32), TensorInfo(TensorShape(33U, 13U), 1, DataType::S32), TensorInfo(TensorShape(8U, 11U), 1, DataType::S32), TensorInfo(TensorShape(64U, 32U), 1, DataType::S32), - })), - framework::dataset::make("Expected", { true, false, false, false, true })), + }), + make("Expected", { true, false, false, false, true })), a_info, b_info, output_info, expected) { // Lock tensors @@ -109,444 +167,224 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( // clang-format on // *INDENT-ON* -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset()) +/** Test case for memory injection in @ref cpu::CpuGemmLowpMatrixMultiplyCore. + * + * Configure the operator once and inject memory at run-time in multiple executions. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) { - // Validate output - validate(Accessor(_target), _reference); + auto gemm = std::make_unique<cpu::CpuGemmLowpMatrixMultiplyCore>(); + auto a_info = TensorInfo(TensorShape(32U, 72U), 1, DataType::QASYMM8); + auto b_info = TensorInfo(TensorShape(17U, 32U), 1, DataType::QASYMM8); + auto dst_info = TensorInfo(TensorShape(17U, 72U), 1, DataType::S32); + a_info.set_quantization_info(QuantizationInfo(1.0f / 255, -9)); + b_info.set_quantization_info(QuantizationInfo(1.0f / 255, 1)); + const auto gemm_info = GEMMInfo{}; + gemm->configure(&a_info, &b_info, nullptr, &dst_info, gemm_info); + + // telhs are newly created every call of this lambda function + auto a = create_tensor<Tensor>(a_info); + auto b = create_tensor<Tensor>(b_info); + auto dst = create_tensor<Tensor>(dst_info); + a.allocator()->allocate(); + b.allocator()->allocate(); + dst.allocator()->allocate(); + + ITensorPack run_pack = + { + { TensorType::ACL_SRC_0, &a }, + { TensorType::ACL_SRC_1, &b }, + { TensorType::ACL_DST, &dst } + }; + ITensorPack prep_pack = + { + { TensorType::ACL_SRC_1, &b }, + }; + + auto mg = MemoryGroup{}; + auto ws = manage_workspace<Tensor>(gemm->workspace(), mg, run_pack, prep_pack); + + auto run_conv = [&]() -> Tensor + { + auto dst = create_tensor<Tensor>(dst_info); + dst.allocator()->allocate(); + run_pack.add_tensor(TensorType::ACL_DST, &dst); + + library->fill_tensor_value(Accessor(a), static_cast<uint8_t>(1)); + library->fill_tensor_value(Accessor(b), static_cast<uint8_t>(2)); + // This operator is configured once and captured by this lambda. + gemm->prepare(prep_pack); + gemm->run(run_pack); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((uint8_t *)result_0.buffer())[i] == ((uint8_t *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpDataset()) +/** Test case for memory injection in @ref NEGEMMLowpMatrixMultiplyCore. + * + * Make sure @ref NEGEMMLowpMatrixMultiplyCore still works through injecting the memory at configure time using the old API. + * + * Checks performed in order: + * - Both runs compute the same output + */ +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) { - // Validate output - validate(Accessor(_target), _reference); + auto gemm = std::make_unique<NEGEMMLowpMatrixMultiplyCore>(); + auto a_info = TensorInfo(TensorShape(32U, 72U), 1, DataType::QASYMM8); + auto b_info = TensorInfo(TensorShape(17U, 32U), 1, DataType::QASYMM8); + auto dst_info = TensorInfo(TensorShape(17U, 72U), 1, DataType::S32); + a_info.set_quantization_info(QuantizationInfo(1.0f / 255, -9)); + b_info.set_quantization_info(QuantizationInfo(1.0f / 255, 1)); + const auto gemm_info = GEMMInfo{}; + auto run_conv = [&]() + { + auto a = create_tensor<Tensor>(a_info); + auto b = create_tensor<Tensor>(b_info); + auto dst = create_tensor<Tensor>(dst_info); + gemm->configure(&a, &b, nullptr, &dst, gemm_info); + a.allocator()->allocate(); + b.allocator()->allocate(); + dst.allocator()->allocate(); + library->fill_tensor_value(Accessor(a), static_cast<uint8_t>(1)); + library->fill_tensor_value(Accessor(b), static_cast<uint8_t>(2)); + gemm->run(); + return dst; + }; + auto result_0 = run_conv(); + auto result_1 = run_conv(); + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); ++i) + { + ARM_COMPUTE_EXPECT(((uint8_t *)result_0.buffer())[i] == ((uint8_t *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } } -using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>; -TEST_SUITE(FusedOffsetOutput) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(), - framework::dataset::make("DataType", { DataType::QASYMM8 }))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset()) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeGEMMLowpFusedOffsetOutputUint8Dataset(), - framework::dataset::make("DataType", { DataType::QASYMM8 }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpDataset()) { // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() // FusedOffsetOutput -TEST_SUITE_END() // MatrixMultiplyCore - -TEST_SUITE(OutputStage) - -TEST_SUITE(QuantizeDownInt32Scale) +TEST_SUITE(BatchedMatMul) TEST_SUITE(QASYMM8) - -const auto quantize_down_int32_to_uint8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2, - 3) - * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_uint8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, - 2) - * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true }); - -using NEGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpOutputStage>; - -// *INDENT-OFF* -// clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16 - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type - }), - framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32), - TensorInfo(TensorShape(20U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8), - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), - })), - framework::dataset::make("Min",{ 0, - 13, - })), - framework::dataset::make("Max",{ 205, - 180, - })), - framework::dataset::make("Expected", { true, false })), - a_info, b_info, output_info, min, max, expected) -{ - - GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo(); - output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN; - output_stage.gemmlowp_min_bound = min; - output_stage.gemmlowp_max_bound = max; - output_stage.output_data_type = DataType::QASYMM8; - - // Lock tensors - Status status = NEGEMMLowpOutputStage::validate(&a_info.clone()->set_is_resizable(false), - &b_info.clone()->set_is_resizable(false), - &output_info.clone()->set_is_resizable(false), - output_stage); - ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); -} -// clang-format on -// *INDENT-ON* - -TEST_CASE(NoPaddingAdded, framework::DatasetMode::PRECOMMIT) +using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedUnsigned = + GEMMLowpBatchedMatrixMultiplyCoreFusedOffsetOutputFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore, false, false, uint8_t, uint8_t, true>; +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedUnsigned, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedBatchedMatMulDataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { false }))) { - Tensor input1 = create_tensor<Tensor>(TensorShape(21U, 13U), DataType::S32); - Tensor input2 = create_tensor<Tensor>(TensorShape(21U, 1U), DataType::S32); - Tensor output = create_tensor<Tensor>(TensorShape(21U, 13U), DataType::QASYMM8); - - GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo(); - output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN; - output_stage.gemmlowp_min_bound = 0; - output_stage.gemmlowp_max_bound = 205; - output_stage.output_data_type = DataType::QASYMM8; - - NEGEMMLowpOutputStage f; - f.configure(&input1, &input2, &output, output_stage); - - // Validate padding is zero - validate(input1.info()->padding(), PaddingSize()); - validate(input2.info()->padding(), PaddingSize()); - validate(output.info()->padding(), PaddingSize()); + validate(Accessor(_target), _reference, tolerance_batched); } - -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_cases)) -{ - // Validate output - validate(Accessor(_target), _reference); -} - -TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_uint8_scale_relu_cases)) -{ - // Validate output - validate(Accessor(_target), _reference); -} - -TEST_SUITE_END() // BoundedReLu - TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) - -const auto quantize_down_int32_to_int8_scale_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, 2) * framework::dataset::make("result_shift", 2, - 3) - * framework::dataset::make("min", 0) * framework::dataset::make("max", 0) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int8_scale_relu_cases = framework::dataset::make("result_offset", -2, 1) * framework::dataset::make("result_mult_int", 1, - 2) - * framework::dataset::make("result_shift", 2, 3) * framework::dataset::make("min", -100, -98) * framework::dataset::make("max", 71, 74) * framework::dataset::make("addBias", { false, true }); - -using NEGEMMLowpQuantizeDownInt32ScaleFixture = GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture<Tensor, Accessor, NEGEMMLowpOutputStage>; - -// *INDENT-OFF* -// clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16 - TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Invalid min and max - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type - }), - framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32), - TensorInfo(TensorShape(21U), 1, DataType::S32), - TensorInfo(TensorShape(20U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED), - TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED), - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), - })), - framework::dataset::make("Min",{ -10, - -200, - -113, - })), - framework::dataset::make("Max",{ 105, - 300, - -18, - })), - framework::dataset::make("Expected", { true, false, false })), - a_info, b_info, output_info, min, max, expected) +using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedSigned = + GEMMLowpBatchedMatrixMultiplyCoreFusedOffsetOutputFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore, false, false, int8_t, int8_t, true>; +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixtureBatchedSigned, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedBatchedMatMulDataset(), + make("DataType", { DataType::QASYMM8_SIGNED }), + make("reshape_b_only_on_first_run", { false }))) { - GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo(); - output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN; - output_stage.gemmlowp_min_bound = min; - output_stage.gemmlowp_max_bound = max; - output_stage.output_data_type = DataType::QASYMM8_SIGNED; - - // Lock tensors - Status status = NEGEMMLowpOutputStage::validate(&a_info.clone()->set_is_resizable(false), - &b_info.clone()->set_is_resizable(false), - &output_info.clone()->set_is_resizable(false), - output_stage); - ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + validate(Accessor(_target), _reference, tolerance_batched); } -// clang-format on -// *INDENT-ON* +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // BatchedMatMul -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_cases)) +TEST_SUITE(FusedOffsetOutput) +using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture<Tensor, Accessor, NEGEMMLowpMatrixMultiplyCore>; +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, + combine(datasets::SmallGEMMLowpFusedOffsetOutputUint8Dataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { false }))) { // Validate output - validate(Accessor(_target), _reference); + validate(Accessor(_target), _reference, tolerance_quant); } - -TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ScaleFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), quantize_down_int32_to_int8_scale_relu_cases)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::NIGHTLY, + combine(datasets::LargeGEMMLowpFusedOffsetOutputUint8Dataset(), + make("DataType", { DataType::QASYMM8 }), + make("reshape_b_only_on_first_run", { false }))) { // Validate output - validate(Accessor(_target), _reference); + validate(Accessor(_target), _reference, tolerance_quant); } +TEST_SUITE_END() // FusedOffsetOutput -TEST_SUITE_END() // BoundedReLu - -TEST_SUITE_END() // QASYMM8_SIGNED - -TEST_SUITE_END() // QuantizeDownInt32Scale - -TEST_SUITE(QuantizeDownInt32ToUint8ScaleByFixedPoint) - -const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0) * framework::dataset::make("max", 255) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", 0, 2) * framework::dataset::make("max", 171, 174) * framework::dataset::make("addBias", { false, true }); - -using NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint>; - -using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>; - -// *INDENT-OFF* -// clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16 - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type - }), - framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32), - TensorInfo(TensorShape(20U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8), - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), - })), - framework::dataset::make("Min",{ 0, - 13, - })), - framework::dataset::make("Max",{ 205, - 180, - })), - framework::dataset::make("Expected", { true, false })), - a_info, b_info, output_info, min, max, expected) -{ - // Lock tensors - Status status = NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false), - &b_info.clone()->set_is_resizable(false), - &output_info.clone()->set_is_resizable(false), - min, - max); - ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); -} -// clang-format on -// *INDENT-ON* - -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_cases)) +// accumulation is not supported for Int8/UInt8 in aarch32 +#ifdef __aarch64__ +TEST_SUITE(ACCUMULATION) +TEST_SUITE(S32) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreAccumulateFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset()) { // Validate output validate(Accessor(_target), _reference); } - -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_cases)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreAccumulateFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpDataset()) { // Validate output validate(Accessor(_target), _reference); } +TEST_SUITE_END() // S32 +TEST_SUITE_END() // ACCUMULATION +#endif // __arch64__ -TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases)) +TEST_SUITE(DynamicQuantization) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreDynamicQuantizationFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpDataset()) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeShapes(), - quantize_down_int32_to_uint8_scale_by_fixedpoint_relu_cases)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreDynamicQuantizationFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpDataset()) { // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() // BoundedReLu - -TEST_SUITE_END() // QuantizeDownInt32ToUint8ScaleByFixedPoint - -TEST_SUITE(QuantizeDownInt32ToInt8ScaleByFixedPoint) - -const auto quantize_down_int32_to_int8_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -128) * framework::dataset::make("max", 128) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("result_offset_after_shift", 2, 3) * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true }); - -using NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint>; - -// *INDENT-OFF* -// clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::F32), // Invalid input data type - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type - TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), - }), - framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32), - TensorInfo(TensorShape(20U), 1, DataType::S32), - TensorInfo(TensorShape(21U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED), - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), - TensorInfo(TensorShape(21U, 13U), 1, DataType::QASYMM8_SIGNED), - })), - framework::dataset::make("Min",{ -110, - -113, - -113, - })), - framework::dataset::make("Max",{ 87, - 97, - 97, - })), - framework::dataset::make("Expected", { false, false, true })), - a_info, b_info, output_info, min, max, expected) -{ - // Lock tensors - Status status = NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false), - &b_info.clone()->set_is_resizable(false), - &output_info.clone()->set_is_resizable(false), - min, - max); - ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); -} -// clang-format on -// *INDENT-ON* - -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int8_scale_by_fixedpoint_cases)) +TEST_SUITE_END() // DynamicQuantization + +#ifdef __aarch64__ +// Deqaunt tests involve returning F32 from the MatrixMultiplyCore kernels and is only implemented in aarch64 +TEST_SUITE(Dequant) +constexpr AbsoluteTolerance<float> tolerance_dequantized(0.01f); +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpDequantizedMatrixMultiplyValidationFixture, framework::DatasetMode::ALL, + combine( + datasets::SmallGEMMLowpDataset(), + make("accumulate", {true, false}) + )) { // Validate output - validate(Accessor(_target), _reference); + validate(Accessor(_target), _reference, tolerance_dequantized); } -TEST_SUITE(BoundedReLu) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int8_scale_by_fixedpoint_relu_cases)) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpDequantizedMatrixMultiplyValidationFixture, framework::DatasetMode::NIGHTLY, + combine( + datasets::LargeGEMMLowpDataset(), + make("accumulate", {false}) + )) { // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // BoundedReLu -TEST_SUITE_END() // QuantizeDownInt32ToInt8ScaleByFixedPoint - -TEST_SUITE(QuantizeDownInt32ToInt16ScaleByFixedPoint) - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("min", -32768) * framework::dataset::make("max", 32767) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, 254601602) * framework::dataset::make("result_shift", 1, - 2) - * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true }); -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_cases = framework::dataset::make("result_fixedpoint_multiplier", 1073741823, - 1073741825) - * framework::dataset::make("result_shift", -3, - -2) - * framework::dataset::make("min", -32768) * framework::dataset::make("max", 32767) * framework::dataset::make("addBias", { false, true }); - -const auto quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_relu_cases = framework::dataset::make("result_fixedpoint_multiplier", 254601600, - 254601602) - * framework::dataset::make("result_shift", -3, - -1) - * framework::dataset::make("min", -2, 0) * framework::dataset::make("max", 1, 3) * framework::dataset::make("addBias", { false, true }); - -using NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture = - GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture<Tensor, Accessor, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint>; - -// *INDENT-OFF* -// clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( - framework::dataset::make("InputAInfo", { TensorInfo(TensorShape(21U, 13U), 1, DataType::S32), // Input not a multiple of 16 - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), // Wrong output data type - }), - framework::dataset::make("InputBInfo",{ TensorInfo(TensorShape(21U), 1, DataType::S32), - TensorInfo(TensorShape(20U), 1, DataType::S32), - })), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(21U, 13U), 1, DataType::QSYMM16), - TensorInfo(TensorShape(20U, 13U), 1, DataType::S32), - })), - framework::dataset::make("Min",{ -205, - -180, - })), - framework::dataset::make("Max",{ 205, - 180, - })), - framework::dataset::make("Expected", { true, false })), - a_info, b_info, output_info, min, max, expected) -{ - // Lock tensors - Status status = NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPoint::validate(&a_info.clone()->set_is_resizable(false), - &b_info.clone()->set_is_resizable(false), - &output_info.clone()->set_is_resizable(false), - min, - max); - ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + validate(Accessor(_target), _reference, tolerance_dequantized); } -// clang-format on -// *INDENT-ON* +TEST_SUITE_END() // Dequant +#endif // __aarch64__ -TEST_SUITE(NoRelu) -TEST_SUITE(MultSmallerEq1) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_cases)) -{ - // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // MultSmallerEq1 -TEST_SUITE(MultGreater1) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_cases)) -{ - // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // MultGreater1 -TEST_SUITE_END() // NoRelu -TEST_SUITE(BoundedReLu) -TEST_SUITE(MultSmallerEq1) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_relu_cases)) -{ - // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // MultSmallerEq1 -TEST_SUITE(MultGreater1) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointFixture, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), - quantize_down_int32_to_int16_scale_by_fixedpoint_multgreat1_relu_cases)) -{ - // Validate output - validate(Accessor(_target), _reference); -} -TEST_SUITE_END() // MultGreater1 -TEST_SUITE_END() // BoundedReLu -TEST_SUITE_END() // QuantizeDownInt32ToInt16ScaleByFixedPoint -TEST_SUITE_END() // OutputStage +TEST_SUITE_END() // MatrixMultiplyCore TEST_SUITE_END() // GEMMLowp -TEST_SUITE_END() // Neon +TEST_SUITE_END() // NEON } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/NEON/Gather.cpp b/tests/validation/NEON/Gather.cpp index ca1e166bd1..0aea19939e 100644 --- a/tests/validation/NEON/Gather.cpp +++ b/tests/validation/NEON/Gather.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -100,12 +100,14 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( template <typename T> using NEGatherFixture = GatherFixture<Tensor, Accessor, NEGather, T>; +const auto gather_small_shapes = arm_compute::test::framework::dataset::concat(datasets::SmallGatherDataset(), datasets::SmallGatherMultiDimIndicesDataset()); + TEST_SUITE(Float) TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, NEGatherFixture<half>, framework::DatasetMode::PRECOMMIT, - combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::F16))) + combine(gather_small_shapes, framework::dataset::make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference); @@ -125,7 +127,7 @@ TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEGatherFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::F32))) + combine(gather_small_shapes, framework::dataset::make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference); @@ -146,7 +148,7 @@ TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, NEGatherFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::U8))) + combine(gather_small_shapes, framework::dataset::make("DataType", DataType::U8))) { // Validate output validate(Accessor(_target), _reference); @@ -166,7 +168,7 @@ TEST_SUITE(U16) FIXTURE_DATA_TEST_CASE(RunSmall, NEGatherFixture<uint16_t>, framework::DatasetMode::PRECOMMIT, - combine(datasets::SmallGatherDataset(), framework::dataset::make("DataType", DataType::U16))) + combine(gather_small_shapes, framework::dataset::make("DataType", DataType::U16))) { // Validate output validate(Accessor(_target), _reference); diff --git a/tests/validation/NEON/Im2Col.cpp b/tests/validation/NEON/Im2Col.cpp index 156957a601..ef5e75c5db 100644 --- a/tests/validation/NEON/Im2Col.cpp +++ b/tests/validation/NEON/Im2Col.cpp @@ -22,7 +22,7 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "src/core/NEON/kernels/NEIm2ColKernel.h" +#include "src/cpu/kernels/CpuIm2ColKernel.h" #include "tests/NEON/Accessor.h" #include "tests/NEON/Helper.h" #include "tests/datasets/ShapeDatasets.h" @@ -57,7 +57,7 @@ const auto conv_args_small = combine(combine(combine(combine(conv_filter TEST_SUITE(NEON) TEST_SUITE(Im2Col) -using NEIm2Col = NESynthetizeFunction<NEIm2ColKernel>; +using CpuIm2Col = NESynthetizeFunctionWithZeroConstantKernelBorder<cpu::kernels::CpuIm2ColKernel>; // *INDENT-OFF* // clang-format off @@ -78,26 +78,26 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( framework::dataset::make("Expected", { false, false, false, false, true })), input_info, output_info, has_bias, expected) { - bool status = bool(NEIm2Col::validate(&input_info, &output_info, Size2D(3U, 3U), PadStrideInfo(), has_bias)); + bool status = bool(cpu::kernels::CpuIm2ColKernel::validate(&input_info, &output_info, Size2D(3U, 3U), PadStrideInfo(), has_bias)); ARM_COMPUTE_EXPECT(status == expected, framework::LogLevel::ERRORS); } // clang-format on // *INDENT-ON* template <typename T> -using NEIm2ColFixture = Im2ColValidationFixture<Tensor, Accessor, NEIm2Col, T, false>; +using CpuIm2ColFixture = Im2ColOpValidationFixture<Tensor, Accessor, CpuIm2Col, T, false>; TEST_SUITE(Float) TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(im2col_shapes, framework::dataset::make("DataType", DataType::F32)), - conv_args_small)) +FIXTURE_DATA_TEST_CASE(RunSmall, CpuIm2ColFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(im2col_shapes, framework::dataset::make("DataType", DataType::F32)), + conv_args_small)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(concat(im2col_shapes, datasets::LargeShapes()), framework::dataset::make("DataType", - DataType::F32)), - conv_args)) +FIXTURE_DATA_TEST_CASE(RunLarge, CpuIm2ColFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(concat(im2col_shapes, datasets::LargeShapes()), framework::dataset::make("DataType", + DataType::F32)), + conv_args)) { // Validate output validate(Accessor(_target), _reference); @@ -107,15 +107,15 @@ TEST_SUITE_END() // FP32 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(im2col_shapes, framework::dataset::make("DataType", DataType::F16)), - conv_args_small)) +FIXTURE_DATA_TEST_CASE(RunSmall, CpuIm2ColFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(im2col_shapes, framework::dataset::make("DataType", DataType::F16)), + conv_args_small)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(concat(im2col_shapes, datasets::LargeShapes()), framework::dataset::make("DataType", - DataType::F16)), - conv_args)) +FIXTURE_DATA_TEST_CASE(RunLarge, CpuIm2ColFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(concat(im2col_shapes, datasets::LargeShapes()), framework::dataset::make("DataType", + DataType::F16)), + conv_args)) { // Validate output validate(Accessor(_target), _reference); @@ -127,15 +127,15 @@ TEST_SUITE_END() // FP16 TEST_SUITE_END() // Float TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEIm2ColFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(im2col_shapes, framework::dataset::make("DataType", DataType::QASYMM8)), - conv_args_small)) +FIXTURE_DATA_TEST_CASE(RunSmall, CpuIm2ColFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(im2col_shapes, framework::dataset::make("DataType", DataType::QASYMM8)), + conv_args_small)) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(concat(im2col_shapes, datasets::LargeShapes()), - framework::dataset::make("DataType", DataType::QASYMM8)), - conv_args)) +FIXTURE_DATA_TEST_CASE(RunLarge, CpuIm2ColFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(concat(im2col_shapes, datasets::LargeShapes()), + framework::dataset::make("DataType", DataType::QASYMM8)), + conv_args)) { // Validate output validate(Accessor(_target), _reference); @@ -165,8 +165,8 @@ TEST_CASE(PaddedChannelNHWC, framework::DatasetMode::PRECOMMIT) Tensor dst_target = create_tensor<Tensor>(dst_shape, data_type, 1, qinfo); // Configure target function - NEIm2Col im2col_func; - im2col_func.configure(&src_target, &dst_target, spatial_kernel, conv_info, has_bias); + CpuIm2Col im2col_func; + im2col_func.configure(src_target.info(), dst_target.info(), spatial_kernel, conv_info, has_bias); // Extend padding src_target.info()->extend_padding(PaddingSize(3, 5, 9, 1)); @@ -185,8 +185,13 @@ TEST_CASE(PaddedChannelNHWC, framework::DatasetMode::PRECOMMIT) // Fill target source library->fill_tensor_uniform(Accessor(src_target), 0); + ITensorPack pack = + { + { TensorType::ACL_SRC, &src_target }, + { TensorType::ACL_DST, &dst_target } + }; // Run target function - im2col_func.run(); + im2col_func.run(pack); // Calculate Reference SimpleTensor<float> src_ref{ src_shape, data_type, 1, qinfo, data_layout }; diff --git a/tests/validation/NEON/LSTMLayerQuantized.cpp b/tests/validation/NEON/LSTMLayerQuantized.cpp index d391267e3e..6b98ee2b67 100644 --- a/tests/validation/NEON/LSTMLayerQuantized.cpp +++ b/tests/validation/NEON/LSTMLayerQuantized.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -64,11 +64,7 @@ inline void fill_tensor(SimpleTensor<T> &tensor, const std::vector<T> &v) } /** Tolerance for quantized asymmetric operations */ -#if defined(__aarch64__) -constexpr AbsoluteTolerance<int16_t> tolerance_qsymm16(0); -#else // defined(__aarch64__) constexpr AbsoluteTolerance<int16_t> tolerance_qsymm16(1); -#endif // defined(__aarch64__) } // namespace diff --git a/tests/validation/NEON/MatMul.cpp b/tests/validation/NEON/MatMul.cpp new file mode 100644 index 0000000000..f22bd9e86a --- /dev/null +++ b/tests/validation/NEON/MatMul.cpp @@ -0,0 +1,467 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/NEON/functions/NEMatMul.h" + +#include "tests/datasets/LargeMatMulDataset.h" +#include "tests/datasets/SmallMatMulDataset.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Macros.h" +#include "tests/NEON/Accessor.h" +#include "tests/validation/fixtures/MatMulFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using framework::dataset::make; + +TEST_SUITE(NEON) +TEST_SUITE(MatMul) + +constexpr AbsoluteTolerance<float> tolerance_fp32( + 0.001f); /**< Tolerance value for comparing reference's output against implementation's output for FP32 data types */ +const AbsoluteTolerance<half> tolerance_fp16(half(0.1f)); +#ifdef __aarch64__ +constexpr AbsoluteTolerance<int32_t> tolerance_qasymm8(1); +constexpr AbsoluteTolerance<int32_t> tolerance_qasymm8_signed(1); +#endif // __aarch64__ + +// clang-format off +// *INDENT-OFF* +// Validation Tests +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, + zip( + make("InputAInfo", { + TensorInfo(TensorShape(9U, 6U), 1, DataType::F32), // Mismatching datatype + TensorInfo(TensorShape(9U, 6U), 1, DataType::S32), // Unsupported datatypes + TensorInfo(TensorShape(9U, 6U, 2U), 1, DataType::F32), // Broadcasting in batch dimension not supported + TensorInfo(TensorShape(9U, 6U), 1, DataType::F32), // Invalid shape for multiplication + TensorInfo(TensorShape(9U, 6U), 1, DataType::F32), + TensorInfo(TensorShape(9U, 6U , 12U) , 1 , DataType::F32), + TensorInfo(TensorShape(9U, 6U , 12U) , 1 , DataType::F32), // Tensors are not dynamic + TensorInfo(TensorShape(9U, 6U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(9U, 6U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(9U, 6U), 1, DataType::QASYMM8_SIGNED), // Mismatching data type + }), + make("InputBInfo", { + TensorInfo(TensorShape(5U, 9U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(5U, 9U), 1, DataType::S32), + TensorInfo(TensorShape(5U, 9U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 12U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 9U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 9U, 12U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 9U, 12U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 9U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(5U, 9U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(5U, 9U), 1, DataType::QASYMM8_SIGNED), + }), + make("OutputInfo", { + TensorInfo(TensorShape(5U, 6U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 6U), 1, DataType::S32), + TensorInfo(TensorShape(5U, 6U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 6U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 6U), 1, DataType::F32), + TensorInfo(TensorShape(5U, 6U, 12U) , 1, DataType::F32), + TensorInfo(TensorShape(5U, 6U, 12U) , 1, DataType::F32), + TensorInfo(TensorShape(5U, 6U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(5U, 6U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(5U, 6U), 1, DataType::QASYMM8), + }), + make("TensorIsConst", {false, false, false, false, false , false, true, false, false, false}), + make("Expected", { false, false, false, false, true, true, false, true, true, false })), + a_info, b_info, output_info, are_tensors_const, expected) +{ + TensorInfo a{a_info}; + TensorInfo b{b_info}; + a.set_are_values_constant(are_tensors_const); + b.set_are_values_constant(are_tensors_const); + Status status = NEMatMul::validate(&a, + &b, + &output_info, + MatMulInfo(), + CpuMatMulSettings()); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); +} +// *INDENT-ON* +// clang-format on + +// Generic Template +template <typename T> +using NEMatMulFixture = MatMulValidationWithActivationFixture<Tensor, Accessor, NEMatMul, CpuMatMulSettings, T>; + +// Fast math Template +template <typename T> +using NEMatMulFastMathFixture = MatMulGenericValidationFixture<Tensor, Accessor, NEMatMul, CpuMatMulSettings, T>; + +template <typename T> +using NEMatMulFixedFormatFixture = MatMulFixedFormatFixture<Tensor, Accessor, NEMatMul, CpuMatMulSettings, T>; + +template <typename T> +using NEMatMulDynamicTensorsFixture = + MatMulValidationWithDynamicTensorsFixture<Tensor, Accessor, NEMatMul, CpuMatMulSettings, T>; + +template <typename T> +using NEQuantizedMatMulFixture = QuantizedMatMulValidationFixture<Tensor, Accessor, NEMatMul, CpuMatMulSettings, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, + NEMatMulFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F32), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + NEMatMulFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F32), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} +FIXTURE_DATA_TEST_CASE(RunHighDimensions, + NEMatMulFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(datasets::HighDimensionalMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F32), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} + +FIXTURE_DATA_TEST_CASE(RunStressDynamicTensors, + NEMatMulDynamicTensorsFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F32), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}), +make("NumberOfRuns", 5))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() // FP32 + +#ifdef ARM_COMPUTE_ENABLE_BF16 +/* Note : MatMul BF16 is enabled by specifying FP32 datatype and enabling the fast math setting */ +constexpr AbsoluteTolerance<float> tolerance_bf16(0.02f); +TEST_SUITE(BF16) +FIXTURE_DATA_TEST_CASE(RunSmall, + NEMatMulFastMathFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F32), + make("ActivationInfo", {ActivationLayerInfo()}), + make("RunTimes", {0}), + make("Settings", {CpuMatMulSettings().fast_math(true)}), + make("LhsQInfo", {QuantizationInfo()}), + make("RhsQInfo", {QuantizationInfo()}), + make("OutQInfo", {QuantizationInfo()}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_bf16); +} + +#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +FIXTURE_DATA_TEST_CASE(RunTinyFixedFormat, + NEMatMulFixedFormatFixture<bfloat16>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::TinyMatMulDataset(), + make("TransposeA", {false}), + make("TransposeB", {false}), + make("DataType", DataType::BFLOAT16), + make("ActivationInfo", {ActivationLayerInfo()}), + make("RunTimes", {0}), + make("Settings", {CpuMatMulSettings().fast_math(true).fixed_format(true)}), + make("LhsQInfo", {QuantizationInfo()}), + make("RhsQInfo", {QuantizationInfo()}), + make("OutQInfo", {QuantizationInfo()}))) +{ + if (CPUInfo::get().has_bf16()) + { + // Validate output + validate(Accessor(_target), _reference, tolerance_bf16); + } +} +#endif /* ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS */ + +FIXTURE_DATA_TEST_CASE(RunLarge, + NEMatMulFastMathFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F32), + make("ActivationInfo", {ActivationLayerInfo()}), + make("RunTimes", {0}), + make("Settings", {CpuMatMulSettings().fast_math(true)}), + make("LhsQInfo", {QuantizationInfo()}), + make("RhsQInfo", {QuantizationInfo()}), + make("OutQInfo", {QuantizationInfo()}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_bf16, 0.01 /* tolerance_num */); +} +TEST_SUITE_END() // BF16 +#endif /* ARM_COMPUTE_ENABLE_BF16 */ + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, + NEMatMulFixture<half>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F16), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp16); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + NEMatMulFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F16), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp16); +} +FIXTURE_DATA_TEST_CASE(RunStressDynamicTensors, + NEMatMulDynamicTensorsFixture<half>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::F16), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}), +make("NumberOfRuns", 5))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp16); +} +TEST_SUITE_END() // FP16 +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + +TEST_SUITE_END() // Float + +#ifdef __aarch64__ // All the GeMM CPU assembly kernels for integer datatypes require aarch64 +TEST_SUITE(Quantized) + +TEST_SUITE(QASYMM8) + +FIXTURE_DATA_TEST_CASE(RunSmall, + NEQuantizedMatMulFixture<uint8_t>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::QASYMM8), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}), +make("NumberOfExtraRuns", {0, 1}), +make("LhsQInfo", {QuantizationInfo(1.f / 50, 1)}), +make("RhsQInfo", {QuantizationInfo(1.f / 30, -1)}), +make("OutQInfo", {QuantizationInfo(1.f, 2)}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +FIXTURE_DATA_TEST_CASE(RunSmallExtraActivation, + NEQuantizedMatMulFixture<uint8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::SmallerMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::QASYMM8), + make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) +}), +make("NumberOfExtraRuns", {0, 1}), +make("LhsQInfo", {QuantizationInfo(1.f / 50, 1)}), +make("RhsQInfo", {QuantizationInfo(1.f / 30, -1)}), +make("OutQInfo", {QuantizationInfo(1.f, 2)}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, + NEQuantizedMatMulFixture<uint8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::QASYMM8), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}), +make("NumberOfExtraRuns", {0, 1}), +make("LhsQInfo", {QuantizationInfo(1.f / 100, 1)}), +make("RhsQInfo", {QuantizationInfo(1.f / 200, -1)}), +make("OutQInfo", {QuantizationInfo(1.f, 2)}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) + +FIXTURE_DATA_TEST_CASE(RunSmall, + NEQuantizedMatMulFixture<int8_t>, + framework::DatasetMode::PRECOMMIT, + combine(datasets::SmallMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::QASYMM8_SIGNED), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}), +make("NumberOfExtraRuns", {0, 1}), +make("LhsQInfo", {QuantizationInfo(1.f / 40, -2)}), +make("RhsQInfo", {QuantizationInfo(1.f / 50, 1)}), +make("OutQInfo", {QuantizationInfo(1.f, 1)}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} + +FIXTURE_DATA_TEST_CASE(RunSmallExtraActivation, + NEQuantizedMatMulFixture<int8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::SmallerMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::QASYMM8_SIGNED), + make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU) +}), +make("NumberOfExtraRuns", {0, 1}), +make("LhsQInfo", {QuantizationInfo(1.f / 40, -2)}), +make("RhsQInfo", {QuantizationInfo(1.f / 50, 1)}), +make("OutQInfo", {QuantizationInfo(1.f, 1)}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, + NEQuantizedMatMulFixture<int8_t>, + framework::DatasetMode::NIGHTLY, + combine(datasets::LargeMatMulDataset(), + make("TransposeA", {false, true}), + make("TransposeB", {false, true}), + make("DataType", DataType::QASYMM8_SIGNED), + make("ActivationInfo", +{ + ActivationLayerInfo(), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU) +}), +make("NumberOfExtraRuns", {0, 1}), +make("LhsQInfo", {QuantizationInfo(1.f / 150, -2)}), +make("RhsQInfo", {QuantizationInfo(1.f / 250, 1)}), +make("OutQInfo", {QuantizationInfo(1.f, 1)}))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_signed); +} + +TEST_SUITE_END() // QASYMM8_SIGNED + +TEST_SUITE_END() // Quantized +#endif // __aarch64__ + +TEST_SUITE_END() // MatMul +TEST_SUITE_END() // NEON +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/MaxUnpoolingLayer.cpp b/tests/validation/NEON/MaxUnpoolingLayer.cpp index 27f131fa51..0eb021fe71 100644 --- a/tests/validation/NEON/MaxUnpoolingLayer.cpp +++ b/tests/validation/NEON/MaxUnpoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2021 Arm Limited. + * Copyright (c) 2020-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,10 +22,12 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/StringUtils.h" #include "arm_compute/runtime/NEON/functions/NEMaxUnpoolingLayer.h" #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/cpu/kernels/CpuMaxUnpoolingLayerKernel.h" #include "tests/NEON/Accessor.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/framework/Asserts.h" @@ -33,7 +35,6 @@ #include "tests/framework/datasets/Datasets.h" #include "tests/validation/Validation.h" #include "tests/validation/fixtures/MaxUnpoolingLayerFixture.h" - namespace arm_compute { namespace test @@ -51,7 +52,7 @@ const auto PoolingLayerIndicesDatasetFPSmall = combine(combine(framework::datase TEST_SUITE(Float) TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(MaxUnpooling, NEMaxUnpoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerIndicesDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(MaxUnpooling, NEMaxUnpoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerIndicesDatasetFPSmall, framework::dataset::make("DataType", DataType::F32))), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }) @@ -63,7 +64,7 @@ FIXTURE_DATA_TEST_CASE(MaxUnpooling, NEMaxUnpoolingLayerFixture<float>, framewor TEST_SUITE_END() // FP32 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(MaxUnpooling, NEMaxUnpoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerIndicesDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(MaxUnpooling, NEMaxUnpoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerIndicesDatasetFPSmall, framework::dataset::make("DataType", DataType::F16))), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }) @@ -74,7 +75,37 @@ FIXTURE_DATA_TEST_CASE(MaxUnpooling, NEMaxUnpoolingLayerFixture<half>, framework } TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ + TEST_SUITE_END() // Float + +TEST_SUITE(KernelSelection) + +DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, + combine(framework::dataset::make("CpuExt", std::string("NEON")), + framework::dataset::make("DataType", { DataType::F32, + DataType::F16, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED + })), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "NEON"); + cpu_isa.sve = (cpu_ext == "SVE"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuMaxUnpoolingLayerKernel::get_implementation(DataTypeISASelectorData{ data_type, cpu_isa }, cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = lower_string(cpu_ext) + "_" + cpu_impl_dt(data_type) + "_maxunpooling"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() // KernelSelection TEST_SUITE_END() // PoolingLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp b/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp index 90d3d05a0d..085f3608a0 100644 --- a/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp +++ b/tests/validation/NEON/MeanStdDevNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,8 @@ namespace #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC RelativeTolerance<half> tolerance_f16(half(0.2f)); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -RelativeTolerance<float> tolerance_f32(1e-8f); +RelativeTolerance<float> tolerance_f32(1e-4f); +RelativeTolerance<uint8_t> tolerance_qasymm8(1); } // namespace TEST_SUITE(NEON) @@ -114,9 +115,23 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEMeanStdDevNormalizationLayerFixture<float>, f // Validate output validate(Accessor(_target), _reference, tolerance_f32); } + TEST_SUITE_END() // FP32 TEST_SUITE_END() // Float +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEMeanStdDevNormalizationLayerFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small2DShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("InPlace", { false, true })), + framework::dataset::make("Epsilon", { 1e-7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() // Quantized +TEST_SUITE_END() // QASYMM8 + TEST_SUITE_END() // MeanStdNormalizationLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/PixelWiseMultiplication.cpp b/tests/validation/NEON/PixelWiseMultiplication.cpp index 1bb0588919..964d1c5deb 100644 --- a/tests/validation/NEON/PixelWiseMultiplication.cpp +++ b/tests/validation/NEON/PixelWiseMultiplication.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -56,6 +56,11 @@ const auto PixelWiseMultiplicationQASYMM8QuantDataset = combine(combine( framework::dataset::make("Src1QInfo", { QuantizationInfo(2.f / 32768.f, 0) })), framework::dataset::make("OutQInfo", { QuantizationInfo(1.f / 32768.f, 0) })); +const auto PixelWiseMultiplicationQASYMM8QuantInPlaceDataset = combine(combine( + framework::dataset::make("Src0QInfo", { QuantizationInfo(5.f / 32768.f, 10) }), + framework::dataset::make("Src1QInfo", { QuantizationInfo(5.f / 32768.f, 10) })), + framework::dataset::make("OutQInfo", { QuantizationInfo(5.f / 32768.f, 10) })); + const auto PixelWiseMultiplicationPolicySTNUDataset = combine( framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE }), framework::dataset::make("RoundingPolicy", { RoundingPolicy::TO_NEAREST_UP })); @@ -75,7 +80,8 @@ const auto PixelWiseMultiplicationPolicySTZDataset = combine( * expected to have either different quantization information, data type * or different shape we are not testing in-place computation. */ -const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto InPlaceDataSet = framework::dataset::make("InPlace", { false, true }); +const auto OutOfPlaceDataSet = framework::dataset::make("InPlace", { false }); #define DEFAULT_VALIDATE validate(Accessor(_target), _reference); #define VALIDATE(TYPE, TOLERANCE) validate(Accessor(_target), _reference, AbsoluteTolerance<TYPE>(TOLERANCE), 0.f); @@ -275,7 +281,19 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQASYMM8SignedFixture, framework::dataset::make("Scale", { scale_unity })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQASYMM8QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunSmallInPlace, NEPixelWiseMultiplicationQASYMM8SignedFixture, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make("DataTypeIn1", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeIn2", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeOut", DataType::QASYMM8_SIGNED)), + framework::dataset::make("Scale", { scale_unity })), + PixelWiseMultiplicationPolicySTZDataset), + PixelWiseMultiplicationQASYMM8QuantInPlaceDataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -292,7 +310,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQASYMM8Fixture, framew framework::dataset::make("Scale", { scale_255 })), PixelWiseMultiplicationPolicySTNUDataset), PixelWiseMultiplicationQASYMM8QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -306,7 +324,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQASYMM8Fixture, framew framework::dataset::make("Scale", { scale_unity })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQASYMM8QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -320,7 +338,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQASYMM8Fixture, framew framework::dataset::make("Scale", { scale_other })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQASYMM8QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -335,7 +353,20 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationBroadcastQASYMM8Fixtur framework::dataset::make("Scale", { scale_other })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQASYMM8QuantDataset), - framework::dataset::make("InPlace", { false }))) + OutOfPlaceDataSet)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunTinyInPlace, NEPixelWiseMultiplicationBroadcastQASYMM8Fixture, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(combine(datasets::TinyShapesBroadcastInplace(), + framework::dataset::make("DataTypeIn1", DataType::QASYMM8)), + framework::dataset::make("DataTypeIn2", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", DataType::QASYMM8)), + framework::dataset::make("Scale", { scale_other })), + PixelWiseMultiplicationPolicySTZDataset), + PixelWiseMultiplicationQASYMM8QuantInPlaceDataset), + InPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -351,7 +382,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQSYMM16Fixture, framew framework::dataset::make("Scale", { scale_255 })), PixelWiseMultiplicationPolicySTNUDataset), PixelWiseMultiplicationQSYMM16QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qsymm16); @@ -365,7 +396,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQSYMM16Fixture, framew framework::dataset::make("Scale", { scale_unity })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQSYMM16QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qsymm16); @@ -379,7 +410,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQSYMM16Fixture, framew framework::dataset::make("Scale", { scale_other })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQSYMM16QuantDataset), - InPlaceDataSet)) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference, tolerance_qsymm16); @@ -394,7 +425,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationQSYMM16ToS32Fixture, f framework::dataset::make("Scale", { scale_unity })), PixelWiseMultiplicationPolicySTZDataset), PixelWiseMultiplicationQSYMM16QuantDataset), - framework::dataset::make("InPlace", { false }))) + OutOfPlaceDataSet)) { // Validate output validate(Accessor(_target), _reference); @@ -411,7 +442,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPixelWiseMultiplicationU8U8ToS16Fixture, fram framework::dataset::make("Scale", { scale_255 })), datasets::ConvertPolicies()), framework::dataset::make("RoundingPolicy", RoundingPolicy::TO_NEAREST_UP)), - framework::dataset::make("InPlace", { false }))) + OutOfPlaceDataSet)) { // Validate output validate_wrap(Accessor(_target), _reference, AbsoluteTolerance<int16_t>(1), 0.f); @@ -451,17 +482,17 @@ TEST_SUITE_END() // U8toU8 TEST_SUITE(U8toS16) TEST_SUITE(Scale255) -PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS16Fixture<uint8_t>, ALL, SmallShapes(), U8, S16, S16, scale_255, TO_NEAREST_UP, framework::dataset::make("InPlace", { false }), +PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS16Fixture<uint8_t>, ALL, SmallShapes(), U8, S16, S16, scale_255, TO_NEAREST_UP, OutOfPlaceDataSet, WRAP_VALIDATE(int16_t, 2)) TEST_SUITE_END() // Scale255 TEST_SUITE(ScaleUnity) -PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS16Fixture<uint8_t>, ALL, SmallShapes(), U8, S16, S16, scale_unity, TO_ZERO, framework::dataset::make("InPlace", { false }), +PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS16Fixture<uint8_t>, ALL, SmallShapes(), U8, S16, S16, scale_unity, TO_ZERO, OutOfPlaceDataSet, DEFAULT_VALIDATE) TEST_SUITE_END() // ScaleUnity TEST_SUITE(ScaleOther) -PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS16Fixture<uint8_t>, ALL, SmallShapes(), U8, S16, S16, scale_other, TO_ZERO, framework::dataset::make("InPlace", { false }), +PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, ToS16Fixture<uint8_t>, ALL, SmallShapes(), U8, S16, S16, scale_other, TO_ZERO, OutOfPlaceDataSet, DEFAULT_VALIDATE) TEST_SUITE_END() // ScaleOther diff --git a/tests/validation/NEON/Pooling3dLayer.cpp b/tests/validation/NEON/Pooling3dLayer.cpp new file mode 100644 index 0000000000..07054462f5 --- /dev/null +++ b/tests/validation/NEON/Pooling3dLayer.cpp @@ -0,0 +1,361 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/runtime/NEON/functions/NEPooling3dLayer.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" +#include "tests/NEON/Accessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/Pooling3dLayerDataset.h" +#include "tests/datasets/PoolingTypesDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/Pooling3dLayerFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +/** Input data sets for floating-point data types */ +const auto Pooling3dLayerDatasetFP = combine(combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size3D(2, 3, 2) })), + framework::dataset::make("Stride", { Size3D(1, 1, 1), Size3D(2, 1, 1), Size3D(1, 2, 1), Size3D(2, 2, 1) })), + framework::dataset::make("Padding", { Padding3D(0, 1, 0), Padding3D(1, 1, 1) })), + framework::dataset::make("ExcludePadding", { true, false })); + +const auto Pooling3dLayerDatasetFPSmall = combine(combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size3D(2, 2, 2), Size3D(3, 3, 3) })), + framework::dataset::make("Stride", { Size3D(2, 2, 2), Size3D(2, 1, 1) })), + framework::dataset::make("Padding", { Padding3D(0, 0, 0), Padding3D(1, 1, 1), Padding3D(1, 0, 0) })), + framework::dataset::make("ExcludePadding", { true, false })); + +const auto Pooling3dLayerDatasetQASYMM8Small = combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), + framework::dataset::make("PoolingSize", { Size3D(3, 3, 3) })), + framework::dataset::make("Stride", { Size3D(1, 1, 1), Size3D(2, 1, 1), Size3D(1, 2, 1), Size3D(2, 2, 1) })), + framework::dataset::make("Padding", { Padding3D(0, 0, 0), Padding3D(1, 1, 1), Padding3D(1, 0, 0) })), + framework::dataset::make("ExcludePadding", { true })); + +const auto Pooling3dLayerDatasetQASYMM8Large = combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), + framework::dataset::make("PoolingSize", { Size3D(3, 3, 3) })), + framework::dataset::make("Stride", { Size3D(1, 1, 1), Size3D(2, 2, 1) })), + framework::dataset::make("Padding", { Padding3D(0, 0, 0), Padding3D(1, 1, 0) })), + framework::dataset::make("ExcludePadding", { true })); + +using ShapeDataset = framework::dataset::ContainerDataset<std::vector<TensorShape>>; + +constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */ +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */ +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for unsigned 8-bit asymmetric type */ +constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_s(1); /**< Tolerance value for comparing reference's output against implementation's output for signed 8-bit asymmetric type */ + +const auto qasymm8_in_qinfo_dataset = framework::dataset::make("InputQuantInfo", { QuantizationInfo(.2f, 10) }); +const auto qasymm8_out_qinfo_dataset = framework::dataset::make("OutputQuantInfo", +{ + QuantizationInfo(.2f, 10), // Same qinfo + QuantizationInfo(.1f, 5), // Multiplier <= 1 + QuantizationInfo(2.f, 3) // Multiplier > 1 +}); + +const auto qasymm8_signed_in_qinfo_dataset = framework::dataset::make("InputQuantInfo", { QuantizationInfo(.2f, -10) }); +const auto qasymm8_signed_out_qinfo_dataset = framework::dataset::make("OutputQuantInfo", +{ + QuantizationInfo(.2f, -10), // Same qinfo + QuantizationInfo(.1f, -5), // Multiplier <= 1 + QuantizationInfo(2.f, -3) // Multiplier > 1 +}); + +} //namespace + +TEST_SUITE(NEON) +TEST_SUITE(Pooling3dLayer) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Mismatching data type + TensorInfo(TensorShape(2U, 27U, 13U, 4U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid pad/size combination + TensorInfo(TensorShape(2U, 27U, 13U, 4U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid pad/size combination + TensorInfo(TensorShape(2U, 27U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Invalid output shape + TensorInfo(TensorShape(5U, 13U, 15U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Global Pooling + TensorInfo(TensorShape(13U,13U, 5U, 1U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid output Global Pooling + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 4U), 1, DataType::F32, DataLayout::NDHWC), // Invalid data type + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 4U), 1, DataType::F32, DataLayout::NHWC), // Invalid data layout + TensorInfo(TensorShape(5U, 13U, 13U, 5U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(1U, 16U, 1U, 3U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 13U, 13U, 4U, 3U), 1, DataType::F32, DataLayout::NDHWC), + }), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(2U, 25U, 11U, 3U, 3U), 1, DataType::F16, DataLayout::NDHWC), + TensorInfo(TensorShape(2U, 30U, 11U, 3U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(2U, 25U, 16U, 3U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(2U, 27U, 13U, 3U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 1U, 1U, 1U, 3U), 1, DataType::F32, DataLayout::NDHWC), // Global pooling applied + TensorInfo(TensorShape(5U, 2U, 2U, 2U, 2U), 1, DataType::F32, DataLayout::NDHWC), // Invalid output Global Pooling + TensorInfo(TensorShape(5U, 12U, 12U, 3U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 12U, 12U, 3U, 4U), 1, DataType::QASYMM8, DataLayout::NDHWC), // Invalid data type + TensorInfo(TensorShape(5U, 12U, 12U, 3U, 4U), 1, DataType::F32, DataLayout::NDHWC), // Invalid data layout + TensorInfo(TensorShape(5U, 1U, 1U, 1U, 4U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(1U, 15U, 1U, 2U, 4U), 1, DataType::F32, DataLayout::NDHWC), // size larger than height + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 2U), 1, DataType::F32, DataLayout::NDHWC), + TensorInfo(TensorShape(5U, 6U, 6U, 2U, 3U), 1, DataType::F32, DataLayout::NDHWC), + })), + framework::dataset::make("PoolInfo", { Pooling3dLayerInfo(PoolingType::AVG, 3, Size3D(1, 1, 1), Padding3D(0, 0, 0)), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1, 1, 1), Padding3D(2, 0, 0)), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1, 1, 1), Padding3D(0, 0, 0)), + Pooling3dLayerInfo(PoolingType::L2, 3, Size3D(1, 1, 1), Padding3D(0, 0, 0)), + Pooling3dLayerInfo(PoolingType::AVG), + Pooling3dLayerInfo(PoolingType::MAX), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1U, 1U, 1U), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(1U, 1U, 1U), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG), + Pooling3dLayerInfo(PoolingType::MAX, 2, Size3D(1, 1, 2), Padding3D(0, 0, 0), false), + Pooling3dLayerInfo(PoolingType::AVG, 2, Size3D(2U, 2U, 2U), Padding3D(), false), + Pooling3dLayerInfo(PoolingType::AVG, 1, Size3D(2U, 2U, 2U), Padding3D(2, 2, 2), true), // pool size is equal to the padding size + Pooling3dLayerInfo(PoolingType::AVG, 1, Size3D(2U, 2U, 2U), Padding3D(2, 2, 2), false), // pool size is equal to the padding size + Pooling3dLayerInfo(PoolingType::AVG, 3, Size3D(2U, 2U, 2U), Padding3D(2,1,2,2,1,2), false, false, DimensionRoundingType::CEIL), // CEIL with asymmetric Padding + })), + framework::dataset::make("Expected", { false, false, false, false, true, false, false, false, false, true , false, true, false, false, false})), + input_info, output_info, pool_info, expected) +{ + bool is_valid = bool(NEPooling3dLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pool_info)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using NEPoolingLayer3dFixture = Pooling3dLayerValidationFixture<Tensor, Accessor, NEPooling3dLayer, T>; + +template <typename T> +using NESpecial3dPoolingLayerFixture = SpecialPooling3dLayerValidationFixture<Tensor, Accessor, NEPooling3dLayer, T>; + +template <typename T> +using NEPooling3dLayerGlobalFixture = Pooling3dLayerGlobalValidationFixture<Tensor, Accessor, NEPooling3dLayer, T>; + +// clang-format on +// *INDENT-ON* +TEST_SUITE(Float) +TEST_SUITE(FP32) + +FIXTURE_DATA_TEST_CASE(RunSpecial, NESpecial3dPoolingLayerFixture<float>, framework::DatasetMode::ALL, datasets::Pooling3dLayerDatasetSpecial() * framework::dataset::make("DataType", DataType::F32)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayer3dFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small5dShapes(), combine(Pooling3dLayerDatasetFPSmall, + framework::dataset::make("DataType", DataType::F32)))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayer3dFixture<float>, framework::DatasetMode::NIGHTLY, + combine(datasets::Large5dShapes(), combine(Pooling3dLayerDatasetFPSmall, framework::dataset::make("DataType", DataType::F32)))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE(GlobalPooling) +// *INDENT-OFF* +// clang-format off +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayer3dFixture<float>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(3U, 27U, 13U, 4U), + TensorShape(4U, 27U, 13U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(27, 13, 4) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", {false, true})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunGlobalSmall, NEPooling3dLayerGlobalFixture<float>, framework::DatasetMode::ALL, + combine(combine( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 4U, 3U), + TensorShape(27U, 13U, 4U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayer3dFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(4U, 79U, 37U, 11U), + TensorShape(4U, 79U, 37U, 11U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(79, 37, 11) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", {false, true})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() // GlobalPooling +TEST_SUITE_END() // FP32 + +#ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +TEST_SUITE(FP16) + +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayer3dFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small5x5Shapes(), combine(Pooling3dLayerDatasetFPSmall, + framework::dataset::make("DataType", DataType::F16)))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} + + +FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayer3dFixture<half>, framework::DatasetMode::NIGHTLY, combine(datasets::Large5dShapes(), combine(Pooling3dLayerDatasetFP, + framework::dataset::make("DataType", + DataType::F16)))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE(GlobalPooling) +// *INDENT-OFF* +// clang-format off +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayer3dFixture<half>, framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(3U, 27U, 13U, 4U), + TensorShape(4U, 27U, 13U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(27, 13, 4) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", {false, true})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} + + +FIXTURE_DATA_TEST_CASE(RunSmallGlobal, NEPooling3dLayerGlobalFixture<half>, framework::DatasetMode::ALL, + combine(combine( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 4U, 3U), + TensorShape(27U, 13U, 4U, 4U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayer3dFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine( + framework::dataset::make("InputShape", { TensorShape(4U, 79U, 37U, 11U), + TensorShape(4U, 79U, 37U, 11U, 2U) + }), + framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size3D(79, 37, 11) })), + framework::dataset::make("Strides", Size3D(1, 1, 1))), + framework::dataset::make("Paddings", Padding3D(0, 0, 0))), + framework::dataset::make("ExcludePadding", false)), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} + +// clang-format on +// *INDENT-ON* +TEST_SUITE_END() // GlobalPooling +TEST_SUITE_END() // FP16 +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +TEST_SUITE_END() // Float +TEST_SUITE(Quantized) + +template <typename T> +using NEPooling3dLayerQuantizedFixture = Pooling3dLayerValidationQuantizedFixture<Tensor, Accessor, NEPooling3dLayer, T>; + +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEPooling3dLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small5dShapes(), + combine(Pooling3dLayerDatasetQASYMM8Small, + framework::dataset::make("DataType", DataType::QASYMM8))), + qasymm8_in_qinfo_dataset), + qasymm8_out_qinfo_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEPooling3dLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::Large5dShapes(), + combine(Pooling3dLayerDatasetQASYMM8Large, + framework::dataset::make("DataType", DataType::QASYMM8))), + qasymm8_in_qinfo_dataset), + qasymm8_out_qinfo_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} + +TEST_SUITE_END() // QASYMM8 + +TEST_SUITE(QASYMM8_SIGNED) + +FIXTURE_DATA_TEST_CASE(RunSmall, NEPooling3dLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small5dShapes(), + combine(Pooling3dLayerDatasetQASYMM8Small, + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))), + qasymm8_signed_in_qinfo_dataset), + qasymm8_signed_out_qinfo_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8_s); +} + +TEST_SUITE_END() // QASYMM8_SIGNED +TEST_SUITE_END() // Quantized +TEST_SUITE_END() // Pooling3dLayer +TEST_SUITE_END() // NEON +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/PoolingLayer.cpp b/tests/validation/NEON/PoolingLayer.cpp index b70a18907f..161fe627cc 100644 --- a/tests/validation/NEON/PoolingLayer.cpp +++ b/tests/validation/NEON/PoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,7 +24,6 @@ #include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/functions/NEPoolingLayer.h" #include "arm_compute/runtime/Tensor.h" -#include "arm_compute/runtime/TensorAllocator.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/PoolingLayerDataset.h" @@ -81,6 +80,14 @@ const auto qasymm8_signed_out_qinfo_dataset = framework::dataset::make("OutputQu QuantizationInfo(.1f, -5), // Multiplier <= 1 QuantizationInfo(2.f, -3) // Multiplier > 1 }); + +// Cases where pooling region is completely outside the input tensor (excluding global pooling) +const auto pool_outside_input_dataset = zip(zip(zip(zip( + framework::dataset::make("Shape", { TensorShape{ 2U, 2U, 1U }, TensorShape{ 2U, 2U, 4U }, TensorShape{ 3U, 5U, 2U }, TensorShape{ 10U, 20U, 3U } }), + framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG, PoolingType::L2, PoolingType::MAX })), + framework::dataset::make("PoolingSize", { Size2D{ 2, 2 }, Size2D{ 3, 3 }, Size2D{ 2, 2 }, Size2D{ 3, 6 } })), + framework::dataset::make("PadStride", { PadStrideInfo{ 1, 1, 2, 2 }, PadStrideInfo{ 1, 1, 4, 4 }, PadStrideInfo{ 1, 1, 3, 3 }, PadStrideInfo{ 1, 1, 2, 5 } })), + framework::dataset::make("ExcludePadding", { false, false, false, false })); } // namespace TEST_SUITE(NEON) @@ -98,7 +105,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(13U, 13U, 5U), 1, DataType::QASYMM8), // Invalid exclude_padding = false with quantized type, no actual padding and NHWC TensorInfo(TensorShape(13U, 13U, 5U), 1, DataType::F32), TensorInfo(TensorShape(1U, 16U, 1U), 1, DataType::F32), - }), + TensorInfo(TensorShape(112, 112, 64,1), 1, DataType::F32, DataLayout::NHWC), // Mismatching number of channels + TensorInfo(TensorShape(112, 112, 64,1), 1, DataType::F32, DataLayout::NHWC), // Mismatching width + }), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F16), TensorInfo(TensorShape(25U, 10U, 2U), 1, DataType::F32), TensorInfo(TensorShape(30U, 11U, 2U), 1, DataType::F32), @@ -108,7 +117,10 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(12U, 12U, 5U), 1, DataType::QASYMM8), TensorInfo(TensorShape(25U, 11U, 2U), 1, DataType::F32), TensorInfo(TensorShape(1U, 15U, 1U), 1, DataType::F32), - })), + TensorInfo(TensorShape(56, 56, 64,1), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(56, 51, 64,1), 1, DataType::F32, DataLayout::NHWC), + + })), framework::dataset::make("PoolInfo", { PoolingLayerInfo(PoolingType::AVG, 3, DataLayout::NCHW, PadStrideInfo(1, 1, 0, 0)), PoolingLayerInfo(PoolingType::AVG, 3, DataLayout::NCHW, PadStrideInfo(1, 1, 0, 0)), PoolingLayerInfo(PoolingType::AVG, 2, DataLayout::NCHW, PadStrideInfo(1, 1, 2, 0)), @@ -118,8 +130,11 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( PoolingLayerInfo(PoolingType::AVG, 2, DataLayout::NHWC, PadStrideInfo(), false), PoolingLayerInfo(PoolingType::AVG, DataLayout::NCHW), PoolingLayerInfo(PoolingType::MAX, 2, DataLayout::NHWC, PadStrideInfo(1, 1, 0, 0), false), + PoolingLayerInfo(PoolingType::MAX,3,DataLayout::NHWC,PadStrideInfo(2,2,1,1)), + PoolingLayerInfo(PoolingType::MAX,3,DataLayout::NHWC,PadStrideInfo(2,2,1,1)), + })), - framework::dataset::make("Expected", { false, false, false, false, true, false, true, false, false})), + framework::dataset::make("Expected", { false, false, false, false, true, false, true, false, false, false, false})), input_info, output_info, pool_info, expected) { bool is_valid = bool(NEPoolingLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pool_info)); @@ -142,13 +157,26 @@ using NESpecialPoolingLayerFixture = SpecialPoolingLayerValidationFixture<Tensor const auto PoolingLayerIndicesDatasetFPSmall = combine(combine(combine(framework::dataset::make("PoolType", { PoolingType::MAX }), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0) })), framework::dataset::make("ExcludePadding", { true, false })); - +const auto PoolingLayerKernelIndicesDatasetFPSmall = combine(combine(combine(framework::dataset::make("PoolType", { PoolingType::MAX }), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3), Size2D(7, 7) })), + framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 1, 1, 1) })), + framework::dataset::make("ExcludePadding", { false })); TEST_SUITE(Float) TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunIndices, NEPoolingLayerIndicesFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerIndicesDatasetFPSmall, - framework::dataset::make("DataType", - DataType::F32))), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) +FIXTURE_DATA_TEST_CASE(RunIndices, NEPoolingLayerIndicesFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallNoneUnitShapes(), + combine(PoolingLayerIndicesDatasetFPSmall, + framework::dataset::make("DataType", DataType::F32))), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("UseKernelIndices", { false }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); + validate(Accessor(_target_indices), _ref_indices); +} +FIXTURE_DATA_TEST_CASE(RunKernelIndices, NEPoolingLayerIndicesFixture<float>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SmallNoneUnitShapes(), + combine(PoolingLayerKernelIndicesDatasetFPSmall, + framework::dataset::make("DataType", DataType::F32))), + framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("UseKernelIndices", { true }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); @@ -159,7 +187,7 @@ FIXTURE_DATA_TEST_CASE(RunSpecial, NESpecialPoolingLayerFixture<float>, framewor // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetFPSmall, framework::dataset::make("DataType", DataType::F32))), pool_data_layout_dataset)) @@ -167,7 +195,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixture<float>, framework::Datase // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(2, 1, 0, 0) })), @@ -186,22 +214,32 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayerFixture<float>, framework::Datase // Validate output validate(Accessor(_target), _reference, tolerance_f32); } +TEST_SUITE(CornerCases) +FIXTURE_DATA_TEST_CASE(PoolRegionCompletelyOutsideInput, NEPoolingLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(pool_outside_input_dataset, + framework::dataset::make("DataType", + DataType::F32)), + pool_data_layout_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // CornerCases TEST_SUITE_END() // FP32 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunIndices, NEPoolingLayerIndicesFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerIndicesDatasetFPSmall, - framework::dataset::make("DataType", - DataType::F16))), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }) - - )) +FIXTURE_DATA_TEST_CASE(RunIndices, NEPoolingLayerIndicesFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallNoneUnitShapes(), + combine(PoolingLayerIndicesDatasetFPSmall, + framework::dataset::make("DataType", + DataType::F16))), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("UseKernelIndices", { false }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); validate(Accessor(_target_indices), _ref_indices); } -FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFPSmall, +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetFPSmall, framework::dataset::make("DataType", DataType::F16))), pool_data_layout_dataset)) { @@ -215,6 +253,16 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayerFixture<half>, framework::Dataset // Validate output validate(Accessor(_target), _reference, tolerance_f16); } +TEST_SUITE(CornerCases) +FIXTURE_DATA_TEST_CASE(PoolRegionCompletelyOutsideInput, NEPoolingLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(pool_outside_input_dataset, + framework::dataset::make("DataType", + DataType::F16)), + pool_data_layout_dataset)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // CornerCases TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE_END() // Float @@ -227,7 +275,7 @@ template <typename T> using NEPoolingLayerQuantizedMixedDataLayoutFixture = PoolingLayerValidationQuantizedFixture<Tensor, Accessor, NEPoolingLayer, T, true>; TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmallNCHW, NEPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallNCHW, NEPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetQASYMM8Small, framework::dataset::make("DataType", DataType::QASYMM8))), framework::dataset::make("DataLayout", { DataLayout::NCHW })), @@ -237,7 +285,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallNCHW, NEPoolingLayerQuantizedFixture<uint8_t>, fr // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetQASYMM8Small, framework::dataset::make("DataType", DataType::QASYMM8))), framework::dataset::make("DataLayout", { DataLayout::NHWC })), @@ -247,7 +295,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture<uint8_t>, framew // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 2, 1, 1) })), @@ -262,7 +310,7 @@ FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerQuantizedMixedDataLayou } TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(PoolingLayerDatasetQASYMM8Small, framework::dataset::make("DataType", DataType::QASYMM8_SIGNED))), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), @@ -272,7 +320,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerQuantizedFixture<int8_t>, framewo // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_s); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEPoolingLayerQuantizedMixedDataLayoutFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallNoneUnitShapes(), combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 2, 1, 1) })), diff --git a/tests/validation/NEON/QuantizationLayer.cpp b/tests/validation/NEON/QuantizationLayer.cpp index aeee54c835..bab7490762 100644 --- a/tests/validation/NEON/QuantizationLayer.cpp +++ b/tests/validation/NEON/QuantizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -34,6 +34,7 @@ #include "tests/validation/Validation.h" #include "tests/validation/fixtures/QuantizationLayerFixture.h" + namespace arm_compute { namespace test @@ -182,7 +183,16 @@ FIXTURE_DATA_TEST_CASE(RunSmallQASYMM8, NEQuantizationLayerQASYMM8GenFixture<uin framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataTypeOut", { DataType::QASYMM8 })), framework::dataset::make("QuantizationInfoOutput", { QuantizationInfo(0.5f, 10) })), - framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(2.0f, 15) }))) + framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(2.0f, 15), QuantizationInfo(0.5f, 25) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_u8); +} +FIXTURE_DATA_TEST_CASE(ConvertUint8toInt8, NEQuantizationLayerQASYMM8GenFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(QuantizationSmallShapes, + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("DataTypeOut", { DataType::QASYMM8_SIGNED })), + framework::dataset::make("QuantizationInfoOutput", { QuantizationInfo(2.0f, -1) })), + framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(2.0f, 127) }))) { // Validate output validate(Accessor(_target), _reference, tolerance_u8); @@ -191,7 +201,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQASYMM8_SIGNED, NEQuantizationLayerQASYMM8_SIGNED framework::dataset::make("DataTypeIn", DataType::QASYMM8)), framework::dataset::make("DataTypeOut", { DataType::QASYMM8_SIGNED })), framework::dataset::make("QuantizationInfoOutput", { QuantizationInfo(1.0f, 10), QuantizationInfo(2.0f, -25) })), - framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(1.0f, 15) }))) + framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(1.0f, 15), QuantizationInfo(1.0f, 127) }))) { // Validate output validate(Accessor(_target), _reference, tolerance_s8); @@ -211,7 +221,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQASYMM8_SIGNED, NEQuantizationLayerQASYMM8_SIGNED framework::dataset::make("DataTypeIn", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataTypeOut", { DataType::QASYMM8_SIGNED })), framework::dataset::make("QuantizationInfoOutput", { QuantizationInfo(1.0f, 10) })), - framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(2.0f, -5) }))) + framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(2.0f, -5), QuantizationInfo(1.0f, 43) }))) { // Validate output validate(Accessor(_target), _reference, tolerance_s8); @@ -220,11 +230,21 @@ FIXTURE_DATA_TEST_CASE(RunSmallQASYMM8, NEQuantizationLayerQASYMM8GenFixture<int framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataTypeOut", { DataType::QASYMM8 })), framework::dataset::make("QuantizationInfoOutput", { QuantizationInfo(2.0f, 10), QuantizationInfo(2.0f, -25) })), - framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(1.0f, 30) }))) + framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(1.0f, 30), QuantizationInfo(2.0f, -128) }))) { // Validate output validate(Accessor(_target), _reference, tolerance_u8); } +FIXTURE_DATA_TEST_CASE(ConvertInt8toUint8, NEQuantizationLayerQASYMM8_SIGNEDGenFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(combine(QuantizationSmallShapes, + framework::dataset::make("DataTypeIn", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataTypeOut", { DataType::QASYMM8 })), + framework::dataset::make("QuantizationInfoOutput", { QuantizationInfo(1.0f, 0) })), + framework::dataset::make("QuantizationInfoInput", { QuantizationInfo(1.0f, -128) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_s8); +} + TEST_SUITE_END() // QASYMM8_SIGNED TEST_SUITE_END() // Quantized diff --git a/tests/validation/NEON/RNNLayer.cpp b/tests/validation/NEON/RNNLayer.cpp index 14d9a5d14e..979aa0f2c5 100644 --- a/tests/validation/NEON/RNNLayer.cpp +++ b/tests/validation/NEON/RNNLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -139,7 +139,7 @@ TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, NERNNLayerFixture<half>, framework::DatasetMode::ALL, combine(datasets::SmallRNNLayerDataset(), framework::dataset::make("DataType", DataType::F16))) { // Validate output - validate(Accessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); + validate(Accessor(_target), _reference, tolerance_f16, 0.02f, abs_tolerance_f16); } TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ diff --git a/tests/validation/NEON/ReduceMean.cpp b/tests/validation/NEON/ReduceMean.cpp index b4a3f0d399..8ca0bb53a7 100644 --- a/tests/validation/NEON/ReduceMean.cpp +++ b/tests/validation/NEON/ReduceMean.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,10 +46,15 @@ constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value f #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC constexpr AbsoluteTolerance<float> tolerance_f16(0.03f); /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */ #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +#ifdef __aarch64__ constexpr AbsoluteTolerance<uint8_t> tolerance_u8(1); /**< Tolerance value for comparing reference's output against implementation's output for unsigned 8-bit asymmetric quantized type */ +constexpr AbsoluteTolerance<int8_t> tolerance_s8(1); /**< Tolerance value for comparing reference's output against implementation's output for signed 8-bit asymmetric quantized type */ +#else // __aarch64__ +constexpr AbsoluteTolerance<uint8_t> tolerance_u8(2); /**< Tolerance value for comparing reference's output against implementation's output for unsigned 8-bit asymmetric quantized type */ constexpr AbsoluteTolerance<int8_t> tolerance_s8(2); /**< Tolerance value for comparing reference's output against implementation's output for signed 8-bit asymmetric quantized type */ +#endif // __aarch64__ -const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(0, 1, 2, 3) }), +const auto axis_keep = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1, 0), Coordinates(1, 2), Coordinates(0, 2), Coordinates(1, 3), Coordinates(2, 3), Coordinates(0, 1, 2, 3) }), framework::dataset::make("KeepDims", { true })); const auto axis_drop = combine(framework::dataset::make("Axis", { Coordinates(0), Coordinates(1), Coordinates(3) }), framework::dataset::make("KeepDims", { false })); } // namespace diff --git a/tests/validation/NEON/Remap.cpp b/tests/validation/NEON/Remap.cpp deleted file mode 100644 index 3c02f8eece..0000000000 --- a/tests/validation/NEON/Remap.cpp +++ /dev/null @@ -1,77 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "arm_compute/core/Types.h" -#include "arm_compute/runtime/NEON/functions/NERemap.h" -#include "arm_compute/runtime/Tensor.h" -#include "arm_compute/runtime/TensorAllocator.h" -#include "tests/NEON/Accessor.h" -#include "tests/PaddingCalculator.h" -#include "tests/datasets/BorderModeDataset.h" -#include "tests/datasets/ShapeDatasets.h" -#include "tests/framework/Asserts.h" -#include "tests/framework/Macros.h" -#include "tests/framework/datasets/Datasets.h" -#include "tests/validation/Validation.h" -#include "tests/validation/fixtures/RemapFixture.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -namespace -{ -constexpr AbsoluteTolerance<uint8_t> tolerance_value(0); -constexpr float tolerance_number = 0.f; -} // namespace - -TEST_SUITE(NEON) -TEST_SUITE(Remap) - -template <typename T> -using NERemapFixture = RemapValidationFixture<Tensor, Accessor, NERemap, T>; - -FIXTURE_DATA_TEST_CASE(RunSmall, NERemapFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("InterpolationPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR, InterpolationPolicy::BILINEAR })), - framework::dataset::make("DataType", - DataType::U8)), - framework::dataset::make("BorderModes", { BorderMode::UNDEFINED, BorderMode::CONSTANT }))) -{ - // Validate output - validate(Accessor(_target), _reference, _valid_mask, tolerance_value, tolerance_number); -} - -FIXTURE_DATA_TEST_CASE(RunLarge, NERemapFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("InterpolationPolicy", { InterpolationPolicy::NEAREST_NEIGHBOR, InterpolationPolicy::BILINEAR })), - framework::dataset::make("DataType", - DataType::U8)), - framework::dataset::make("BorderModes", { BorderMode::UNDEFINED, BorderMode::CONSTANT }))) -{ - // Validate output - validate(Accessor(_target), _reference, _valid_mask, tolerance_value, tolerance_number); -} -TEST_SUITE_END() -TEST_SUITE_END() -} // namespace validation -} // namespace test -} // namespace arm_compute diff --git a/tests/validation/NEON/ReorderLayer.cpp b/tests/validation/NEON/ReorderLayer.cpp new file mode 100644 index 0000000000..839ad0ac92 --- /dev/null +++ b/tests/validation/NEON/ReorderLayer.cpp @@ -0,0 +1,107 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#if defined(__aarch64__) + +#include "arm_compute/runtime/NEON/functions/NEReorderLayer.h" +#include "arm_compute/runtime/Tensor.h" +#include "tests/NEON/Accessor.h" +#include "tests/datasets/ReorderLayerDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/ReorderFixture.h" +#include "src/core/NEON/kernels/NEReorderKernel.h" +#include "src/core/NEON/kernels/arm_gemm/utils.hpp" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using framework::dataset::make; + +TEST_SUITE(NEON) +TEST_SUITE(ReorderLayer) + +template <typename T> +using NEReorderLayerAlias = ReorderValidationFixture<Tensor, Accessor, NEReorderLayer, T>; + +TEST_SUITE(FP32) +#if defined(ARM_COMPUTE_ENABLE_SVE) +DATA_TEST_CASE(ValidateReorderOHWIo8, framework::DatasetMode::ALL, combine( + zip( + make("InShape",{ TensorShape(10U, 9U), TensorShape(234U, 301U) }), + make("OutShape", { TensorShape(10U, 16U), TensorShape(234U, 304U) }) + ), + zip( + make("InputWeightFormat", {WeightFormat::OHWI}), + make("OutputWeightFormat", {WeightFormat::OHWIo8}) + )), + input_shape, output_shape, input_wf, output_wf) +{ + if(Scheduler::get().cpu_info().has_sve()){ + arm_compute::NEReorderLayer reorder_layer; + int vector_length = arm_gemm::utils::get_vector_length<float>(); + bool expected_bool_status = false; + if (vector_length == 8) + { + expected_bool_status = true; + } + + TensorInfo input_tensor_info(input_shape, 1, DataType::F32); + TensorInfo output_tensor_info(output_shape, 1, DataType::F32); + + Status status = reorder_layer.validate(&input_tensor_info, &output_tensor_info, input_wf, output_wf); + + ARM_COMPUTE_EXPECT((expected_bool_status == bool(status)), framework::LogLevel::ERRORS); + } +} + +FIXTURE_DATA_TEST_CASE(RunBlock8, NEReorderLayerAlias<float>, framework::DatasetMode::ALL, combine(datasets::ReorderLayerDatasetBlock8(), make("DataType", DataType::F32))) +{ + // Validate output + if (_hardware_supports) + { + validate(Accessor(_target), _reference); + } +} +#endif // ARM_COMPUTE_ENABLE_SVE + +FIXTURE_DATA_TEST_CASE(RunBlock4, NEReorderLayerAlias<float>, framework::DatasetMode::ALL, combine(datasets::ReorderLayerDatasetBlock4(), make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +TEST_SUITE_END() // FP32 + +TEST_SUITE_END() // ReorderLayer +TEST_SUITE_END() // NEON +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // defined(__aarch64__) diff --git a/tests/validation/NEON/ReshapeLayer.cpp b/tests/validation/NEON/ReshapeLayer.cpp index bf39c399a5..e9f114d491 100644 --- a/tests/validation/NEON/ReshapeLayer.cpp +++ b/tests/validation/NEON/ReshapeLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 Arm Limited. + * Copyright (c) 2017-2018, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -77,6 +77,9 @@ input_info, output_info, expected) template <typename T> using NEReshapeLayerFixture = ReshapeLayerValidationFixture<Tensor, Accessor, NEReshapeLayer, T>; +template <typename T> +using NEReshapeLayerPaddedFixture = ReshapeLayerPaddedValidationFixture<Tensor, Accessor, NEReshapeLayer, T>; + TEST_SUITE(Float) TEST_SUITE(F32) FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallReshapeLayerDataset(), framework::dataset::make("DataType", DataType::F32))) @@ -84,8 +87,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerFixture<float>, framework::Datase // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() -TEST_SUITE_END() +TEST_SUITE_END() //F32 +TEST_SUITE_END() //Float TEST_SUITE(Integer) TEST_SUITE(S8) @@ -94,7 +97,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerFixture<int8_t>, framework::Datas // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() +TEST_SUITE_END() //S8 TEST_SUITE(S16) FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallReshapeLayerDataset(), framework::dataset::make("DataType", DataType::S16))) @@ -102,11 +105,41 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerFixture<int16_t>, framework::Data // Validate output validate(Accessor(_target), _reference); } -TEST_SUITE_END() -TEST_SUITE_END() +TEST_SUITE_END() //S16 +TEST_SUITE_END() //Integer + +TEST_SUITE(Padded) +TEST_SUITE(Float) +TEST_SUITE(F32) +FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerPaddedFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallReshapeLayerDataset(), framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() //S32 +TEST_SUITE_END() //Float + +TEST_SUITE(Integer) +TEST_SUITE(S8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerPaddedFixture<int8_t>, framework::DatasetMode::ALL, combine(datasets::SmallReshapeLayerDataset(), framework::dataset::make("DataType", DataType::S8))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() //S8 + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEReshapeLayerPaddedFixture<int16_t>, framework::DatasetMode::ALL, combine(datasets::SmallReshapeLayerDataset(), framework::dataset::make("DataType", DataType::S16))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() //S16 +TEST_SUITE_END() //Integer +TEST_SUITE_END() //Padded -TEST_SUITE_END() -TEST_SUITE_END() +TEST_SUITE_END() //ReshapeLayer +TEST_SUITE_END() //NEON } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/NEON/Reverse.cpp b/tests/validation/NEON/Reverse.cpp index 3dc3eeee80..7b5337f14b 100644 --- a/tests/validation/NEON/Reverse.cpp +++ b/tests/validation/NEON/Reverse.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,8 @@ namespace validation { namespace { -auto run_small_dataset = combine(datasets::SmallShapes(), datasets::Tiny1DShapes()); +using framework::dataset::make; +auto run_small_dataset = combine(datasets::Small3DShapes(), datasets::Tiny1DShapes()); auto run_large_dataset = combine(datasets::LargeShapes(), datasets::Tiny1DShapes()); } // namespace @@ -53,28 +54,31 @@ TEST_SUITE(Reverse) // *INDENT-OFF* // clang-format off DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), // Invalid axis datatype + make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), // Invalid axis datatype TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid axis shape TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Invalid axis length (> 4) TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Mismatching shapes + TensorInfo(TensorShape(32U, 13U, 17U, 3U, 2U), 1, DataType::U8), // Unsupported source dimensions (>4) TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(2U), 1, DataType::U8), }), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), + make("OutputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(2U, 13U, 2U), 1, DataType::U8), + TensorInfo(TensorShape(32U, 13U, 17U, 3U, 2U), 1, DataType::U8), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), TensorInfo(TensorShape(2U), 1, DataType::U8), })), - framework::dataset::make("AxisInfo", { TensorInfo(TensorShape(3U), 1, DataType::U8), + make("AxisInfo", { TensorInfo(TensorShape(3U), 1, DataType::U8), TensorInfo(TensorShape(2U, 10U), 1, DataType::U32), TensorInfo(TensorShape(8U), 1, DataType::U32), TensorInfo(TensorShape(2U), 1, DataType::U32), TensorInfo(TensorShape(2U), 1, DataType::U32), TensorInfo(TensorShape(2U), 1, DataType::U32), + TensorInfo(TensorShape(2U), 1, DataType::U32), })), - framework::dataset::make("Expected", { false, false, false, false, true, true})), + make("Expected", { false, false, false, false, false, true, true})), src_info, dst_info, axis_info, expected) { Status s = NEReverse::validate(&src_info.clone()->set_is_resizable(false), @@ -95,7 +99,11 @@ TEST_SUITE(F16) FIXTURE_DATA_TEST_CASE(RunSmall, NEReverseFixture<half>, framework::DatasetMode::PRECOMMIT, - combine(run_small_dataset, framework::dataset::make("DataType", DataType::F16))) + combine( + run_small_dataset, + make("DataType", DataType::F16), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(Accessor(_target), _reference); @@ -104,7 +112,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, FIXTURE_DATA_TEST_CASE(RunLarge, NEReverseFixture<half>, framework::DatasetMode::NIGHTLY, - combine(run_large_dataset, framework::dataset::make("DataType", DataType::F16))) + combine( + run_large_dataset, + make("DataType", DataType::F16), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(Accessor(_target), _reference); @@ -116,7 +128,11 @@ TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEReverseFixture<float>, framework::DatasetMode::PRECOMMIT, - combine(run_small_dataset, framework::dataset::make("DataType", DataType::F32))) + combine( + run_small_dataset, + make("DataType", DataType::F32), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(Accessor(_target), _reference); @@ -125,7 +141,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, FIXTURE_DATA_TEST_CASE(RunLarge, NEReverseFixture<float>, framework::DatasetMode::NIGHTLY, - combine(run_large_dataset, framework::dataset::make("DataType", DataType::F32))) + combine( + run_large_dataset, + make("DataType", DataType::F32), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(Accessor(_target), _reference); @@ -138,7 +158,11 @@ TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEReverseFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, - combine(run_small_dataset, framework::dataset::make("DataType", DataType::QASYMM8))) + combine( + run_small_dataset, + make("DataType", DataType::QASYMM8), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(Accessor(_target), _reference); @@ -147,7 +171,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, FIXTURE_DATA_TEST_CASE(RunLarge, NEReverseFixture<uint8_t>, framework::DatasetMode::NIGHTLY, - combine(run_large_dataset, framework::dataset::make("DataType", DataType::QASYMM8))) + combine( + run_large_dataset, + make("DataType", DataType::QASYMM8), + make("use_negative_axis", { true, false }), + make("use_inverted_axis", { true, false }))) { // Validate output validate(Accessor(_target), _reference); diff --git a/tests/validation/NEON/Scale.cpp b/tests/validation/NEON/Scale.cpp index 64427ae34f..f1209a21ac 100644 --- a/tests/validation/NEON/Scale.cpp +++ b/tests/validation/NEON/Scale.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,16 +22,10 @@ * SOFTWARE. */ #include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/functions/NEScale.h" -#include "arm_compute/runtime/Tensor.h" -#include "arm_compute/runtime/TensorAllocator.h" #include "tests/NEON/Accessor.h" -#include "tests/PaddingCalculator.h" #include "tests/datasets/ScaleValidationDataset.h" -#include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" -#include "tests/validation/Helpers.h" #include "tests/validation/Validation.h" #include "tests/validation/fixtures/ScaleFixture.h" @@ -51,7 +45,7 @@ using datasets::ScaleAlignCornersSamplingPolicySet; /** We consider vector size in byte 64 since the maximum size of * a vector used by the kernel is currently 64-byte (float32x4x4). - * There are possibility to reduce test time further by using + * There is possibility to reduce test time further by using * smaller vector sizes for different data types where applicable. */ constexpr uint32_t vector_byte = 64; @@ -62,25 +56,31 @@ constexpr uint32_t num_elements_per_vector() return vector_byte / sizeof(T); } -/** Scale data types */ -const auto ScaleDataTypes = framework::dataset::make("DataType", +/** Quantization information data set */ +const auto QuantizationInfoSet = framework::dataset::make("QuantizationInfo", { - DataType::U8, - DataType::S16, - DataType::F32, + QuantizationInfo(0.5f, -10), }); /** Quantization information data set */ -const auto QuantizationInfoSet = framework::dataset::make("QuantizationInfo", +const auto InputQuantizationInfoSet = framework::dataset::make("InputQuantizationInfo", { QuantizationInfo(0.5f, -10), }); +/** Quantization information data set */ +const auto OutputQuantizationInfoSet = framework::dataset::make("OutputQuantizationInfo", +{ + QuantizationInfo(0.2f, 20), +}); + /** Tolerance */ constexpr AbsoluteTolerance<uint8_t> tolerance_u8(1); +constexpr AbsoluteTolerance<int8_t> tolerance_s8(1); constexpr AbsoluteTolerance<int16_t> tolerance_s16(1); RelativeTolerance<float> tolerance_f32(0.05); #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC +constexpr float abs_tolerance_f16(0.01f); RelativeTolerance<half> tolerance_f16(half(0.1)); #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ @@ -155,8 +155,6 @@ TEST_CASE(SupportDataType, framework::DatasetMode::ALL) { DataType::BFLOAT16, false }, #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC { DataType::F16, true }, -#else // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC - { DataType::F16, false }, #endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC { DataType::F32, true }, { DataType::F64, false }, @@ -320,11 +318,14 @@ using NEScaleMixedDataLayoutFixture = ScaleValidationFixture<Tensor, Accessor, N template <typename T> using NEScaleQuantizedFixture = ScaleValidationQuantizedFixture<Tensor, Accessor, NEScale, T>; template <typename T> +using NEScaleDifferentOutputQuantizedFixture = ScaleValidationDifferentOutputQuantizedFixture<Tensor, Accessor, NEScale, T>; +template <typename T> using NEScaleQuantizedMixedDataLayoutFixture = ScaleValidationQuantizedFixture<Tensor, Accessor, NEScale, T, true>; TEST_SUITE(Float) TEST_SUITE(FP32) -const auto f32_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<float>())), framework::dataset::make("DataType", DataType::F32)); +const auto f32_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<float>())), framework::dataset::make("DataType", DataType::F32)); +const auto f32_shape_nhwc = combine(datasets::Small3DShapes(), framework::dataset::make("DataType", DataType::F32)); FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<float>, framework::DatasetMode::ALL, ASSEMBLE_DATASET(f32_shape, ScaleSamplingPolicySet)) { //Create valid region @@ -352,10 +353,38 @@ FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<float>, framework::D // Validate output validate(Accessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32); } +FIXTURE_DATA_TEST_CASE(RunMediumNHWC, NEScaleFixture<float>, framework::DatasetMode::ALL, ASSEMBLE_NHWC_DATASET(f32_shape_nhwc, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32); +} +FIXTURE_DATA_TEST_CASE(RunMediumMixedDataLayoutNHWC, NEScaleMixedDataLayoutFixture<float>, framework::DatasetMode::PRECOMMIT, ASSEMBLE_NHWC_DATASET(f32_shape_nhwc, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32); +} +FIXTURE_DATA_TEST_CASE(RunMediumAlignCornersNHWC, NEScaleFixture<float>, framework::DatasetMode::ALL, ASSEMBLE_NHWC_DATASET(f32_shape_nhwc, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32); +} TEST_SUITE_END() // FP32 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -const auto f16_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<half>())), framework::dataset::make("DataType", DataType::F16)); +const auto f16_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<half>())), framework::dataset::make("DataType", DataType::F16)); +const auto f16_shape_nhwc = combine(datasets::Small3DShapes(), framework::dataset::make("DataType", DataType::F16)); FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<half>, framework::DatasetMode::ALL, ASSEMBLE_DATASET(f16_shape, ScaleSamplingPolicySet)) { //Create valid region @@ -363,7 +392,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<half>, framework::DatasetMode::A const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); // Validate output - validate(Accessor(_target), _reference, valid_region, tolerance_f16); + validate(Accessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); } FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<half>, framework::DatasetMode::ALL, ASSEMBLE_DATASET(f16_shape, ScaleAlignCornersSamplingPolicySet)) { @@ -372,7 +401,34 @@ FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<half>, framework::Da const ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); // Validate output - validate(Accessor(_target), _reference, valid_region, tolerance_f16); + validate(Accessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunMediumNHWC, NEScaleFixture<half>, framework::DatasetMode::ALL, ASSEMBLE_NHWC_DATASET(f16_shape_nhwc, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunMediumMixedDataLayoutNHWC, NEScaleMixedDataLayoutFixture<half>, framework::DatasetMode::PRECOMMIT, ASSEMBLE_NHWC_DATASET(f16_shape_nhwc, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunMediumAlignCornersNHWC, NEScaleFixture<half>, framework::DatasetMode::ALL, ASSEMBLE_NHWC_DATASET(f16_shape_nhwc, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); } TEST_SUITE_END() // FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ @@ -400,6 +456,27 @@ FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<uint8_t>, framework: validate(Accessor(_target), _reference, valid_region, tolerance_u8); } TEST_SUITE_END() // U8 +TEST_SUITE(S8) +const auto s8_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<int8_t>())), framework::dataset::make("DataType", DataType::S8)); +FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<int8_t>, framework::DatasetMode::ALL, ASSEMBLE_S8_DATASET(s8_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_s8); +} +FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleFixture<int8_t>, framework::DatasetMode::ALL, ASSEMBLE_S8_DATASET(s8_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_s8); +} +TEST_SUITE_END() // S8 TEST_SUITE(S16) const auto s16_shape = combine((SCALE_SHAPE_DATASET(num_elements_per_vector<int16_t>())), framework::dataset::make("DataType", DataType::S16)); FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleFixture<int16_t>, framework::DatasetMode::ALL, ASSEMBLE_DATASET(s16_shape, ScaleSamplingPolicySet)) @@ -435,7 +512,18 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleQuantizedFixture<uint8_t>, framework::Da // Validate output validate(Accessor(_target), _reference, valid_region, tolerance_u8); } -FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEScaleQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET(qasymm8_shape, ScaleSamplingPolicySet, QuantizationInfoSet)) +FIXTURE_DATA_TEST_CASE(RunSmallDifferentOutputQuantization, NEScaleDifferentOutputQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, + ASSEMBLE_DIFFERENTLY_QUANTIZED_DATASET(qasymm8_shape, ScaleSamplingPolicySet, InputQuantizationInfoSet, OutputQuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_u8); +} +FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEScaleQuantizedMixedDataLayoutFixture<uint8_t>, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET(qasymm8_shape, ScaleSamplingPolicySet, + QuantizationInfoSet)) { //Create valid region TensorInfo src_info(_shape, 1, _data_type); @@ -467,6 +555,16 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEScaleQuantizedFixture<int8_t>, framework::Dat // Validate output validate(Accessor(_target), _reference, valid_region, tolerance_qasymm8_signed); } +FIXTURE_DATA_TEST_CASE(RunSmallDifferentOutputQuantization, NEScaleDifferentOutputQuantizedFixture<int8_t>, framework::DatasetMode::ALL, + ASSEMBLE_DIFFERENTLY_QUANTIZED_DATASET(qasymm8_signed_shape, ScaleSamplingPolicySet, InputQuantizationInfoSet, OutputQuantizationInfoSet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + ValidRegion valid_region = calculate_valid_region_scale(src_info, _reference.shape(), _policy, _sampling_policy, (_border_mode == BorderMode::UNDEFINED)); + + // Validate output + validate(Accessor(_target), _reference, valid_region, tolerance_qasymm8_signed); +} FIXTURE_DATA_TEST_CASE(RunSmallAlignCorners, NEScaleQuantizedFixture<int8_t>, framework::DatasetMode::ALL, ASSEMBLE_QUANTIZED_DATASET(qasymm8_signed_shape, ScaleAlignCornersSamplingPolicySet, QuantizationInfoSet)) { diff --git a/tests/validation/NEON/SoftmaxLayer.cpp b/tests/validation/NEON/SoftmaxLayer.cpp index 2a9e30604e..94d0866c38 100644 --- a/tests/validation/NEON/SoftmaxLayer.cpp +++ b/tests/validation/NEON/SoftmaxLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,21 +25,22 @@ #include "arm_compute/runtime/NEON/functions/NESoftmaxLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/common/cpuinfo/CpuIsaInfo.h" +#include "src/cpu/kernels/CpuSoftmaxKernel.h" #include "tests/NEON/Accessor.h" -#include "tests/PaddingCalculator.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" #include "tests/validation/Validation.h" #include "tests/validation/fixtures/SoftmaxLayerFixture.h" - namespace arm_compute { namespace test { namespace validation { +using framework::dataset::make; namespace { /** Tolerance for float operations */ @@ -51,7 +52,7 @@ constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); constexpr AbsoluteTolerance<int8_t> tolerance_qasymm8_signed(1); /** CNN data types */ -const auto CNNDataTypes = framework::dataset::make("DataType", +const auto CNNDataTypes = make("DataType", { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC DataType::F16, @@ -62,56 +63,55 @@ const auto CNNDataTypes = framework::dataset::make("DataType", TEST_SUITE(NEON) TEST_SUITE(SoftmaxLayer) - // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching data types - TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching shapes - TensorInfo(TensorShape(27U, 13U), 1, DataType::QASYMM8, // Invalid output quantization info - QuantizationInfo(1.f/256, 12)), - TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, - QuantizationInfo(1.f/256, 12)), - TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, //Invalid axis high - QuantizationInfo(1.f/256, 12)), - TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, //Invalid axis low - QuantizationInfo(1.f/256, 12)), - }), - framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U), 1, DataType::F16), - TensorInfo(TensorShape(27U, 11U), 1, DataType::F32), - TensorInfo(TensorShape(27U, 13U), 1, DataType::QASYMM8, - QuantizationInfo(1.f/256, 12)), - TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, - QuantizationInfo(1.f/256, 0)), - TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), - TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, - QuantizationInfo(1.f/256, 0)), - TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, - QuantizationInfo(1.f/256, 0)), - })), - framework::dataset::make("beta", { 1.0, - 2.0, - 1.0, - 2.0, - 1.0, - 1.0, - 2.0, - 1.0, - })), - framework::dataset::make("axis", { 0, - 0, - 0, - 1, - 0, - -1, - 2, - -3, - })), - framework::dataset::make("Expected", { false, false, false, true, true, true, false, false })), - input_info, output_info, beta, axis, expected) +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + make("InputInfo", { TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching data types + TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(27U, 13U), 1, DataType::QASYMM8, // Invalid output quantization info + QuantizationInfo(1.f/256, 12)), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, + QuantizationInfo(1.f/256, 12)), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, //Invalid axis high + QuantizationInfo(1.f/256, 12)), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, //Invalid axis low + QuantizationInfo(1.f/256, 12)), + }), + make("OutputInfo",{ TensorInfo(TensorShape(27U, 13U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 11U), 1, DataType::F32), + TensorInfo(TensorShape(27U, 13U), 1, DataType::QASYMM8, + QuantizationInfo(1.f/256, 12)), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, + QuantizationInfo(1.f/256, 0)), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, + QuantizationInfo(1.f/256, 0)), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM8, + QuantizationInfo(1.f/256, 0)), + }), + make("beta", { 1.0, + 2.0, + 1.0, + 2.0, + 1.0, + 1.0, + 2.0, + 1.0, + }), + make("axis", { 0, + 0, + 0, + 1, + 0, + -1, + 2, + -3, + }), + make("Expected", { false, false, false, true, true, true, false, false })), + input_info, output_info, beta, axis, expected) { ARM_COMPUTE_EXPECT(bool(NESoftmaxLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), beta, axis)) == expected, framework::LogLevel::ERRORS); } @@ -121,29 +121,80 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( template <typename T> using NESoftmaxLayerFixture = SoftmaxValidationFixture<Tensor, Accessor, NESoftmaxLayer, T>; +DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, + concat( + combine( + make("CpuExt", std::string("neon")), + make("DataType", { DataType::F32, + DataType::F16, + DataType::QASYMM8, + DataType::QASYMM8_SIGNED}) + ), + combine( + make("CpuExt", std::string("sme2")), + make("DataType", { DataType::F32, + DataType::F16})) + ), + cpu_ext, data_type) +{ + using namespace cpu::kernels; + + cpuinfo::CpuIsaInfo cpu_isa{}; + cpu_isa.neon = (cpu_ext == "neon"); + cpu_isa.sme2 = (cpu_ext == "sme2"); + cpu_isa.fp16 = (data_type == DataType::F16); + + const auto *selected_impl = CpuSoftmaxKernel::get_implementation( + SoftmaxKernelDataTypeISASelectorData{ data_type, cpu_isa, false /* is_log */, 0 /* axis */, CPUInfo::get().get_sme2_vector_length()}, + cpu::KernelSelectionType::Preferred); + + ARM_COMPUTE_ERROR_ON_NULLPTR(selected_impl); + + std::string expected = cpu_ext + "_" + cpu_impl_dt(data_type) + "_softmax"; + std::string actual = selected_impl->name; + + ARM_COMPUTE_EXPECT_EQUAL(expected, actual, framework::LogLevel::ERRORS); +} + TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small4DShapes(), - framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("Beta", { 1.0f, 2.0f })), - framework::dataset::make("Axis", { 0, 1 }))) +FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine( + datasets::SoftmaxLayerSmallShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0, -1 }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine( + datasets::SmallShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0, 1 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small4DShapes(), - framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("Beta", { 1.0f, 2.0f })), - framework::dataset::make("Axis", { 0, 2, -1 }))) +FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerFixture<half>, framework::DatasetMode::PRECOMMIT, + combine( + datasets::Small4DShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f }), + make("Axis", { 0, 2, -1 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SoftmaxLayerLargeShapes(), - framework::dataset::make("DataType", DataType::F16)), - framework::dataset::make("Beta", { 1.0f, 2.0f })), - framework::dataset::make("Axis", { 0 }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<half>, framework::DatasetMode::NIGHTLY, + combine( + datasets::SoftmaxLayerLargeShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); @@ -152,26 +203,30 @@ TEST_SUITE_END() //FP16 #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("Beta", { 1.0f, 2.0f })), - framework::dataset::make("Axis", { 0, -1 }))) +FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine( + datasets::SoftmaxLayerSmallShapes(), + make("DataType", DataType::F32), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0, -1 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::Small4DShapes(), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("Beta", { 1.0f, 2.0f })), - framework::dataset::make("Axis", { 0, -2, 3 }))) +FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerFixture<float>, framework::DatasetMode::PRECOMMIT, + combine(datasets::Small4DShapes(), + make("DataType", DataType::F32), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0, -2, 3 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SoftmaxLayerLargeShapes(), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("Beta", { 1.0f, 2.0f })), - framework::dataset::make("Axis", { 0 }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixture<float>, framework::DatasetMode::NIGHTLY, + combine(datasets::SoftmaxLayerLargeShapes(), + make("DataType", DataType::F32), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); @@ -184,29 +239,40 @@ using NESoftmaxLayerQuantizedFixture = SoftmaxValidationQuantizedFixture<Tensor, TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(), - framework::dataset::make("DataType", DataType::QASYMM8)), - combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), - framework::dataset::make("Beta", { 1.0f, 2.f }))), - framework::dataset::make("Axis", { 0, -1 }))) +FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, + combine( + datasets::SoftmaxLayerSmallShapes(), + make("DataType", DataType::QASYMM8), + combine( + make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + make("Beta", { 1.0f, 2.f }) + ), + make("Axis", { 0, -1 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::Small4DShapes(), - framework::dataset::make("DataType", DataType::QASYMM8)), - combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), - framework::dataset::make("Beta", { 1.0f, 2.f }))), - framework::dataset::make("Axis", { 0, 1, -2 }))) +FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, + combine( + datasets::Small4DShapes(), + make("DataType", DataType::QASYMM8), + combine( + make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + make("Beta", { 1.0f, 2.f })), + make("Axis", { 0, 1, -2 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SoftmaxLayerLargeShapes(), - framework::dataset::make("DataType", DataType::QASYMM8)), - combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), - framework::dataset::make("Beta", { 1.0f, 2.0f }))), - framework::dataset::make("Axis", { 0 }))) +FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, + combine( + datasets::SoftmaxLayerLargeShapes(), + make("DataType", DataType::QASYMM8), + combine( + make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + make("Beta", { 1.0f, 2.0f }) + ), + make("Axis", { 0 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -214,20 +280,28 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerQuantizedFixture<uint8_t>, framew TEST_SUITE_END() //QASYMM8 TEST_SUITE(QASYMM8_SIGNED) -FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::SoftmaxLayerSmallShapes(), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), - framework::dataset::make("Beta", { 1.0f, 2.f }))), - framework::dataset::make("Axis", { 0, -1 }))) +FIXTURE_DATA_TEST_CASE(RunSmall2D, NESoftmaxLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, + combine( + datasets::SoftmaxLayerSmallShapes(), + make("DataType", DataType::QASYMM8_SIGNED), + combine( + make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + make("Beta", { 1.0f, 2.f }) + ), + make("Axis", { 0, -1 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_signed); } -FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::Small4DShapes(), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), - framework::dataset::make("Beta", { 1.0f, 2.f }))), - framework::dataset::make("Axis", { 0, 1, -1 }))) +FIXTURE_DATA_TEST_CASE(RunSmall4D, NESoftmaxLayerQuantizedFixture<int8_t>, framework::DatasetMode::ALL, + combine( + datasets::Small4DShapes(), + make("DataType", DataType::QASYMM8_SIGNED), + combine( + make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + make("Beta", { 1.0f, 2.f }) + ), + make("Axis", { 0, 1, -1 }))) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8_signed); diff --git a/tests/validation/NEON/StackLayer.cpp b/tests/validation/NEON/StackLayer.cpp index d88f713ccd..3828010c7b 100644 --- a/tests/validation/NEON/StackLayer.cpp +++ b/tests/validation/NEON/StackLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,69 +44,74 @@ namespace test { namespace validation { + +using framework::dataset::make; namespace { // *INDENT-OFF* // clang-format off /** Data types */ -const auto data_types = framework::dataset::make("DataType", { DataType::QASYMM8, DataType::F16, DataType::F32 }); +const auto data_types = make("DataType", { DataType::QASYMM8, DataType::F16, DataType::F32 }); /** Num tensors values to test */ -const auto n_values = framework::dataset::make("NumTensors", { 3, 4 }); +const auto n_values = make("NumTensors", { 3, 4 }); /** Shapes 1D to test */ -const auto shapes_1d_small = combine(datasets::Small1DShapes(), framework::dataset::make("Axis", -1, 2)); +const auto shapes_1d_small = combine(datasets::Small1DShapes(), make("Axis", -1, 2)); /** Shapes 2D to test */ -const auto shapes_2d_small = combine(datasets::Small2DShapes(), framework::dataset::make("Axis", -2, 3)); +const auto shapes_2d_small = combine(datasets::Small2DShapes(), make("Axis", -2, 3)); /** Shapes 3D to test */ -const auto shapes_3d_small = combine(datasets::Small3DShapes(), framework::dataset::make("Axis", -3, 4)); +const auto shapes_3d_small = combine(datasets::Small3DShapes(), make("Axis", -3, 4)); /** Shapes 4D to test */ -const auto shapes_4d_small = combine(datasets::Small4DShapes(), framework::dataset::make("Axis", -4, 5)); +const auto shapes_4d_small = combine(datasets::Small4DShapes(), make("Axis", -4, 5)); /** Shapes 1D to test */ -const auto shapes_1d_large = combine(datasets::Large1DShapes(), framework::dataset::make("Axis", -1, 2)); +const auto shapes_1d_large = combine(datasets::Large1DShapes(), make("Axis", -1, 2)); /** Shapes 2D to test */ -const auto shapes_2d_large = combine(datasets::Medium2DShapes(), framework::dataset::make("Axis", -2, 3)); +const auto shapes_2d_large = combine(datasets::Medium2DShapes(), make("Axis", -2, 3)); /** Shapes 3D to test */ -const auto shapes_3d_large = combine(datasets::Medium3DShapes(), framework::dataset::make("Axis", -3, 4)); +const auto shapes_3d_large = combine(datasets::Medium3DShapes(), make("Axis", -3, 4)); /** Shapes 4D to test */ -const auto shapes_4d_large = combine(datasets::Medium4DShapes(), framework::dataset::make("Axis", -4, 5)); +const auto shapes_4d_large = combine(datasets::Medium4DShapes(), make("Axis", -4, 5)); } // namespace /** Fixture to use */ template<typename T> using NEStackLayerFixture = StackLayerValidationFixture<Tensor, ITensor, Accessor, NEStackLayer, T>; +template<typename T> +using NEStackLayerWithPaddingFixture = StackLayerWithPaddingValidationFixture<Tensor, ITensor, Accessor, NEStackLayer, T>; + using namespace arm_compute::misc::shape_calculator; TEST_SUITE(NEON) TEST_SUITE(StackLayer) -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( - framework::dataset::make("InputInfo", +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( +make("InputInfo", { std::vector<TensorInfo>{ TensorInfo(TensorShape(9U, 8U), 1, DataType::U8) }, - std::vector<TensorInfo>{ TensorInfo(TensorShape(1U, 2U), 1, DataType::U8) , TensorInfo(TensorShape(1U, 2U), 1, DataType::U8), TensorInfo(TensorShape(1U, 2U), 1, DataType::U8)}, + std::vector<TensorInfo>{ TensorInfo(TensorShape(1U, 2U), 1, DataType::U8) , TensorInfo(TensorShape(1U, 2U), 1, DataType::U8), TensorInfo(TensorShape(1U, 2U), 1, DataType::U8)}, std::vector<TensorInfo>{ TensorInfo(TensorShape(2U, 3U), 1, DataType::S32) }, - std::vector<TensorInfo>{ TensorInfo(TensorShape(7U, 5U, 3U, 8U, 2U), 1, DataType::S32), TensorInfo(TensorShape(7U, 5U, 3U, 8U, 2U), 1, DataType::S32)}, + std::vector<TensorInfo>{ TensorInfo(TensorShape(7U, 5U, 3U, 8U, 2U), 1, DataType::S32), TensorInfo(TensorShape(7U, 5U, 3U, 8U, 2U), 1, DataType::S32)}, std::vector<TensorInfo>{ TensorInfo(TensorShape(9U, 8U), 1, DataType::S32) }, }), -framework::dataset::make("OutputInfo", +make("OutputInfo", { TensorInfo(TensorShape(1U, 9U, 8U), 1, DataType::U8), // Passes, stack 1 tensor on x axis TensorInfo(TensorShape(1U, 3U, 2U), 1, DataType::U8), // Passes, stack 3 tensors on y axis TensorInfo(TensorShape(1U, 2U, 3U), 1, DataType::S32), // fails axis < (- input's rank) TensorInfo(TensorShape(3U, 7U, 5U), 1, DataType::S32), // fails, input dimensions > 4 TensorInfo(TensorShape(1U, 2U, 3U), 1, DataType::U8), // fails mismatching data types -})), -framework::dataset::make("Axis", { -3, 1, -4, -3, 1 })), -framework::dataset::make("Expected", { true, true, false, false, false })), +}), +make("Axis", { -3, 1, -4, -3, 1 }), +make("Expected", { true, true, false, false, false })), input_info, output_info, axis, expected) { std::vector<TensorInfo> ti(input_info); @@ -121,18 +126,18 @@ input_info, output_info, axis, expected) TEST_SUITE(Shapes1D) TEST_SUITE(S32) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<int>, framework::DatasetMode::ALL, - combine(combine(shapes_1d_small, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_1d_small, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<int>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_1d_large, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_1d_large, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -141,18 +146,18 @@ TEST_SUITE_END() // S32 TEST_SUITE(S16) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<short>, framework::DatasetMode::ALL, - combine(combine(shapes_1d_small, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_1d_small, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<short>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_1d_large, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_1d_large, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -161,18 +166,18 @@ TEST_SUITE_END() // S16 TEST_SUITE(S8) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<char>, framework::DatasetMode::ALL, - combine(combine(shapes_1d_small, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_1d_small, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<char>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_1d_large, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_1d_large, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -183,18 +188,18 @@ TEST_SUITE_END() // Shapes1D TEST_SUITE(Shapes2D) TEST_SUITE(S32) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<int>, framework::DatasetMode::ALL, - combine(combine(shapes_2d_small, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_2d_small, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<int>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_2d_large, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_2d_large, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -203,18 +208,18 @@ TEST_SUITE_END() // S32 TEST_SUITE(S16) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<short>, framework::DatasetMode::ALL, - combine(combine(shapes_2d_small, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_2d_small, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<short>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_2d_large, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_2d_large, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -223,18 +228,18 @@ TEST_SUITE_END() // S16 TEST_SUITE(S8) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<char>, framework::DatasetMode::ALL, - combine(combine(shapes_2d_small, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_2d_small, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<char>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_2d_large, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_2d_large, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -245,18 +250,18 @@ TEST_SUITE_END() // Shapes2D TEST_SUITE(Shapes3D) TEST_SUITE(S32) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<int>, framework::DatasetMode::ALL, - combine(combine(shapes_3d_small, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_3d_small, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<int>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_3d_large, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_3d_large, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -265,18 +270,18 @@ TEST_SUITE_END() // S32 TEST_SUITE(S16) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<short>, framework::DatasetMode::ALL, - combine(combine(shapes_3d_small, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_3d_small, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<short>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_3d_large, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_3d_large, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -285,18 +290,18 @@ TEST_SUITE_END() // S16 TEST_SUITE(S8) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<char>, framework::DatasetMode::ALL, - combine(combine(shapes_3d_small, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_3d_small, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<char>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_3d_large, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_3d_large, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -307,18 +312,29 @@ TEST_SUITE_END() // Shapes3D TEST_SUITE(Shapes4D) TEST_SUITE(S32) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<int>, framework::DatasetMode::ALL, - combine(combine(shapes_4d_small, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_4d_small, + make("DataType", { DataType::S32 }), + n_values)) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +// Testing the case with padding for only 4d shapes and for one data type. This is because the underlying code +// path depends only on the padding, which isn't affected by the shapes or data types. +FIXTURE_DATA_TEST_CASE(RunSmallWithPadding, NEStackLayerWithPaddingFixture<int>, framework::DatasetMode::ALL, + combine(shapes_4d_small, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<int>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_4d_large, - framework::dataset::make("DataType", { DataType::S32 })), - n_values)) + combine(shapes_4d_large, + make("DataType", { DataType::S32 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -327,18 +343,18 @@ TEST_SUITE_END() // S32 TEST_SUITE(S16) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<short>, framework::DatasetMode::ALL, - combine(combine(shapes_4d_small, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_4d_small, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<short>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_4d_large, - framework::dataset::make("DataType", { DataType::S16 })), - n_values)) + combine(shapes_4d_large, + make("DataType", { DataType::S16 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); @@ -347,24 +363,37 @@ TEST_SUITE_END() // S16 TEST_SUITE(S8) FIXTURE_DATA_TEST_CASE(RunSmall, NEStackLayerFixture<char>, framework::DatasetMode::ALL, - combine(combine(shapes_4d_small, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_4d_small, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } FIXTURE_DATA_TEST_CASE(RunLarge, NEStackLayerFixture<char>, framework::DatasetMode::NIGHTLY, - combine(combine(shapes_4d_large, - framework::dataset::make("DataType", { DataType::S8 })), - n_values)) + combine(shapes_4d_large, + make("DataType", { DataType::S8 }), + n_values)) { // Validate output validate(Accessor(_target), _reference); } TEST_SUITE_END() // S8 TEST_SUITE_END() // Shapes4D + +TEST_SUITE(HighDimensional) +// The Cpu implementation supports tensors of size 4D+, but reference implementation does not. +FIXTURE_DATA_TEST_CASE(RunHighDimensional, NEStackLayerFixture<char>, framework::DatasetMode::DISABLED, + combine(make("Shape", { TensorShape{2U, 3U, 4U, 5U, 3U} }), + make("Axis", { 5, 0, -3, 2 }), + make("DataType", { DataType::S8 }), + make("NumTensors", { 3 }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() // HighDimensional TEST_SUITE_END() // StackLayer TEST_SUITE_END() // Neon } // namespace validation diff --git a/tests/validation/NEON/UNIT/RuntimeContext.cpp b/tests/validation/NEON/UNIT/RuntimeContext.cpp index 819811943d..e126aded28 100644 --- a/tests/validation/NEON/UNIT/RuntimeContext.cpp +++ b/tests/validation/NEON/UNIT/RuntimeContext.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,6 +48,24 @@ namespace validation { TEST_SUITE(NEON) TEST_SUITE(UNIT) +#if defined(ARM_COMPUTE_OPENMP_SCHEDULER) && !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \ + (defined(__arm__) || defined(__aarch64__)) && defined(__ANDROID__) +TEST_CASE(CpuCapacity, framework::DatasetMode::ALL) +{ + CPUInfo& ci = arm_compute::Scheduler::get().cpu_info(); + const uint32_t total_num_cpus = ci.get_cpu_num(); + const uint32_t nonlittle_num_cpus = ci.get_cpu_num_excluding_little(); + const bool has_lmb = ci.cpu_has_little_mid_big(); + const uint32_t num_threads = arm_compute::Scheduler::get().num_threads(); + + if(has_lmb){ + ARM_COMPUTE_EXPECT(total_num_cpus!=nonlittle_num_cpus , framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(num_threads==nonlittle_num_cpus , framework::LogLevel::ERRORS); + } +} +#endif /* defined(ARM_COMPUTE_OPENMP_SCHEDULER) && !defined(_WIN64) && !defined(BARE_METAL) && !defined(__APPLE__) && !defined(__OpenBSD__) && \ + (defined(__arm__) || defined(__aarch64__)) && defined(__ANDROID__)*/ + TEST_SUITE(RuntimeContext) TEST_CASE(Scheduler, framework::DatasetMode::ALL) diff --git a/tests/validation/NEON/UNIT/TensorAllocator.cpp b/tests/validation/NEON/UNIT/TensorAllocator.cpp index d84bcd4a20..0aab9ef9b5 100644 --- a/tests/validation/NEON/UNIT/TensorAllocator.cpp +++ b/tests/validation/NEON/UNIT/TensorAllocator.cpp @@ -193,7 +193,7 @@ TEST_CASE(ImportMemoryMallocPadded, framework::DatasetMode::ALL) ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); } -#if !defined(BARE_METAL) +#if !defined(_WIN64) && !defined(BARE_METAL) TEST_CASE(ImportMemoryMappedFile, framework::DatasetMode::ALL) { const ActivationLayerInfo act_info(ActivationLayerInfo::ActivationFunction::RELU); @@ -250,7 +250,7 @@ TEST_CASE(ImportMemoryMappedFile, framework::DatasetMode::ALL) tensor.allocator()->free(); ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); } -#endif // !defined(BARE_METAL) +#endif // !defined(_WIN64) && !defined(BARE_METAL) TEST_CASE(AlignedAlloc, framework::DatasetMode::ALL) { diff --git a/tests/validation/UNIT/CPPScheduler.cpp b/tests/validation/UNIT/CPPScheduler.cpp new file mode 100644 index 0000000000..6a3f6819fc --- /dev/null +++ b/tests/validation/UNIT/CPPScheduler.cpp @@ -0,0 +1,91 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/runtime/CPP/CPPScheduler.h" + +#include "arm_compute/core/CPP/ICPPKernel.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" + +#include <stdexcept> + +using namespace arm_compute; +using namespace arm_compute::test; + +namespace +{ +class TestException: public std::exception +{ +public: + const char* what() const noexcept override + { + return "Expected test exception"; + } +}; + +class TestKernel: public ICPPKernel +{ +public: + TestKernel() + { + Window window; + window.set(0, Window::Dimension(0, 2)); + configure(window); + } + + const char* name() const override + { + return "TestKernel"; + } + + void run(const Window &, const ThreadInfo &) override + { + throw TestException(); + } + +}; +} + +TEST_SUITE(UNIT) +TEST_SUITE(CPPScheduler) +#if defined(ARM_COMPUTE_CPP_SCHEDULER) && !defined(BARE_METAL) +TEST_CASE(RethrowException, framework::DatasetMode::ALL) +{ + CPPScheduler scheduler; + CPPScheduler::Hints hints(0); + TestKernel kernel; + + scheduler.set_num_threads(2); + try + { + scheduler.schedule(&kernel, hints); + } + catch(const TestException&) + { + return; + } + ARM_COMPUTE_EXPECT_FAIL("Expected exception not caught", framework::LogLevel::ERRORS); +} +#endif // defined(ARM_COMPUTE_CPP_SCHEDULER) && !defined(BARE_METAL) +TEST_SUITE_END() +TEST_SUITE_END() diff --git a/tests/validation/UNIT/GPUTarget.cpp b/tests/validation/UNIT/GPUTarget.cpp index d2c81cf778..2e64635b7a 100644 --- a/tests/validation/UNIT/GPUTarget.cpp +++ b/tests/validation/UNIT/GPUTarget.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -37,6 +37,7 @@ TEST_SUITE(GPUTarget) TEST_CASE(GetGPUTargetFromName, framework::DatasetMode::ALL) { + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-T000") == GPUTarget::MIDGARD, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-T600") == GPUTarget::T600, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-T700") == GPUTarget::T700, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-T800") == GPUTarget::T800, framework::LogLevel::ERRORS); @@ -45,15 +46,24 @@ TEST_CASE(GetGPUTargetFromName, framework::DatasetMode::ALL) ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G51") == GPUTarget::G51, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G51BIG") == GPUTarget::G51BIG, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G51LIT") == GPUTarget::G51LIT, framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G52") == GPUTarget::G52, framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G52LIT") == GPUTarget::G52LIT, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G31") == GPUTarget::G31, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G76") == GPUTarget::G76, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G76 r0p0") == GPUTarget::G76, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G52") == GPUTarget::G52, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G52LIT") == GPUTarget::G52LIT, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G77") == GPUTarget::G77, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G57") == GPUTarget::G57, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G78") == GPUTarget::G78, framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G78AE") == GPUTarget::G78, framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(get_target_from_name("Mali-TODX") == GPUTarget::TODX, framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(get_target_from_name("Mali-T000") == GPUTarget::MIDGARD, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G68") == GPUTarget::G68, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G78AE") == GPUTarget::G78AE, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G710") == GPUTarget::G710, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G610") == GPUTarget::G610, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G510") == GPUTarget::G510, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G310") == GPUTarget::G310, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G715") == GPUTarget::G715, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G615") == GPUTarget::G615, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G720") == GPUTarget::G720, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(get_target_from_name("Mali-G620") == GPUTarget::G620, framework::LogLevel::ERRORS); } TEST_CASE(GPUTargetIsIn, framework::DatasetMode::ALL) diff --git a/tests/validation/UNIT/SubTensorInfo.cpp b/tests/validation/UNIT/SubTensorInfo.cpp index 5a930620ce..ca5e46550c 100644 --- a/tests/validation/UNIT/SubTensorInfo.cpp +++ b/tests/validation/UNIT/SubTensorInfo.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020 Arm Limited. + * Copyright (c) 2020-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -69,6 +69,7 @@ TEST_CASE(SubTensorCreation, framework::DatasetMode::ALL) * - A) Extend padding when SubTensor XY does not match parent tensor should fail * B) Extend with zero padding when SubTensor XY does not match parent tensor should succeed * - C) Extend padding when SubTensor XY matches parent tensor should succeed + * - D) Set lock padding to true, so that extend padding would fail */ TEST_CASE(SubTensorPaddingExpansion, framework::DatasetMode::ALL) { @@ -95,6 +96,14 @@ TEST_CASE(SubTensorPaddingExpansion, framework::DatasetMode::ALL) ARM_COMPUTE_EXPECT(tensor_info.padding().top == 2, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(tensor_info.padding().right == 1, framework::LogLevel::ERRORS); } + + // Test D + { + TensorInfo tensor_info(TensorShape(23U, 17U, 3U), 1, DataType::F32); + SubTensorInfo sub_tensor_info(&tensor_info, TensorShape(4U, 3U, 1U), Coordinates(5, 2, 1)); + sub_tensor_info.set_lock_paddings(true); + ARM_COMPUTE_EXPECT_THROW(sub_tensor_info.extend_padding(PaddingSize(2, 1)), framework::LogLevel::ERRORS); + } } TEST_SUITE_END() // SubTensorInfo diff --git a/tests/validation/UNIT/TensorInfo.cpp b/tests/validation/UNIT/TensorInfo.cpp index 50b26293c9..b79c1e9253 100644 --- a/tests/validation/UNIT/TensorInfo.cpp +++ b/tests/validation/UNIT/TensorInfo.cpp @@ -184,8 +184,17 @@ TEST_CASE(SymmPerChannelQuantizationInfo, framework::DatasetMode::ALL) ARM_COMPUTE_EXPECT(info.quantization_info().offset().empty(), framework::LogLevel::ERRORS); } -TEST_SUITE_END() // TensorInfoValidation -TEST_SUITE_END() +/** Validates lock paddings flag*/ +TEST_CASE(SubTensorPaddingExpansion, framework::DatasetMode::ALL) +{ + TensorInfo tensor_info(TensorShape(23U, 17U, 3U), 1, DataType::F32); + tensor_info.set_lock_paddings(true); + + // Now lock padding is set to true, therefore the extend padding would fail + ARM_COMPUTE_EXPECT_THROW(tensor_info.extend_padding(PaddingSize(2, 1)), framework::LogLevel::ERRORS); +} +TEST_SUITE_END() // TensorInfo +TEST_SUITE_END() // UNIT } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/Validation.h b/tests/validation/Validation.h index f1ce0fecc7..289aca4d08 100644 --- a/tests/validation/Validation.h +++ b/tests/validation/Validation.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,6 +45,17 @@ namespace test { namespace validation { +namespace +{ +// Compare if 2 values are both infinities and if they are "equal" (has the same sign) +template <typename T> +inline bool are_equal_infs(T val0, T val1) +{ + const auto same_sign = support::cpp11::signbit(val0) == support::cpp11::signbit(val1); + return (!support::cpp11::isfinite(val0)) && (!support::cpp11::isfinite(val1)) && same_sign; +} +} // namespace + /** Class reprensenting an absolute tolerance value. */ template <typename T> class AbsoluteTolerance @@ -140,7 +151,7 @@ bool compare_dimensions(const Dimensions<T> &dimensions1, const Dimensions<T> &d { ARM_COMPUTE_ERROR_ON(data_layout == DataLayout::UNKNOWN); - if(data_layout == DataLayout::NCHW) + if(data_layout != DataLayout::NHWC) { if(dimensions1.num_dimensions() != dimensions2.num_dimensions()) { @@ -296,9 +307,9 @@ struct compare<AbsoluteTolerance<U>> : public compare_base<AbsoluteTolerance<U>> /** Perform comparison */ operator bool() const { - if(!support::cpp11::isfinite(this->_target) || !support::cpp11::isfinite(this->_reference)) + if(are_equal_infs(this->_target, this->_reference)) { - return false; + return true; } else if(this->_target == this->_reference) { @@ -322,9 +333,9 @@ struct compare<RelativeTolerance<U>> : public compare_base<RelativeTolerance<U>> /** Perform comparison */ operator bool() const { - if(!support::cpp11::isfinite(this->_target) || !support::cpp11::isfinite(this->_reference)) + if(are_equal_infs(this->_target, this->_reference)) { - return false; + return true; } else if(this->_target == this->_reference) { @@ -494,9 +505,9 @@ void validate_wrap(const IAccessor &tensor, const SimpleTensor<T> &reference, co // check for wrapping if(!equal) { - if(!support::cpp11::isfinite(target_value) || !support::cpp11::isfinite(reference_value)) + if(are_equal_infs(target_value, reference_value)) { - equal = false; + equal = true; } else { diff --git a/tests/validation/cpu/unit/Context.cpp b/tests/validation/cpu/unit/Context.cpp index 57ca866032..42247ba1da 100644 --- a/tests/validation/cpu/unit/Context.cpp +++ b/tests/validation/cpu/unit/Context.cpp @@ -94,13 +94,13 @@ TEST_CASE(CpuCapabilities, framework::DatasetMode::ALL) opts.copts.capabilities = AclCpuCapabilitiesDot | AclCpuCapabilitiesMmlaInt8 | AclCpuCapabilitiesSve2; arm_compute::cpu::CpuContext ctx(&opts.copts); - ARM_COMPUTE_ASSERT(ctx.capabilities().dot == true); - ARM_COMPUTE_ASSERT(ctx.capabilities().mmla_int8 == true); - ARM_COMPUTE_ASSERT(ctx.capabilities().sve2 == true); - ARM_COMPUTE_ASSERT(ctx.capabilities().fp16 == false); + ARM_COMPUTE_ASSERT(ctx.capabilities().cpu_info.has_dotprod() == true); + ARM_COMPUTE_ASSERT(ctx.capabilities().cpu_info.has_i8mm() == true); + ARM_COMPUTE_ASSERT(ctx.capabilities().cpu_info.has_sve2() == true); + ARM_COMPUTE_ASSERT(ctx.capabilities().cpu_info.has_fp16() == false); arm_compute::cpu::CpuContext ctx_legacy(nullptr); - ARM_COMPUTE_ASSERT(ctx_legacy.capabilities().neon == true); + ARM_COMPUTE_ASSERT(ctx_legacy.capabilities().cpu_info.has_neon() == true); } TEST_SUITE_END() // Context diff --git a/tests/validation/dynamic_fusion/Utils.h b/tests/validation/dynamic_fusion/Utils.h new file mode 100644 index 0000000000..72e9ec5955 --- /dev/null +++ b/tests/validation/dynamic_fusion/Utils.h @@ -0,0 +1,73 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef TESTS_VALIDATION_DYNAMIC_FUSION_UTILS +#define TESTS_VALIDATION_DYNAMIC_FUSION_UTILS + +#include "tests/AssetsLibrary.h" +#include "utils/Utils.h" + +#include <chrono> +#include <limits> +#include <type_traits> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace utils +{ +/** A pair of macros which measures the wall clock time, and records it into a map measurement_map with name clock_name + * + */ +#define TICK(clock_name) \ + auto clock_name##_tick = std::chrono::high_resolution_clock::now(); +#define TOCK(clock_name, measurement_map) \ + auto clock_name##_tock = std::chrono::high_resolution_clock::now(); \ + measurement_map["\"" #clock_name "\""] = duration_cast<microseconds>(clock_name##_tock - clock_name##_tick); +#define TOCK_AVG(clock_name, measurement_map, num_iterations) \ + auto clock_name##_tock = std::chrono::high_resolution_clock::now(); \ + measurement_map["\"" #clock_name "\""] = duration_cast<microseconds>((clock_name##_tock - clock_name##_tick) / (num_iterations)); + +template <typename T, typename U> +void fill(U &&tensor, int seed, AssetsLibrary *library) +{ + static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); + using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; + + DistributionType distribution{ T(-1.0f), T(1.0f) }; + library->fill(tensor, distribution, seed); + + // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) + DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; + library->fill_borders_with_garbage(tensor, distribution_inf, seed); +} +} // namespace utils +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif /* TESTS_VALIDATION_DYNAMIC_FUSION_UTILS */ diff --git a/tests/validation/dynamic_fusion/gpu/Integration.cpp b/tests/validation/dynamic_fusion/gpu/Integration.cpp new file mode 100644 index 0000000000..453983c077 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/Integration.cpp @@ -0,0 +1,642 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/QuantizationInfo.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/CastAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/Conv2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuAdd.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuCast.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuMul.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Macros.h" +#include "tests/validation/dynamic_fusion/Utils.h" +#include "tests/validation/reference/ActivationLayer.h" +#include "tests/validation/reference/ConvolutionLayer.h" +#include "tests/validation/reference/DepthConvertLayer.h" +#include "tests/validation/reference/DepthwiseConvolutionLayer.h" +#include "tests/validation/reference/ElementwiseOperations.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/reference/PixelWiseMultiplication.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; +using namespace arm_compute::test::validation::utils; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(INTEGRATION) +TEST_SUITE(DYNAMIC_FUSION) + +TEST_CASE(Conv2d, framework::DatasetMode::ALL) +{ + /* Computation: + * out = conv2d1x1(direct_conv)(input, weights, bias) + */ + CLScheduler::get().default_reinit(); + + const auto data_type = DataType::F32; + const auto data_layout = DataLayout::NHWC; + const auto t_input_shape = TensorShape(384, 12, 12); + const auto t_weight_shape = TensorShape(384, 1, 1, 16); + const auto t_dst_shape = TensorShape(16, 12, 12); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Fuse conv2d + Conv2dAttributes conv2d_attr{}; + ITensorInfo *input_info = context.create_tensor_info(t_input_shape, 1, data_type, data_layout); + ITensorInfo *weight_info = context.create_tensor_info(TensorInfo(t_weight_shape, 1, data_type, data_layout)); + + ITensorInfo *conv_out_info = GpuConv2d::create_op(sketch, input_info, weight_info, nullptr, conv2d_attr); + + ITensorInfo *dst_info = context.create_tensor_info(); + GpuOutput::create_op(sketch, conv_out_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + // Instead of using ACL allocated memory, the user can choose to import memory into the tensors + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + // auto buf = cl::Buffer(); + // tensor->allocator()->import_memory(buf); // Or, import external memory + } + + // Construct user tensors + CLTensor t_input{}; + CLTensor t_weight{}; + CLTensor t_dst{}; + + // Initialize user tensors + t_input.allocator()->init(*input_info); + t_weight.allocator()->init(*weight_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + // Instead of using ACL allocator, the user can choose to import memory into the tensors + t_input.allocator()->allocate(); + t_weight.allocator()->allocate(); + t_dst.allocator()->allocate(); + fill<float>(CLAccessor(t_input), 0, library.get()); + fill<float>(CLAccessor(t_weight), 1, library.get()); + + // Run runtime + runtime.run({&t_input, &t_weight, &t_dst}); + + // Create reference + SimpleTensor<float> ref_t_input{t_input_shape, data_type, 1, QuantizationInfo(), DataLayout::NHWC}; + SimpleTensor<float> ref_t_weight{t_weight_shape, data_type, 1, QuantizationInfo(), DataLayout::NHWC}; + SimpleTensor<float> ref_t_bias_placeholder{t_dst_shape, data_type, 1, QuantizationInfo(), DataLayout::NHWC}; + + // Fill reference + fill<float>(ref_t_input, 0, library.get()); + fill<float>(ref_t_weight, 1, library.get()); + + auto ref_t_input_nchw = reference::permute(ref_t_input, PermutationVector(1U, 2U, 0U)); + auto ref_t_weight_nchw = reference::permute(ref_t_weight, PermutationVector(1U, 2U, 0U)); + auto ref_t_bias_placeholder_nchw = reference::permute(ref_t_bias_placeholder, PermutationVector(1U, 2U, 0U)); + auto t_dst_shape_nchw = t_dst_shape; + permute(t_dst_shape_nchw, PermutationVector(1U, 2U, 0U)); + + PadStrideInfo legacy_pad_stride(conv2d_attr.stride().x(), conv2d_attr.stride().y(), conv2d_attr.pad().left, + conv2d_attr.pad().right, conv2d_attr.pad().top, conv2d_attr.pad().bottom, + DimensionRoundingType{}); + auto ref_t_dst_nchw = reference::convolution_layer(ref_t_input_nchw, ref_t_weight_nchw, ref_t_bias_placeholder_nchw, + t_dst_shape_nchw, legacy_pad_stride, conv2d_attr.dilation()); + const auto ref_t_dst = reference::permute(ref_t_dst_nchw, PermutationVector(2U, 0U, 1U)); + + RelativeTolerance<float> tolerance_f32( + 0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ + validate(CLAccessor(t_dst), ref_t_dst_nchw, tolerance_f32); +} + +TEST_CASE(Add_Output_Add_Output, framework::DatasetMode::ALL) +{ + /* Computation: + * out_0 = in_0 + in_1 + * out_1 = out_0 + in_2 + */ + CLScheduler::get().default_reinit(); + + const auto data_type = DataType::F32; + const auto t_input_shape = TensorShape(33, 3, 2); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + ITensorInfo *in_0_info = context.create_tensor_info(t_input_shape, 1, data_type); + ITensorInfo *in_1_info = context.create_tensor_info(t_input_shape, 1, data_type); + ITensorInfo *in_2_info = context.create_tensor_info(t_input_shape, 1, data_type); + + ITensorInfo *out_0_info = context.create_tensor_info(); + ITensorInfo *out_1_info = context.create_tensor_info(); + + ITensorInfo *ans_0_info = GpuAdd::create_op(sketch, in_0_info, in_1_info); + GpuOutput::create_op(sketch, ans_0_info, out_0_info); + ITensorInfo *ans_1_info = GpuAdd::create_op(sketch, ans_0_info, in_2_info); + GpuOutput::create_op(sketch, ans_1_info, out_1_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + // Instead of using ACL allocated memory, the user can choose to import memory into the tensors + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + // auto buf = cl::Buffer(); + // tensor->allocator()->import_memory(buf); // Or, import external memory + } + + // Construct user tensors + CLTensor t_in_0{}; + CLTensor t_in_1{}; + CLTensor t_in_2{}; + + CLTensor t_out_0{}; + CLTensor t_out_1{}; + + // Initialize user tensors + t_in_0.allocator()->init(*in_0_info); + t_in_1.allocator()->init(*in_1_info); + t_in_2.allocator()->init(*in_2_info); + + t_out_0.allocator()->init(*out_0_info); + t_out_1.allocator()->init(*out_1_info); + + // Allocate and fill user tensors + // Instead of using ACL allocator, the user can choose to import memory into the tensors + t_in_0.allocator()->allocate(); + t_in_1.allocator()->allocate(); + t_in_2.allocator()->allocate(); + + t_out_0.allocator()->allocate(); + t_out_1.allocator()->allocate(); + + fill<float>(CLAccessor(t_in_0), 0, library.get()); + fill<float>(CLAccessor(t_in_1), 1, library.get()); + fill<float>(CLAccessor(t_in_2), 2, library.get()); + + // Run runtime + runtime.run({&t_in_0, &t_in_1, &t_in_2, &t_out_0, &t_out_1}); + + // Create reference + SimpleTensor<float> ref_t_in_0{t_input_shape, data_type, 1, QuantizationInfo()}; + SimpleTensor<float> ref_t_in_1{t_input_shape, data_type, 1, QuantizationInfo()}; + SimpleTensor<float> ref_t_in_2{t_input_shape, data_type, 1, QuantizationInfo()}; + + SimpleTensor<float> ref_t_out_0{t_input_shape, data_type, 1, QuantizationInfo()}; + SimpleTensor<float> ref_t_out_1{t_input_shape, data_type, 1, QuantizationInfo()}; + + // Fill reference + fill<float>(ref_t_in_0, 0, library.get()); + fill<float>(ref_t_in_1, 1, library.get()); + fill<float>(ref_t_in_2, 2, library.get()); + + reference::arithmetic_operation(ArithmeticOperation::ADD, ref_t_in_0, ref_t_in_1, ref_t_out_0, ConvertPolicy::WRAP); + reference::arithmetic_operation(ArithmeticOperation::ADD, ref_t_out_0, ref_t_in_2, ref_t_out_1, + ConvertPolicy::WRAP); + + RelativeTolerance<float> tolerance_f32( + 0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ + validate(CLAccessor(t_out_0), ref_t_out_0, tolerance_f32); + validate(CLAccessor(t_out_1), ref_t_out_1, tolerance_f32); +} +TEST_CASE(Add_Output_Add_Cast_Cast_Output, framework::DatasetMode::ALL) +{ + /* Computation: + * out_0 = in_0 + in_1 + * out_1 = float(int32_t(out_0 + in_2)) + */ + CLScheduler::get().default_reinit(); + + const auto data_type = DataType::F32; + const auto t_input_shape = TensorShape(3, 8, 5); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + ITensorInfo *in_0_info = context.create_tensor_info(t_input_shape, 1, data_type); + ITensorInfo *in_1_info = context.create_tensor_info(t_input_shape, 1, data_type); + ITensorInfo *in_2_info = context.create_tensor_info(t_input_shape, 1, data_type); + + ITensorInfo *out_0_info = context.create_tensor_info(); + ITensorInfo *out_1_info = context.create_tensor_info(); + + CastAttributes cast_0_attr; + cast_0_attr.data_type(DataType::F16); + + CastAttributes cast_1_attr; + cast_1_attr.data_type(DataType::F32); + + ITensorInfo *ans_0_info = GpuAdd::create_op(sketch, in_0_info, in_1_info); + GpuOutput::create_op(sketch, ans_0_info, out_0_info); + ITensorInfo *ans_1_info = GpuAdd::create_op(sketch, ans_0_info, in_2_info); + ITensorInfo *ans_2_info = GpuCast::create_op(sketch, ans_1_info, cast_0_attr); + ITensorInfo *ans_3_info = GpuCast::create_op(sketch, ans_2_info, cast_1_attr); + GpuOutput::create_op(sketch, ans_3_info, out_1_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + // Instead of using ACL allocated memory, the user can choose to import memory into the tensors + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + // auto buf = cl::Buffer(); + // tensor->allocator()->import_memory(buf); // Or, import external memory + } + + // Construct user tensors + CLTensor t_in_0{}; + CLTensor t_in_1{}; + CLTensor t_in_2{}; + + CLTensor t_out_0{}; + CLTensor t_out_1{}; + + // Initialize user tensors + t_in_0.allocator()->init(*in_0_info); + t_in_1.allocator()->init(*in_1_info); + t_in_2.allocator()->init(*in_2_info); + + t_out_0.allocator()->init(*out_0_info); + t_out_1.allocator()->init(*out_1_info); + + // Allocate and fill user tensors + // Instead of using ACL allocator, the user can choose to import memory into the tensors + t_in_0.allocator()->allocate(); + t_in_1.allocator()->allocate(); + t_in_2.allocator()->allocate(); + + t_out_0.allocator()->allocate(); + t_out_1.allocator()->allocate(); + + fill<float>(CLAccessor(t_in_0), 0, library.get()); + fill<float>(CLAccessor(t_in_1), 1, library.get()); + fill<float>(CLAccessor(t_in_2), 2, library.get()); + + // Run runtime + runtime.run({&t_in_0, &t_in_1, &t_in_2, &t_out_0, &t_out_1}); + + // Create reference + SimpleTensor<float> ref_t_in_0{t_input_shape, data_type, 1, QuantizationInfo()}; + SimpleTensor<float> ref_t_in_1{t_input_shape, data_type, 1, QuantizationInfo()}; + SimpleTensor<float> ref_t_in_2{t_input_shape, data_type, 1, QuantizationInfo()}; + + SimpleTensor<float> ref_t_out_0{t_input_shape, data_type, 1, QuantizationInfo()}; + SimpleTensor<float> ref_t_ans_1{t_input_shape, data_type, 1, QuantizationInfo()}; + + // Fill reference + fill<float>(ref_t_in_0, 0, library.get()); + fill<float>(ref_t_in_1, 1, library.get()); + fill<float>(ref_t_in_2, 2, library.get()); + + reference::arithmetic_operation(ArithmeticOperation::ADD, ref_t_in_0, ref_t_in_1, ref_t_out_0, ConvertPolicy::WRAP); + reference::arithmetic_operation(ArithmeticOperation::ADD, ref_t_out_0, ref_t_in_2, ref_t_ans_1, + ConvertPolicy::WRAP); + const auto ref_t_ans_2 = + reference::depth_convert<float, int32_t>(ref_t_ans_1, DataType::S32, ConvertPolicy::SATURATE, 0); + const auto ref_t_out_1 = + reference::depth_convert<int32_t, float>(ref_t_ans_2, DataType::F32, ConvertPolicy::SATURATE, 0); + + RelativeTolerance<float> tolerance_add_f32(0.001f); + AbsoluteTolerance<float> tolerance_cast_f32(1.0f); + validate(CLAccessor(t_out_0), ref_t_out_0, tolerance_add_f32); + validate(CLAccessor(t_out_1), ref_t_out_1, tolerance_cast_f32); +} + +/// TODO: COMPMID-6593 : This integration test fails with CKW backend. +/// It was not enabled for CKW before, therefore went unnoticed. +TEST_CASE(Conv2d_Sigmoid_DepthwiseConv2d_Mul, framework::DatasetMode::DISABLED) +{ + // (tensor0) + // | + // ======|============================================== Sketch 0 + // | (tensor1) +---- (tensor2) + // | | | | + // +-- input -- weights -- biases --+ | + // | | | + // | Conv2d | | + // | | | + // +----------- output -------------+ | + // | | + // +-- input ---+ | + // | | | + // | Sigmoid | | + // | | | + // +-- output --+ | + // | | + // +-- input ---+ | + // | | | + // | Output | | + // | | | + // +-- output --+ | + // | | + // (tensor5) | + // | | + // +--------+ | + // ======|=============================|================ Sketch 1 + // | (tensor3) (tensor4) | + // | | | | + // +-- input -- weights -- biases --+ | + // | | | + // | DepthwiseConv2d | | + // | | | + // +----------- output -------------+ | + // | | + // +--+ +----------------+ + // | | + // +-- lhs -- rhs --+ + // | | + // | Multiply | + // | | + // +---- output ----+ + // | + // +-- input ---+ + // | | + // | Output | + // | | + // +-- output --+ + // | + // (tensor6) + + TensorShape conv2d_src_shape(10, 20, 30); + TensorShape conv2d_wei_shape(10, 3, 3, 5); + TensorShape conv2d_bia_shape(5); + TensorShape conv2d_dst_shape(5, 18, 28); + TensorShape dwc_wei_shape(5, 3, 3); + TensorShape dwc_bia_shape(5); + TensorShape dwc_dst_shape(5, 16, 26); + + // Initialize the context. + CLScheduler::get().default_reinit(); + + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context(&cl_compile_ctx); + + auto tensor0_info = context.create_tensor_info(conv2d_src_shape, 1, DataType::F32, DataLayout::NHWC); + + // Create the first sketch: conv2d + cast + output. + GpuWorkloadSketch sketch0(&context); + + Conv2dAttributes conv2d_attr; + auto tensor1_info = context.create_tensor_info(conv2d_wei_shape, 1, DataType::F32, DataLayout::NHWC); + auto tensor2_info = context.create_tensor_info(conv2d_bia_shape, 1, DataType::F32, DataLayout::NHWC); + ARM_COMPUTE_EXPECT(GpuConv2d::validate_op(sketch0, tensor0_info, tensor1_info, tensor2_info, conv2d_attr), + framework::LogLevel::ERRORS); + auto ans_info = GpuConv2d::create_op(sketch0, tensor0_info, tensor1_info, tensor2_info, conv2d_attr); + + ARM_COMPUTE_EXPECT(GpuSigmoid::validate_op(sketch0, ans_info), framework::LogLevel::ERRORS); + ans_info = GpuSigmoid::create_op(sketch0, ans_info); + + DepthwiseConv2dAttributes dwc_attr; + auto tensor3_info = context.create_tensor_info(dwc_wei_shape, 1, DataType::F32, DataLayout::NHWC); + auto tensor4_info = context.create_tensor_info(dwc_bia_shape, 1, DataType::F32, DataLayout::NHWC); + ARM_COMPUTE_EXPECT(!GpuDepthwiseConv2d::validate_op(sketch0, ans_info, tensor3_info, tensor4_info, dwc_attr), + framework::LogLevel::ERRORS); + + auto tensor5_info = context.create_tensor_info(); + ARM_COMPUTE_EXPECT(GpuOutput::validate_op(sketch0, ans_info, tensor5_info), framework::LogLevel::ERRORS); + GpuOutput::create_op(sketch0, ans_info, tensor5_info); + + // Create the first workload runtime. + ClWorkloadRuntime runtime0; + runtime0.configure(sketch0); + + // Create the second sketch: dwc + sigmoid + output. + GpuWorkloadSketch sketch1(&context); + + ARM_COMPUTE_EXPECT(GpuDepthwiseConv2d::validate_op(sketch1, tensor5_info, tensor3_info, tensor4_info, dwc_attr), + framework::LogLevel::ERRORS); + ans_info = GpuDepthwiseConv2d::create_op(sketch1, tensor5_info, tensor3_info, tensor4_info, dwc_attr); + + ARM_COMPUTE_EXPECT(GpuMul::validate_op(sketch1, ans_info, tensor2_info), framework::LogLevel::ERRORS); + ans_info = GpuMul::create_op(sketch1, ans_info, tensor2_info); + + auto tensor6_info = context.create_tensor_info(); + ARM_COMPUTE_EXPECT(GpuOutput::validate_op(sketch1, ans_info, tensor6_info), framework::LogLevel::ERRORS); + GpuOutput::create_op(sketch1, ans_info, tensor6_info); + + // Create the second workload runtime. + ClWorkloadRuntime runtime1; + runtime1.configure(sketch1); + + // Create the user tensors. + CLTensor tensor0; + CLTensor tensor1; + CLTensor tensor2; + CLTensor tensor3; + CLTensor tensor4; + CLTensor tensor5; + CLTensor tensor6; + + tensor0.allocator()->init(*tensor0_info); + tensor1.allocator()->init(*tensor1_info); + tensor2.allocator()->init(*tensor2_info); + tensor3.allocator()->init(*tensor3_info); + tensor4.allocator()->init(*tensor4_info); + tensor5.allocator()->init(*tensor5_info); + tensor6.allocator()->init(*tensor6_info); + + tensor0.allocator()->allocate(); + tensor1.allocator()->allocate(); + tensor2.allocator()->allocate(); + tensor3.allocator()->allocate(); + tensor4.allocator()->allocate(); + tensor5.allocator()->allocate(); + tensor6.allocator()->allocate(); + + // Allocate the auxiliary tensors. + for (auto &data : runtime0.get_auxiliary_tensors()) + { + auto tensor = std::get<0>(data); + auto &tensor_info = std::get<1>(data); + auto mem_req = std::get<2>(data); + + tensor->allocator()->init(tensor_info, mem_req.alignment); + tensor->allocator()->allocate(); + } + + for (auto &data : runtime1.get_auxiliary_tensors()) + { + auto tensor = std::get<0>(data); + auto &tensor_info = std::get<1>(data); + auto mem_req = std::get<2>(data); + + tensor->allocator()->init(tensor_info, mem_req.alignment); + tensor->allocator()->allocate(); + } + + // Fill the input tensors with random data. + fill<float>(CLAccessor(tensor0), 0, library.get()); + fill<float>(CLAccessor(tensor1), 1, library.get()); + fill<float>(CLAccessor(tensor2), 2, library.get()); + fill<float>(CLAccessor(tensor3), 3, library.get()); + fill<float>(CLAccessor(tensor4), 4, library.get()); + + // Run each runtime. + runtime0.run({&tensor0, &tensor1, &tensor2, &tensor5}); + runtime1.run({&tensor5, &tensor3, &tensor4, &tensor2, &tensor6}); + + // Compute the reference result. + SimpleTensor<float> ref_conv2d_src(conv2d_src_shape, DataType::F32, 1, QuantizationInfo(), DataLayout::NHWC); + SimpleTensor<float> ref_conv2d_wei(conv2d_wei_shape, DataType::F32, 1, QuantizationInfo(), DataLayout::NHWC); + SimpleTensor<float> ref_conv2d_bia(conv2d_bia_shape, DataType::F32, 1, QuantizationInfo(), DataLayout::NHWC); + SimpleTensor<float> ref_dwc_wei(dwc_wei_shape, DataType::F32, 1, QuantizationInfo(), DataLayout::NHWC); + SimpleTensor<float> ref_dwc_bia(dwc_bia_shape, DataType::F32, 1, QuantizationInfo(), DataLayout::NHWC); + + fill<float>(ref_conv2d_src, 0, library.get()); + fill<float>(ref_conv2d_wei, 1, library.get()); + fill<float>(ref_conv2d_bia, 2, library.get()); + fill<float>(ref_dwc_wei, 3, library.get()); + fill<float>(ref_dwc_bia, 4, library.get()); + + PermutationVector nhwc_to_nchw(1, 2, 0); + + auto conv2d_dst_shape_nchw = conv2d_dst_shape; + permute(conv2d_dst_shape_nchw, nhwc_to_nchw); + const auto ref_conv2d_src_nchw = reference::permute(ref_conv2d_src, nhwc_to_nchw); + const auto ref_conv2d_wei_nchw = reference::permute(ref_conv2d_wei, nhwc_to_nchw); + const auto ref_conv2d_bia_nchw = reference::permute(ref_conv2d_bia, nhwc_to_nchw); + const auto ref_conv2d_dst_nchw = reference::convolution_layer( + ref_conv2d_src_nchw, ref_conv2d_wei_nchw, ref_conv2d_bia_nchw, conv2d_dst_shape_nchw, PadStrideInfo()); + + const auto ref_sigmoid_dst_nchw = reference::activation_layer( + ref_conv2d_dst_nchw, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + + auto dwc_dst_shape_nchw = dwc_dst_shape; + permute(dwc_dst_shape_nchw, nhwc_to_nchw); + const auto ref_dwc_wei_nchw = reference::permute(ref_dwc_wei, nhwc_to_nchw); + const auto ref_dwc_bia_nchw = reference::permute(ref_dwc_bia, nhwc_to_nchw); + const auto ref_dwc_dst_nchw = reference::depthwise_convolution( + ref_sigmoid_dst_nchw, ref_dwc_wei_nchw, ref_dwc_bia_nchw, dwc_dst_shape_nchw, PadStrideInfo(), 1); + + const auto ref_mul_dst_nchw = reference::pixel_wise_multiplication<float, float, float>( + ref_dwc_dst_nchw, ref_conv2d_bia_nchw, 1.0, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_UP, + DataType::F32); + + constexpr RelativeTolerance<float> tolerance(0.001f); + validate(CLAccessor(tensor6), ref_mul_dst_nchw, tolerance); +} + +TEST_SUITE(Invalid_Fusion_Should_Fail) +TEST_CASE(Multiple_Complex_Ops_0, framework::DatasetMode::ALL) +{ + /* Computation: + * out = conv2d(conv2d(l0_input, l0_weight), l1_weight) + */ + CLScheduler::get().default_reinit(); + + const auto data_type = DataType::F32; + const auto data_layout = DataLayout::NHWC; + const auto t_input_shape = TensorShape(384, 12, 12); + const auto t_weight_shape = TensorShape(384, 1, 1, 16); + auto t_input_info = TensorInfo(t_input_shape, 1, data_type, data_layout); + auto t_weight_info = TensorInfo(t_weight_shape, 1, data_type, data_layout); + auto t_dst_info = TensorInfo(); + + Conv2dAttributes conv2d_attr{}; + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create tensor infos + ITensorInfo *input_info = context.create_tensor_info(t_input_shape, 1, data_type, data_layout); + ITensorInfo *weight_info = context.create_tensor_info(TensorInfo(t_weight_shape, 1, data_type, data_layout)); + ITensorInfo *dst_info; + + // Fuse conv2d into the workload + { + // Validate operator + const Status success = GpuConv2d::validate_op(sketch, input_info, weight_info, nullptr, conv2d_attr); + ARM_COMPUTE_EXPECT(bool(success), framework::LogLevel::ERRORS); + + dst_info = GpuConv2d::create_op(sketch, input_info, weight_info, nullptr, conv2d_attr); + } + + // Create tensor infos + ITensorInfo *weight_info_2 = context.create_tensor_info(t_weight_info); + + // Fuse conv2d into the workload + { + // Validate operator, should fail + const Status success = GpuConv2d::validate_op(sketch, dst_info, weight_info_2, nullptr, conv2d_attr); + const auto expected_error_str = "Operator fusion test failed. This operator cannot be fused into the workload"; + + ARM_COMPUTE_EXPECT(!bool(success), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT((success.error_description().find(expected_error_str) != std::string::npos), + framework::LogLevel::ERRORS); + } +} +TEST_SUITE_END() // Invalid_Fusion_Should_Fail +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // INTEGRATION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Add.cpp b/tests/validation/dynamic_fusion/gpu/cl/Add.cpp new file mode 100644 index 0000000000..9bfdc961fe --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Add.cpp @@ -0,0 +1,264 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuAdd.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/DynamicFusionDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/* Synced with tests/validation/CL/ArithmeticAddition.cpp from the standard interface. + * + * Difference | Why the difference + * No quantized tests | Not supported yet + * No in place tests | Not supported yet + * No activation tests | Not needed in dynamic fusion interface + * + */ +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(ADD) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( + framework::dataset::make("LhsInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), // S16 is valid data type for Add + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), // S32 is valid data type for Add + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(32U, 1U, 1U), 1, DataType::F32), // Broadcasting allowed for lhs + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8_SIGNED), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(15U, 23U, 3U), 1, DataType::F32), // Broadcast Y dimension is not allowed + TensorInfo(TensorShape( 3U, 8U, 9U), 1, DataType::S16), // Broadcast Z dimension is not allowed + TensorInfo(TensorShape(32U, 13U, 2U, 2), 1, DataType::F32), // Batching is allowed + }), + framework::dataset::make("RhsInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), + TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8_SIGNED), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 1U, 1U), 1, DataType::F32), // Broadcasting allowed for rhs + TensorInfo(TensorShape(15U, 1U, 3U), 1, DataType::F32), + TensorInfo(TensorShape( 3U, 8U, 1U), 1, DataType::S16), + TensorInfo(TensorShape(32U, 13U, 2U, 2), 1, DataType::F32), + })), + framework::dataset::make("Expected", { true, false, true, true, false, true, false, false, true, false, false, true})), + input1_info, input2_info, expected) +{ + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Validate Elementwise Add + auto lhs_info = context.create_tensor_info(input1_info); + auto rhs_info = context.create_tensor_info(input2_info); + + bool res = bool(GpuAdd::validate_op(sketch, lhs_info, rhs_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +constexpr AbsoluteTolerance<float> tolerance_f( + 0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 and DataType::F16 */ +constexpr float tolerance_num = 0.0001f; /**< Tolerance number */ + +template <typename T> +using DynamicFusionCLAddFixture = + DynamicFusionGpuElementwiseBinaryOneOpValidationFixture<CLTensor, CLAccessor, GpuAdd, T>; + +template <typename T> +using DynamicFusionCLAddBroadcastFixture = + DynamicFusionGpuElementwiseBinaryBroadcastOneOpValidationFixture<CLTensor, CLAccessor, GpuAdd, T>; + +template <typename T> +using DynamicFusionCLAddTwoOpsFixture = + DynamicFusionGpuElementwiseBinaryTwoOpsValidationFixture<CLTensor, CLAccessor, GpuAdd, T>; + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionCLAddFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f); +} +FIXTURE_DATA_TEST_CASE(RunLargeOneOp, + DynamicFusionCLAddFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::LargeShapes()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f); +} +FIXTURE_DATA_TEST_CASE(RunSmallBroadcastOneOp, + DynamicFusionCLAddBroadcastFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::TemporaryLimitedSmallShapesBroadcast()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f); +} + +FIXTURE_DATA_TEST_CASE(RunLargeBroadcastOneOp, + DynamicFusionCLAddBroadcastFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::TemporaryLimitedLargeShapesBroadcast()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f); +} +FIXTURE_DATA_TEST_CASE( + RunSmallTwoOps, + DynamicFusionCLAddTwoOpsFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::DynamicFusionElementwiseBinaryTwoOpsSmallShapes()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false})), + framework::dataset::make("FuseTwoOps", {true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionCLAddFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f, tolerance_num); +} + +FIXTURE_DATA_TEST_CASE(RunSmallBroadcastOneOp, + DynamicFusionCLAddBroadcastFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::TemporaryLimitedSmallShapesBroadcast()), + framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f, tolerance_num); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(S32) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionCLAddFixture<int32_t>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::S32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // S32 + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionCLAddFixture<int16_t>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::S16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionCLAddFixture<int16_t>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::LargeShapes()), + framework::dataset::make("DataType", {DataType::S16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // S16 + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionCLAddFixture<uint8_t>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::ADD}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::U8})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // U8 + +TEST_SUITE_END() // ADD +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Cast.cpp b/tests/validation/dynamic_fusion/gpu/cl/Cast.cpp new file mode 100644 index 0000000000..4ef359e74d --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Cast.cpp @@ -0,0 +1,97 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuCast.h" +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ConvertPolicyDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +// Tolerance +constexpr AbsoluteTolerance<float> zero_tolerance(0); + +/** Input data sets **/ + +// F16 +const auto CastF16toF32Dataset = combine(framework::dataset::make("DataType", DataType::F16), framework::dataset::make("DataType", DataType::F32)); + +// F32 +const auto CastF32toF16Dataset = combine(framework::dataset::make("DataType", DataType::F32), framework::dataset::make("DataType", DataType::F16)); + +class DFConvertPolicies final : public framework::dataset::ContainerDataset<std::vector<ConvertPolicy>> +{ +public: + DFConvertPolicies() + : ContainerDataset("ConvertPolicy", + { + ConvertPolicy::WRAP + }) + { + } +}; +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(CAST) + +template <typename T> +using DynamicFusionCLCastToF16Fixture = DynamicFusionCastValidationFixture<CLTensor, CLAccessor, GpuCast, T, half>; +template <typename T> +using DynamicFusionCLCastToF32Fixture = DynamicFusionCastValidationFixture<CLTensor, CLAccessor, GpuCast, T, float>; + +#define CAST_SUITE(NAME, idt, odt, type, dataset, tolerance) \ + TEST_SUITE(NAME) \ + FIXTURE_DATA_TEST_CASE(RunSmall, type, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapes(), dataset), \ + DFConvertPolicies())) \ + { \ + validate(CLAccessor(_target), _reference, tolerance); \ + } \ + TEST_SUITE_END() + +// F16 +CAST_SUITE(F16_to_F32, DataType::F16, DataType::F32, DynamicFusionCLCastToF32Fixture<half>, CastF16toF32Dataset, zero_tolerance) + +// F32 +CAST_SUITE(F32_to_F16, DataType::F32, DataType::F16, DynamicFusionCLCastToF16Fixture<float>, CastF32toF16Dataset, zero_tolerance) + +TEST_SUITE_END() // CAST +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp b/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp new file mode 100644 index 0000000000..cef8b87c3f --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Clamp.cpp @@ -0,0 +1,184 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/ClampAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuClamp.h" +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr float epsilon = 1e-6f; +constexpr AbsoluteTolerance<float> tolerance(epsilon); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(CLAMP) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Minimum value larger than maximum value + }), + framework::dataset::make("MinVal", { 0.2f, + 1.5f, + 9.0f, + })), + framework::dataset::make("MaxVal", { 0.5f, + 2.0f, + 1.0f, + })), + framework::dataset::make("Expected", { true, true, false })), + input_info, min_val, max_val, expected) +{ + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Fuse Clamp + const ITensorInfo* src_info = context.create_tensor_info(input_info); + + ClampAttributes attributes {}; + attributes.min_val(min_val) + .max_val(max_val); + + const bool res = static_cast<bool>(GpuClamp::validate_op(sketch, src_info, attributes)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using DynamicFusionClampOpFixture = DynamicFusionClampValidationFixture<CLTensor, CLAccessor, GpuClamp, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionClampOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make( + "ClampAttributes", {ClampAttributes().min_val(0.1f).max_val(0.6f)})), + framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionClampOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::Small5dShapes(), + framework::dataset::make( + "ClampAttributes", {ClampAttributes().min_val(0.1f).max_val(0.6f)})), + framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + ARM_COMPUTE_TEST_INFO("Currently 5D+ tensors are unsupported for this operation."); + framework::ARM_COMPUTE_PRINT_INFO(); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionClampOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make( + "ClampAttributes", {ClampAttributes().min_val(0.2f).max_val(0.4f)})), + framework::dataset::make("Fuse", {true})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionClampOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make( + "ClampAttributes", {ClampAttributes().min_val(0.3f).max_val(0.7f)})), + framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionClampOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::Small5dShapes(), + framework::dataset::make( + "ClampAttributes", {ClampAttributes().min_val(0.3f).max_val(0.7f)})), + framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + ARM_COMPUTE_TEST_INFO("Currently 5D+ tensors are unsupported for this operation."); + framework::ARM_COMPUTE_PRINT_INFO(); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionClampOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallShapes(), + framework::dataset::make( + "ClampAttributes", {ClampAttributes().min_val(0.1f).max_val(0.9f)})), + framework::dataset::make("Fuse", {true})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance); +} + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float + +TEST_SUITE_END() // CLAMP +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp new file mode 100644 index 0000000000..2f8c639cea --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/DepthwiseConv2d.cpp @@ -0,0 +1,474 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/DepthwiseConvolutionLayerDataset.h" +#include "tests/datasets/DilatedDepthwiseConvolutionLayerDataset.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +const auto depth_multipliers = framework::dataset::make("DepthMultiplier", {1U, 4U}); +const auto large_depth_multipliers = framework::dataset::make("DepthMultiplier", {1, 2, 5, 8}); + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(DEPTHWISE_CONV2D) + +RelativeTolerance<float> tolerance_f32( + 0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ +RelativeTolerance<half_float::half> tolerance_f16(half_float::half( + 0.1)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr float tolerance_num = 0.02f; /**< Tolerance number */ + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( // Explanations of failing tests + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Mismatching data type input/weights + TensorInfo(TensorShape(3U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Mismatching input feature maps + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Mismatching depth multiplier + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Invalid biases size + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Invalid biases dimensions + TensorInfo(TensorShape(8U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // dilation < 1 + TensorInfo(TensorShape(8U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QSYMM16, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QSYMM8, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QSYMM8_PER_CHANNEL, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::QASYMM16, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::U8, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::S8, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::U16, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::S16, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::U32, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(8U, 32U, 13U), 1, DataType::S32, DataLayout::NHWC), // Unsupported data type + TensorInfo(TensorShape(32U, 13U, 8U), 1, DataType::F32, DataLayout::NCHW), // Unsupported data layout + TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), // weight dimension > 3 + TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(8U, 32U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), + }), + framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(16U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(16U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QASYMM8, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QASYMM8_SIGNED, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QSYMM16, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QSYMM8, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QSYMM8_PER_CHANNEL, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::QASYMM16, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::U8, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::S8, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::U16, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::S16, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::U32, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 3U, 24U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U, 5U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 3U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U, 4U, 3U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(16U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(16U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::S32, DataLayout::NCHW), + TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(24U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("Padding", { Padding2D(0, 0, 0, 0), + Padding2D(0, 0, 0, 0), + Padding2D(0, 0, 0, 0), + Padding2D(0, 0, 0, 0), + Padding2D(0, 0, 0, 0), + Padding2D(0, 0, 0, 0), + Padding2D(0, 0, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(1, 1, 0, 0), + Padding2D(2, 1, 2, 1), + Padding2D(2, 1, 2, 1), + Padding2D(2, 1, 2, 1), + })), + framework::dataset::make("Stride", { Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(1, 1), + Size2D(2, 3), + Size2D(2, 3), + })), + framework::dataset::make("DepthMultiplier", { 1, + 1, + 3, + 1, + 1, + 2, + 2, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + 3, + })), + framework::dataset::make("Dilation", { Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(0U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(1U, 1U), + Size2D(2U, 3U), + })), + framework::dataset::make("Expected", { false, false, false, false, false, false, true, false, + false, false, false, false, false, false, false, false, false, false, + false, false, true, false, true, true, true })), + input_info, weights_info, biases_info, padding, stride, depth_multiplier, dilation, expected) +{ + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + const ITensorInfo* sketch_input_info = context.create_tensor_info(input_info); + const ITensorInfo* sketch_weights_info = context.create_tensor_info(weights_info); + const ITensorInfo* sketch_biases_info = context.create_tensor_info(biases_info); + + DepthwiseConv2dAttributes attributes {}; + attributes.pad(padding) + .stride(stride) + .dilation(dilation) + .depth_multiplier(depth_multiplier); + + const Status status = GpuDepthwiseConv2d::validate_op(sketch, sketch_input_info, sketch_weights_info, sketch_biases_info, attributes); + const bool res = bool(status); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using DynamicFusionGpuDepthwiseConv2dFixture = + DynamicFusionGpuDepthwiseConv2dValidationFixture<CLTensor, CLAccessor, GpuDepthwiseConv2d, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +TEST_SUITE(W3x3) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE(Dilation) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // Dilation +TEST_SUITE_END() // W3x3 + +TEST_SUITE(Generic) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} + +TEST_SUITE(Dilation) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} +TEST_SUITE_END() // Dilation +TEST_SUITE_END() // Generic +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +TEST_SUITE(W3x3) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE(Dilation) + +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset3x3(), + depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // Dilation +TEST_SUITE_END() // W3x3 + +TEST_SUITE(Generic) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLargeKernelSize, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::LargeKernelSizeDepthwiseConvolutionLayerNHWCDataset(), + framework::dataset::make("DepthMultiplier", {1})), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE(Dilation) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallDepthwiseDilatedConvolutionLayerDataset(), + depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuDepthwiseConv2dFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(datasets::LargeDepthwiseDilatedConvolutionLayerDataset3x3(), + large_depth_multipliers), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", {DataLayout::NHWC}))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // Dilation +TEST_SUITE_END() // Generic +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float +TEST_SUITE_END() // DEPTHWISE_CONV2D +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp new file mode 100644 index 0000000000..b843764786 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp @@ -0,0 +1,260 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "tests/AssetsLibrary.h" +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/SmallConvolutionLayerDataset.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h" +#include "tests/validation/reference/ConvolutionLayer.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +/** Tolerances from tests/validation/CL/DirectConvolutionLayer.cpp + */ +RelativeTolerance<float> tolerance_f32( + 0.05f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ +RelativeTolerance<half_float::half> tolerance_f16(half_float::half( + 0.2)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr float abs_tolerance_f32(0.0001f); /**< Absolute tolerance for FP32 tests*/ +constexpr float tolerance_num = 0.07f; /**< Tolerance number */ +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +/** Synced with tests/validation/CL/ConvolutionLayer.cpp + * + * Difference | Why the difference + * f32 tolerance here is smaller | To use the same tolerance as that of DirectConv2d; lowering tolerance is safe + * No quantized tests | Not supported yet + * No grouped CNN tests | Not supported yet + * No mixed layout tests | Not needed; only NHWC is supported + * No activation | Not needed in fusion + * No ValidateConvolutionMethod | Only a single method (direct conv2d) is supported + * No ReshapeWeights = true tests | Not applicable yet. This parameter only concerns gemm-based conv2d + * No RunSmallWithPadding tests | Padding is removed + * + */ +TEST_SUITE(CONV2D) + +template <typename T> +using DynamicFusionGpuConv2dFixture = DynamicFusionGpuConv2dValidationFixture<CLTensor, CLAccessor, GpuConv2d, T>; +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuConv2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", {DataLayout::NHWC})), + framework::dataset::make("QuantizationInfo", QuantizationInfo()))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuConv2dFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", {DataLayout::NHWC})), + framework::dataset::make("QuantizationInfo", QuantizationInfo()))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} +TEST_SUITE_END() // FP16 + +// Tests for specific conv2d methods +/** Synced with tests/validation/CL/DirectConvolutionLayer.cpp + * + * Difference | Why the difference + * No quantized tests | Not supported yet + * No Invalid output size test | Not applicable. Output is removed from the interface + * No mixed layout/NCHW tests | Not needed; only NHWC is supported + * No activation tests | Not needed in fusion + */ +TEST_SUITE(DIRECT_CONV2D) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Invalid: Mismatching data type input/weights + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Invalid: Mismatching input feature maps + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Invalid weights dimensions + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Unsupported biases size + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Unsupported biases dimensions + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32, DataLayout::NCHW), // Unsupported data layout: NCHW + TensorInfo(TensorShape(2U, 32U, 16U), 1, DataType::QASYMM8, DataLayout::NHWC), // Unsupported data type: quantized + TensorInfo(TensorShape(2U, 32U, 16U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Arbitrary weight sizes for NHWC are supported + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Non-rectangular weights dimensions for NHWC are supported + TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC), // Strides > 2 for any kernel sizes for NHWC are supported + }), + framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(2U, 3U, 3U, 4U), 1, DataType::F16, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 3U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 4U, 3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U, 3U, 2U, 4U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(2U, 1U, 1U, 4U), 1, DataType::QASYMM8, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 1U, 1U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 13U, 13U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 5U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(2U, 3U, 3U, 4U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("BiasesInfo",{ TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U, 2U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(25U), 1, DataType::F32, DataLayout::NCHW), + TensorInfo(TensorShape(4U), 1, DataType::QASYMM8, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + TensorInfo(TensorShape(4U), 1, DataType::F32, DataLayout::NHWC), + })), + framework::dataset::make("Conv2dAttributes", { + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({1, 1}).pad({0, 0, 0, 0}), + Conv2dAttributes().stride({3, 3}).pad({0, 0, 0, 0}), + })), + framework::dataset::make("Expected", { false, false, false, false, false, false, false, true, true, true, true })), + input_info, weights_info, biases_info, conv2d_attrs, expected) +{ + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + const ITensorInfo* sketch_input_info = context.create_tensor_info(input_info); + const ITensorInfo* sketch_weights_info = context.create_tensor_info(weights_info); + const ITensorInfo* sketch_biases_info = context.create_tensor_info(biases_info); + bool is_valid = bool(GpuConv2d::validate_op(sketch, sketch_input_info, sketch_weights_info, sketch_biases_info, conv2d_attrs)); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +template <typename T> +using DynamicFusionGpuDirectConv2dFixture = DynamicFusionDirectConv2dValidationFixture<CLTensor, CLAccessor, GpuConv2d, T>; + +TEST_SUITE(FP16) +/// TODO: COMPMID-6877: Once the issue in Conv2d is resolved, re-enable these +FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDirectConv2dFixture<half>, framework::DatasetMode::DISABLED, + combine(combine(combine(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U), + TensorShape(32U, 37U, 13U) } ), + framework::dataset::make("StrideX", { 1, 3, 1, 1 })), + framework::dataset::make("StrideY", { 1, 3, 2, 1 })), + framework::dataset::make("PadX", { 1, 3, 0, 4 })), + framework::dataset::make("PadY", { 1, 3, 0, 4 })), + framework::dataset::make("KernelSize", { 3, 8, 1, 9 })), + framework::dataset::make("NumKernels", { 17, 3, 1, 19 })), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDirectConv2dFixture<half>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(800U, 800U, 3U) } ), + framework::dataset::make("StrideX", { 1 })), + framework::dataset::make("StrideY", { 1 })), + framework::dataset::make("PadX", { 1 })), + framework::dataset::make("PadY", { 1 })), + framework::dataset::make("KernelSize", { 9 })), + framework::dataset::make("NumKernels", { 3 })), + framework::dataset::make("DataType", DataType::F16)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +/// TODO: COMPMID-6877: Once the issue in Conv2d is resolved, re-enable these +FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuDirectConv2dFixture<float>, framework::DatasetMode::DISABLED, + combine(combine(combine(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(27U, 13U, 23U), + TensorShape(19U, 5U, 16U, 4U), + TensorShape(13U, 5U, 17U, 2U), + TensorShape(32U, 37U, 13U) } ), + framework::dataset::make("StrideX", { 1, 3, 1, 1 })), + framework::dataset::make("StrideY", { 1, 3, 2, 1 })), + framework::dataset::make("PadX", { 1, 3, 0, 4 })), + framework::dataset::make("PadY", { 1, 3, 0, 4 })), + framework::dataset::make("KernelSize", { 3, 8, 1, 9 })), + framework::dataset::make("NumKernels", { 17, 3, 1, 19 })), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32, 0.0, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuDirectConv2dFixture<float>, framework::DatasetMode::NIGHTLY, + combine(combine(combine(zip(zip(zip(zip(zip( + framework::dataset::make("InputShape", { TensorShape(800U, 800U, 3U) } ), + framework::dataset::make("StrideX", { 1 })), + framework::dataset::make("StrideY", { 1 })), + framework::dataset::make("PadX", { 1 })), + framework::dataset::make("PadY", { 1 })), + framework::dataset::make("KernelSize", { 9 })), + framework::dataset::make("NumKernels", { 3 })), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", DataLayout::NHWC))) +{ + validate(CLAccessor(_target), _reference, tolerance_f32, 0.0, abs_tolerance_f32); +} +// clang-format on +// *INDENT-ON* + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // DIRECT_CONV2D +TEST_SUITE_END() // CONV2D +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/MatMul.cpp b/tests/validation/dynamic_fusion/gpu/cl/MatMul.cpp new file mode 100644 index 0000000000..82d66ca6ce --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/MatMul.cpp @@ -0,0 +1,335 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "tests/AssetsLibrary.h" +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/LargeMatMulDataset.h" +#include "tests/datasets/MatMulDataset.h" +#include "tests/datasets/SmallMatMulDataset.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/MatMulKernelFixture.h" +#include "tests/validation/reference/GEMM.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/Validation.h" + +#include <tuple> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance<float> tolerance_f32( + 0.001f); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +constexpr float abs_tolerance_f32( + 0.0001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for floating point data types in case using relative tolerance fails because of small values */ +constexpr float abs_tolerance_f16( + 0.001f); /**< Absolute tolerance value for comparing reference's output against implementation's output for fp16 data types in case using relative tolerance fails because of small values */ +RelativeTolerance<half_float::half> tolerance_f16(half( + 0.02)); /**< Tolerance value for comparing reference's output against implementation's output for floating point data types */ +} // namespace + +/** M0 values to test - precommit */ +const auto m0_values_lhs_nt_precommit = framework::dataset::make("M0", {1, 2, 3}); + +/** N0 values to test - precommit */ +const auto n0_values_rhs_t_precommit = framework::dataset::make("N0", {1, 2, 4}); + +/** K0 values to test - precommit */ +const auto k0_values_rhs_t_precommit = framework::dataset::make("K0", {1, 2, 4}); + +/** M0 values to test - nightly */ +const auto m0_values_lhs_nt_nightly = framework::dataset::make("M0", {1, 2, 3, 4}); + +/** N0 values to test - nightly */ +const auto n0_values_rhs_t_nightly = framework::dataset::make("N0", {1, 2, 3, 4, 8}); + +/** K0 values to test - nightly */ +const auto k0_values_rhs_t_nightly = framework::dataset::make("K0", {1, 2, 3, 4, 8}); + +class DFMatMulDataset final : public datasets::MatMulDataset +{ +public: + DFMatMulDataset() + { + // LHS = [K, M], RHS = [N, K], DST = [N, M] + add_config(TensorShape(1U, 1U), TensorShape(1U, 1U), TensorShape(1U, 1U)); + add_config(TensorShape(1U, 2U), TensorShape(2U, 1U), TensorShape(2U, 2U)); + add_config(TensorShape(9U, 6U), TensorShape(5U, 9U), TensorShape(5U, 6U)); + add_config(TensorShape(32U, 37U), TensorShape(17U, 32U), TensorShape(17U, 37U)); + } +}; + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) + +TEST_SUITE(MatMul) + +TEST_SUITE(Validate) +TEST_CASE(SupportedBlockSizes, framework::DatasetMode::ALL) +{ + using MatMulConfigurationPair = std::pair<MatMulKernelInfo, bool>; + + const std::vector<MatMulConfigurationPair> supported_block_sizes = { + // MatMulKernelInfo(adj_lhs, adj_rhs, M0, N0, K0, export_rhs_to_cl_image = false) + + // Lhs not-transposed, Rhs transposed + {MatMulKernelInfo(false, true, 0, 1, 1), false}, // M0 should be > 0 + {MatMulKernelInfo(false, true, 3, 11, 1), false}, // N0 not in {1, 2, 3, 4, 8, 16} + {MatMulKernelInfo(false, true, 3, 7, 1), false}, // N0 not in {1, 2, 3, 4, 8, 16} + {MatMulKernelInfo(false, true, 3, 3, 12), false}, // K0 not in {1, 2, 3, 4, 8, 16} + {MatMulKernelInfo(false, true, 3, 3, 6), false}, // K0 not in {1, 2, 3, 4, 8, 16} + {MatMulKernelInfo(false, true, 5, 1, 2), true}, {MatMulKernelInfo(false, true, 3, 3, 3), true}, + {MatMulKernelInfo(false, true, 2, 4, 8), true}, + + }; + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Set big enough shapes so that block sizes are not truncated. Also, set all dimensions equal + // so that it doesn't fail for different NT/T configurations. We aim to test the block sizes here, + // not the shapes themselves. + const ITensorInfo *lhs_info = context.create_tensor_info(TensorInfo(TensorShape(100U, 100U), 1, DataType::F32)); + const ITensorInfo *rhs_info = context.create_tensor_info(TensorInfo(TensorShape(100U, 100U), 1, DataType::F32)); + + for (auto &pair : supported_block_sizes) + { + MatMulAttributes matmul_attr{}; + matmul_attr.adj_lhs(pair.first.adj_lhs); + matmul_attr.adj_rhs(pair.first.adj_rhs); + + GpuMatMulSettings matmul_settings{}; + matmul_settings.m0(pair.first.m0); + matmul_settings.n0(pair.first.n0); + matmul_settings.k0(pair.first.k0); + + Status status = GpuMatMul::validate_op(sketch, lhs_info, rhs_info, matmul_attr, matmul_settings); + ARM_COMPUTE_EXPECT(bool(status) == pair.second, framework::LogLevel::ERRORS); + } +} + +TEST_CASE(ValidateInputShapes, framework::DatasetMode::ALL) +{ + // Create a sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using ShapeConfigurationTuple = std::tuple<TensorShape, TensorShape, bool>; + const std::vector<ShapeConfigurationTuple> shape_configurations = { + {TensorShape(5U, 1U), TensorShape(3U, 5U), true}, + {TensorShape(10U, 12U), TensorShape(3U, 10U), true}, + {TensorShape(8U, 4U), TensorShape(2U, 8U), true}, + {TensorShape(8U, 4U), TensorShape(2U, 5U), false}, // Mismatch in the K dimension + {TensorShape(5U, 0U), TensorShape(2U, 5U), false}, // Invalid dimension + {TensorShape(5U, 4U, 3U, 4U, 5U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), true}, + {TensorShape(5U, 4U, 3U, 4U, 5U, 1U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), false}, // no batch broadcasting + {TensorShape(5U, 4U, 3U, 4U, 9U, 6U), TensorShape(2U, 5U, 3U, 4U, 5U, 6U), + false}, // mismatch in batch dimension + }; + + for (auto &tuple : shape_configurations) + { + const bool expected = std::get<2>(tuple); + + for (bool adj_lhs : {false}) + { + for (bool adj_rhs : {true}) + { + TensorShape lhs_shape = std::get<0>(tuple); + TensorShape rhs_shape = std::get<1>(tuple); + + if (adj_lhs) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + + if (adj_rhs) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + const ITensorInfo *lhs_info = context.create_tensor_info(TensorInfo(lhs_shape, 1, DataType::F32)); + const ITensorInfo *rhs_info = context.create_tensor_info(TensorInfo(rhs_shape, 1, DataType::F32)); + + MatMulAttributes matmul_attr{}; + matmul_attr.adj_lhs(adj_lhs); + matmul_attr.adj_rhs(adj_rhs); + + GpuMatMulSettings matmul_settings{}; + matmul_settings.m0(1); + matmul_settings.n0(1); + matmul_settings.k0(1); + + Status status = GpuMatMul::validate_op(sketch, lhs_info, rhs_info, matmul_attr, matmul_settings); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } + } + } +} + +TEST_CASE(ValidateDataTypes, framework::DatasetMode::ALL) +{ + // Configurations are assumed to be Nt/Nt, but will be transposed inside the test to test other configurations + using DataTypeConfigurationTuple = std::tuple<DataType, DataType, DataType, bool>; + const std::vector<DataTypeConfigurationTuple> data_type_configurations = { + {DataType::F32, DataType::F32, DataType::F32, true}, + {DataType::F16, DataType::F16, DataType::F16, true}, + {DataType::F16, DataType::F32, DataType::F32, false}, // no mixed precision + {DataType::F64, DataType::F64, DataType::F64, false}, // no double precision + {DataType::QASYMM8, DataType::QASYMM8, DataType::QASYMM8, false}, // no quantized types + {DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, false}, // no quantized types + {DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, DataType::QSYMM8_PER_CHANNEL, + false}, // no quantized types + {DataType::QASYMM16, DataType::QASYMM16, DataType::QASYMM16, false}, // no quantized types + {DataType::QSYMM16, DataType::QSYMM16, DataType::QSYMM16, false}, // no quantized types + {DataType::QSYMM8, DataType::QSYMM8, DataType::QSYMM8, false}, // no quantized types + {DataType::S64, DataType::S64, DataType::S64, false}, // no integral types + {DataType::S32, DataType::S32, DataType::S32, false}, // no integral types + {DataType::S16, DataType::S16, DataType::S16, false}, // no integral types + {DataType::S8, DataType::S8, DataType::S8, false}, // no integral types + {DataType::U64, DataType::U64, DataType::U64, false}, // no integral types + {DataType::U32, DataType::U32, DataType::U32, false}, // no integral types + {DataType::U16, DataType::U16, DataType::U16, false}, // no integral types + {DataType::U8, DataType::U8, DataType::U8, false}, // no integral types + }; + // Create a sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + const TensorShape shape = TensorShape(10U, 10U); + MatMulAttributes matmul_attr{}; + matmul_attr.adj_lhs(false); + matmul_attr.adj_rhs(false); + GpuMatMulSettings matmul_settings{}; + matmul_settings.m0(1); + matmul_settings.n0(1); + matmul_settings.k0(1); + + for (auto &tuple : data_type_configurations) + { + const bool expected = std::get<3>(tuple); + + const ITensorInfo *lhs_info = context.create_tensor_info(TensorInfo(shape, 1, std::get<0>(tuple))); + const ITensorInfo *rhs_info = context.create_tensor_info(TensorInfo(shape, 1, std::get<1>(tuple))); + + Status status = GpuMatMul::validate_op(sketch, lhs_info, rhs_info, matmul_attr, matmul_settings); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); + } +} + +TEST_SUITE_END() // Validate + +template <typename T> +using DynamicFusionGpuMatmulFixture = DynamicFusionGpuMatMulValidationFixture<CLTensor, CLAccessor, GpuMatMul, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP32) + +FIXTURE_DATA_TEST_CASE(RunPrecommit, + DynamicFusionGpuMatmulFixture<float>, + framework::DatasetMode::ALL, + combine(DFMatMulDataset(), + framework::dataset::make("TransposeA", {false}), + framework::dataset::make("TransposeB", {true}), + m0_values_lhs_nt_precommit, + n0_values_rhs_t_precommit, + k0_values_rhs_t_precommit, + framework::dataset::make("ExportRhsToCLImage", {false}), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunNightly, + DynamicFusionGpuMatmulFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(DFMatMulDataset(), + framework::dataset::make("TransposeA", {false}), + framework::dataset::make("TransposeB", {true}), + m0_values_lhs_nt_nightly, + n0_values_rhs_t_nightly, + k0_values_rhs_t_nightly, + framework::dataset::make("ExportRhsToCLImage", {false}), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32, 0.f, abs_tolerance_f32); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) + +FIXTURE_DATA_TEST_CASE(RunPrecommit, + DynamicFusionGpuMatmulFixture<half>, + framework::DatasetMode::ALL, + combine(DFMatMulDataset(), + framework::dataset::make("TransposeA", {false}), + framework::dataset::make("TransposeB", {true}), + m0_values_lhs_nt_precommit, + n0_values_rhs_t_precommit, + k0_values_rhs_t_precommit, + framework::dataset::make("ExportRhsToCLImage", {false}), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunNightly, + DynamicFusionGpuMatmulFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(DFMatMulDataset(), + framework::dataset::make("TransposeA", {false}), + framework::dataset::make("TransposeB", {true}), + m0_values_lhs_nt_nightly, + n0_values_rhs_t_nightly, + k0_values_rhs_t_nightly, + framework::dataset::make("ExportRhsToCLImage", {false}), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16, 0.f, abs_tolerance_f16); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE_END() // Float +TEST_SUITE_END() // MatMul +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp b/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp new file mode 100644 index 0000000000..af02ce3eaa --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Mul.cpp @@ -0,0 +1,221 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuMul.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/DynamicFusionDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/* Synced with tests/validation/CL/PixelwiseMultiplication.cpp from the standard interface. + * + * Difference | Why the difference + * No integer tests | Not supported yet + * No quantized tests | Not supported yet + * No convert policy tests | Not needed as convert policy is ignored by floating types + * No scale tests | Not supported yet + * No rounding modes tests | Not supported yet + * No in place tests | Not supported yet + * No activation tests | Not needed in dynamic fusion interface + * + */ +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_f16( + 0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr AbsoluteTolerance<float> tolerance_f32( + 0.0001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ +} // namespace +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(MUL) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( + framework::dataset::make("LhsInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), // Unsupported data type U8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), // Unsupported data type S8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), // Unsupported data type S16 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), // Unsupported data type S32 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8_SIGNED), // Unsupported data type QASYMM8_SIGNED + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(32U, 1U, 1U), 1, DataType::F32), // Broadcasting allowed for lhs + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(15U, 23U, 3U), 1, DataType::F32), // Broadcast Y dimension is not allowed + TensorInfo(TensorShape( 3U, 8U, 9U), 1, DataType::F32), // Broadcast Z dimension is not allowed + TensorInfo(TensorShape(32U, 13U, 2U, 2), 1, DataType::F32), // Batching is allowed + }), + framework::dataset::make("RhsInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U8), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S8), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 1U, 1U), 1, DataType::F32), // Broadcasting allowed for rhs + TensorInfo(TensorShape(15U, 1U, 3U), 1, DataType::F32), + TensorInfo(TensorShape( 3U, 8U, 1U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U, 2), 1, DataType::F32), + })), + framework::dataset::make("Expected", { true, true, false, false, false, false, false, false, false, false, true, true, false, false, true })), + input1_info, input2_info, expected) +{ + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Validate Elementwise Mul + auto lhs_info = context.create_tensor_info(input1_info); + auto rhs_info = context.create_tensor_info(input2_info); + + bool res = bool(GpuMul::validate_op(sketch, lhs_info, rhs_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using DynamicFusionCLMulFixture = DynamicFusionMulOneOpValidationFixture<CLTensor, CLAccessor, GpuMul, T>; +template <typename T> +using DynamicFusionCLMulBroadcastFixture = DynamicFusionMulBroadcastValidationFixture<CLTensor, CLAccessor, GpuMul, T>; +template <typename T> +using DynamicFusionCLMulTwoOpsFixture = DynamicFusionMulTwoOpsValidationFixture<CLTensor, CLAccessor, GpuMul, T>; + +TEST_SUITE(F16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionCLMulFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmallBroadcastOneOp, + DynamicFusionCLMulBroadcastFixture<half>, + framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::TemporaryLimitedSmallShapesBroadcast(), + framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunLargeBroadcastOneOp, + DynamicFusionCLMulBroadcastFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(combine(datasets::TemporaryLimitedLargeShapesBroadcast(), + framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // F16 + +TEST_SUITE(F32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionCLMulFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLargeOneOp, + DynamicFusionCLMulFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmallBroadcastOneOp, + DynamicFusionCLMulBroadcastFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::TemporaryLimitedSmallShapesBroadcast(), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLargeBroadcastOneOp, + DynamicFusionCLMulBroadcastFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(datasets::TemporaryLimitedLargeShapesBroadcast(), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionCLMulTwoOpsFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(datasets::DynamicFusionElementwiseBinaryTwoOpsSmallShapes(), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false})), + framework::dataset::make("FuseTwoOps", {true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // F32 + +TEST_SUITE_END() // MUL +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp new file mode 100644 index 0000000000..be816b32b3 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp @@ -0,0 +1,219 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuPool2d.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/dynamic_fusion/PoolingLayerDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(POOL2D) + +constexpr AbsoluteTolerance<float> tolerance_f32( + 0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */ +constexpr AbsoluteTolerance<float> tolerance_f16( + 0.01f); /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */ + +const auto PoolingLayerDatasetFP = + combine(combine(combine(combine(framework::dataset::make("PoolingType", {PoolingType::MAX, PoolingType::AVG}), + framework::dataset::make("PoolingSize", {Size2D(2, 2), Size2D(3, 3)})), + framework::dataset::make("Pad", {Padding2D()})), + framework::dataset::make("Stride", {Size2D(1, 1), Size2D(2, 1), Size2D(5, 7)})), + framework::dataset::make("ExcludePadding", {true})); + +template <typename T> +using DynamicFusionGpuPool2dFixture = DynamicFusionGpuPool2dValidationFixture<CLTensor, CLAccessor, GpuPool2d, T>; + +template <typename T> +using DFSpecialGpuPool2dFixture = DynamicFusionGpuPool2dSpecialValidationFixture<CLTensor, CLAccessor, GpuPool2d, T>; +// *INDENT-OFF* +// clang-format off + +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC), // Invalid parameters, unsupported pooling + TensorInfo(TensorShape(5U, 15U, 13U), 1, DataType::F32, DataLayout::NHWC), // Valid Non-rectangular Global Pooling + TensorInfo(TensorShape(5U, 13U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC), // Invalid - Quantized not supported. + TensorInfo(TensorShape(5U, 13U, 13U), 1, DataType::F32, DataLayout::NHWC), // Valid global pooling + TensorInfo(TensorShape(13U, 13U, 5U), 1, DataType::F32, DataLayout::NCHW), // Unsupported data layout + }), + framework::dataset::make("Pool2dAttributes", { + Pool2dAttributes().pool_type(PoolingType::L2).pool_size(Size2D(3,3)).pad(Padding2D(0,0,0,0)).stride(Size2D(1,1)), + Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(15U, 13U)), + Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(2,2)).pad(Padding2D()).stride(Size2D(1,1)), + Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(13U,13U)), + Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(13U,13U)), + })), + framework::dataset::make("Expected", { false, true, false, true, false })), + input_info, pool2d_attr, expected) +{ + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Declare GpuPool2d settings + const GpuPool2dSettings &settings = GpuPool2dSettings(); + + // Validate Pool2d Configuration + auto src_info = context.create_tensor_info(input_info); + bool res = bool(GpuPool2d::validate_op(sketch, src_info, pool2d_attr, settings)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} + +// clang-format on +// *INDENT-ON* + +TEST_SUITE(Float) +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuPool2dFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(datasets::SmallNoneUnitShapes(), PoolingLayerDatasetFP), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionGpuPool2dFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(datasets::LargeShapes(), PoolingLayerDatasetFP), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +FIXTURE_DATA_TEST_CASE(RunSpecial, + DFSpecialGpuPool2dFixture<float>, + framework::DatasetMode::ALL, + combine(datasets::PoolingLayerDatasetSpecialDynamicFusion(), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE(GlobalPooling) +FIXTURE_DATA_TEST_CASE( + RunSmall, + DynamicFusionGpuPool2dFixture<float>, + framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(framework::dataset::make("InputShape", + {TensorShape(27U, 13U, 2U), + TensorShape(27U, 13U, 2U, 4U)}), + framework::dataset::make("PoolingType", + {PoolingType::AVG, PoolingType::MAX})), + framework::dataset::make("PoolingSize", {Size2D(27, 13)})), + framework::dataset::make("Pad", {Padding2D()})), + framework::dataset::make("Stride", {Size2D(1, 1)})), + framework::dataset::make("ExcludePadding", true)), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE( + RunLarge, + DynamicFusionGpuPool2dFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(framework::dataset::make("InputShape", + {TensorShape(79U, 37U, 11U), + TensorShape(79U, 37U, 11U, 4U)}), + framework::dataset::make("PoolingType", + {PoolingType::AVG, PoolingType::MAX})), + framework::dataset::make("PoolingSize", {Size2D(79, 37)})), + framework::dataset::make("Pad", {Padding2D()})), + framework::dataset::make("Stride", {Size2D(1, 1)})), + framework::dataset::make("ExcludePadding", true)), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // GlobalPooling +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +TEST_SUITE(GlobalPooling) +FIXTURE_DATA_TEST_CASE( + RunSmall, + DynamicFusionGpuPool2dFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(combine(combine(combine(framework::dataset::make("InputShape", + {TensorShape(27U, 13U, 2U), + TensorShape(27U, 13U, 2U, 4U)}), + framework::dataset::make("PoolingType", + {PoolingType::AVG, PoolingType::MAX})), + framework::dataset::make("PoolingSize", {Size2D(27, 13)})), + framework::dataset::make("Pad", {Padding2D()})), + framework::dataset::make("Stride", {Size2D(1, 1)})), + framework::dataset::make("ExcludePadding", true)), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE( + RunLarge, + DynamicFusionGpuPool2dFixture<half>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(combine(combine(combine(framework::dataset::make("InputShape", + {TensorShape(79U, 37U, 11U), + TensorShape(79U, 37U, 11U, 4U)}), + framework::dataset::make("PoolingType", + {PoolingType::AVG, PoolingType::MAX})), + framework::dataset::make("PoolingSize", {Size2D(79, 37)})), + framework::dataset::make("Pad", {Padding2D()})), + framework::dataset::make("Stride", {Size2D(1, 1)})), + framework::dataset::make("ExcludePadding", true)), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // GlobalPooling +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // FLOAT + +TEST_SUITE_END() // POOL2D +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp b/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp new file mode 100644 index 0000000000..d46754ccca --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Reshape.cpp @@ -0,0 +1,147 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ReshapeLayerDataset.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(RESHAPE) + +DATA_TEST_CASE(Validate, + framework::DatasetMode::DISABLED, + zip(zip(framework::dataset::make( + "InputInfo", + { + TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), + TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32) /*mismatching dimensions*/, + }), + framework::dataset::make("OutputShape", + { + TensorShape(9U, 5U, 21U), + TensorShape(8U, 24U, 4U), + TensorShape(192U, 192U), + })), + framework::dataset::make("Expected", {true, true, false})), + input_info, + output_shape, + expected) +{ + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + TensorShape input_shape = input_info.tensor_shape(); + ARM_COMPUTE_UNUSED(input_shape); + ITensorInfo *src_info = context.create_tensor_info(input_info); + + ReshapeAttributes attributes; + attributes.shape(output_shape); + Status status = GpuReshape::validate_op(sketch, src_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); +} + +template <typename T> +using DynamicFusionGpuReshapeLayerFixture = + DynamicFusionGpuReshapeLayerValidationFixture<CLTensor, CLAccessor, GpuReshape, T>; + +TEST_SUITE(F32) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuReshapeLayerFixture<float>, + framework::DatasetMode::DISABLED, + combine(datasets::SmallReshapeLayerDataset(), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // F32 + +TEST_SUITE(F16) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuReshapeLayerFixture<half>, + framework::DatasetMode::DISABLED, + combine(datasets::SmallReshapeLayerDataset(), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // F16 + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuReshapeLayerFixture<uint8_t>, + framework::DatasetMode::DISABLED, + combine(datasets::SmallReshapeLayerDataset(), + framework::dataset::make("DataType", DataType::U8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // U8 + +TEST_SUITE(S8) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuReshapeLayerFixture<int8_t>, + framework::DatasetMode::DISABLED, + combine(datasets::SmallReshapeLayerDataset(), + framework::dataset::make("DataType", DataType::S8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // S8 + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionGpuReshapeLayerFixture<int16_t>, + framework::DatasetMode::DISABLED, + combine(datasets::SmallReshapeLayerDataset(), + framework::dataset::make("DataType", DataType::S16))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // S16 + +TEST_SUITE_END() // RESHAPE +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp b/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp new file mode 100644 index 0000000000..a6bcf4ae26 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Resize.cpp @@ -0,0 +1,359 @@ +/* +* Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuResize.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ScaleValidationDataset.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +using datasets::ScaleAlignCornersSamplingPolicySet; +using datasets::ScaleInterpolationPolicySet; +using datasets::ScaleSamplingPolicySet; +using datasets::ScaleShapesBaseDataSet; + +/** We consider vector size in byte 16 since the maximum size of + * a vector used by @ref CLScaleKernel is currently 16-byte (float4). + */ +constexpr uint32_t vector_byte = 16; + +template <typename T> +constexpr uint32_t num_elements_per_vector() +{ + return vector_byte / sizeof(T); +} + +/** Quantization information data set */ +const auto QuantizationInfoSet = framework::dataset::make("QuantizationInfo", + { + QuantizationInfo(0.5f, -1), + }); + +/** Tolerance */ +constexpr float tolerance_f32_absolute(0.001f); + +RelativeTolerance<float> tolerance_f32(0.05); +constexpr float abs_tolerance_f16(0.1f); +RelativeTolerance<half> tolerance_f16(half(0.1)); + +constexpr float tolerance_num_f32(0.01f); + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(RESIZE) + +TEST_SUITE(Validate) + +const auto default_input_shape = TensorShape{2, 3, 3, 2}; +const auto default_output_shape = TensorShape{4, 6, 3, 2}; + +constexpr auto default_data_type = DataType::U8; +constexpr auto default_data_layout = DataLayout::NHWC; + +TEST_CASE(NullPtr, framework::DatasetMode::ALL) +{ + const TensorInfo input_info = TensorInfo{default_input_shape, 1, default_data_type, default_data_layout}; + const TensorInfo output_info = TensorInfo{default_output_shape, 1, default_data_type, default_data_layout}; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // nullptr is given as input + Status status = GpuResize::validate_op(sketch, nullptr, ResizeAttributes()); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(SupportDataType, framework::DatasetMode::ALL) +{ + const std::map<DataType, bool> supported_data_types = + { + { DataType::U8, false }, + { DataType::S8, false }, + { DataType::QSYMM8, false }, + { DataType::QASYMM8, false }, + { DataType::QASYMM8_SIGNED, false }, + { DataType::QSYMM8_PER_CHANNEL, false }, + { DataType::U16, false }, + { DataType::S16, false }, + { DataType::QSYMM16, false }, + { DataType::QASYMM16, false }, + { DataType::U32, false }, + { DataType::S32, false }, + { DataType::U64, false }, + { DataType::S64, false }, + { DataType::BFLOAT16, false }, + { DataType::F16, true }, + { DataType::F32, true }, + { DataType::F64, false }, + { DataType::SIZET, false }, + }; + + for (auto &kv : supported_data_types) + { + const TensorInfo input_info = TensorInfo{default_input_shape, 1, kv.first, default_data_layout}; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + const ITensorInfo *sketch_input_info = context.create_tensor_info(input_info); + + ResizeAttributes attributes; + attributes.output_width(default_output_shape[0]); // shape is not important unless it's empty + attributes.output_height(default_output_shape[1]); + + Status status = GpuResize::validate_op(sketch, sketch_input_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == kv.second, framework::LogLevel::ERRORS); + } +} + +TEST_CASE(MismatchingDataType, framework::DatasetMode::ALL) +{ + constexpr DataType non_default_data_type = DataType::F32; + + const TensorInfo input_info = TensorInfo{default_input_shape, 1, default_data_type, default_data_layout}; + const TensorInfo output_info = TensorInfo{default_output_shape, 1, non_default_data_type, default_data_layout}; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + const ITensorInfo *sketch_input_info = context.create_tensor_info(input_info); + + Status status = GpuResize::validate_op(sketch, sketch_input_info, ResizeAttributes()); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(AlignedCornerNotSupported, framework::DatasetMode::ALL) +{ + // Aligned corners require sampling policy to be TOP_LEFT. + constexpr InterpolationPolicy interpolation_policy = InterpolationPolicy::BILINEAR; + constexpr bool align_corners = true; + constexpr SamplingPolicy sampling_policy = SamplingPolicy::CENTER; + + const TensorInfo input_info = TensorInfo{default_input_shape, 1, default_data_type, default_data_layout}; + const TensorInfo output_info = TensorInfo{default_output_shape, 1, default_data_type, default_data_layout}; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + const ITensorInfo *sketch_input_info = context.create_tensor_info(input_info); + + ResizeAttributes attributes{}; + attributes.interpolation_policy(interpolation_policy).sampling_policy(sampling_policy).align_corners(align_corners); + + Status status = GpuResize::validate_op(sketch, sketch_input_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(UnsupportedInterpolationPolicy, framework::DatasetMode::ALL) +{ + const TensorInfo input_info = TensorInfo{TensorShape(28U, 33U, 2U), 1, DataType::F32, default_data_layout}; + const TensorInfo output_info = TensorInfo{TensorShape(26U, 21U, 2U), 1, DataType::F32, default_data_layout}; + constexpr auto interpolation_policy = InterpolationPolicy::AREA; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + const ITensorInfo *sketch_input_info = context.create_tensor_info(input_info); + + ResizeAttributes attributes{}; + attributes.interpolation_policy(interpolation_policy); + + Status status = GpuResize::validate_op(sketch, sketch_input_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_CASE(UnsupportedLayout, framework::DatasetMode::ALL) +{ + const TensorInfo input_info = TensorInfo{default_input_shape, 1, default_data_type, DataLayout::NCHW}; + const TensorInfo output_info = TensorInfo{default_output_shape, 1, default_data_type, DataLayout::NCHW}; + constexpr auto interpolation_policy = InterpolationPolicy::BILINEAR; + + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + const ITensorInfo *sketch_input_info = context.create_tensor_info(input_info); + + ResizeAttributes attributes{}; + attributes.interpolation_policy(interpolation_policy); + + Status status = GpuResize::validate_op(sketch, sketch_input_info, attributes); + ARM_COMPUTE_EXPECT(bool(status) == false, framework::LogLevel::ERRORS); +} + +TEST_SUITE_END() // Validate + +template <typename T> +using DynamicFusionResizeFixture = DynamicFusionResizeValidationFixture<CLTensor, CLAccessor, GpuResize, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP32) + +const auto f32_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector<float>())), + framework::dataset::make("DataType", DataType::F32)); + +FIXTURE_DATA_TEST_CASE(Run, + DynamicFusionResizeFixture<float>, + framework::DatasetMode::ALL, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} + +FIXTURE_DATA_TEST_CASE(RunAlignCorners, + DynamicFusionResizeFixture<float>, + framework::DatasetMode::ALL, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} +const auto f32_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector<float>())), + framework::dataset::make("DataType", DataType::F32)); +FIXTURE_DATA_TEST_CASE(RunNightly, + DynamicFusionResizeFixture<float>, + framework::DatasetMode::NIGHTLY, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_nightly_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, + DynamicFusionResizeFixture<float>, + framework::DatasetMode::NIGHTLY, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f32_nightly_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f32, tolerance_num_f32, tolerance_f32_absolute); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +const auto f16_shape = combine((SCALE_PRECOMMIT_SHAPE_DATASET(num_elements_per_vector<half>())), + framework::dataset::make("DataType", DataType::F16)); +FIXTURE_DATA_TEST_CASE(Run, + DynamicFusionResizeFixture<half>, + framework::DatasetMode::ALL, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunAlignCorners, + DynamicFusionResizeFixture<half>, + framework::DatasetMode::ALL, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +const auto f16_nightly_shape = combine((SCALE_NIGHTLY_SHAPE_DATASET(num_elements_per_vector<half>())), + framework::dataset::make("DataType", DataType::F16)); +FIXTURE_DATA_TEST_CASE(RunNightly, + DynamicFusionResizeFixture<half>, + framework::DatasetMode::NIGHTLY, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_nightly_shape, ScaleSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunNightlyAlignCorners, + DynamicFusionResizeFixture<half>, + framework::DatasetMode::NIGHTLY, + ASSEMBLE_DATASET_DYNAMIC_FUSION(f16_nightly_shape, ScaleAlignCornersSamplingPolicySet)) +{ + //Create valid region + TensorInfo src_info(_shape, 1, _data_type); + const ValidRegion valid_region = + calculate_valid_region_scale(src_info, _reference.shape(), _interpolation_policy, _sampling_policy, false); + + // Validate output + validate(CLAccessor(_target), _reference, valid_region, tolerance_f16, 0.0f, abs_tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // Float + +TEST_SUITE_END() // RESIZE +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp b/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp new file mode 100644 index 0000000000..0134a7c11b --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Sigmoid.cpp @@ -0,0 +1,154 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSigmoid.h" +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_f32(1e-6f); +constexpr AbsoluteTolerance<float> tolerance_f16(0.001f); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(SIGMOID) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type + }), + framework::dataset::make("Expected", { true, true, false })), + input_info, expected) +{ + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Fuse sigmoid + const ITensorInfo *src_info = context.create_tensor_info(input_info); + + const bool res = static_cast<bool>(GpuSigmoid::validate_op(sketch, src_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using DynamicFusionSigmoidOpFixture = DynamicFusionSigmoidValidationFixture<CLTensor, CLAccessor, GpuSigmoid, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionSigmoidOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionSigmoidOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + ARM_COMPUTE_TEST_INFO("Currently 5D+ tensors are unsupported for this operation."); + framework::ARM_COMPUTE_PRINT_INFO(); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionSigmoidOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {true})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionSigmoidOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionSigmoidOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + ARM_COMPUTE_TEST_INFO("Currently 5D+ tensors are unsupported for this operation."); + framework::ARM_COMPUTE_PRINT_INFO(); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionSigmoidOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {true})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float + +TEST_SUITE_END() // SIGMOID +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp b/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp new file mode 100644 index 0000000000..8f5a1ed14a --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Softmax.cpp @@ -0,0 +1,219 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSoftmax.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/** Tolerance for float operations */ +RelativeTolerance<half> tolerance_f16(half(0.2)); +RelativeTolerance<float> tolerance_f32(0.001f); + +using framework::dataset::make; + +/// TODO: COMPMID-6713 +/// Softmax is not implemented in CKW. Therefore, the tests are DISABLED. +/// Enable the tests when Softmax is implemented in CKW. + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(SOFTMAX) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::DISABLED, + zip( + make("InputInfo", { + TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching data types + TensorInfo(TensorShape(27U, 13U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::S32), // Unsupported data type + TensorInfo(TensorShape(32U, 13U), 1, DataType::F16), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + }), + make("OutputInfo",{ + TensorInfo(TensorShape(27U, 13U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 11U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::QASYMM16), // Unsupported data type + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U), 1, DataType::F32), + }), + make("beta", { + 1.0, + 2.0, + 2.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + }), + make("axis", { + 0, + 0, + 1, // Invalid as axis != 0 + 0, + 0, + 0, + -3, // Invalid as axis != 0 + 2, // Invalid as axis != 0 + 1, // Invalid as axis != 0 + -1, // Invalid as axis != 0 + }), + make("Expected", { false, false, false, true, false, false, false, false, false, false})), + input_info, output_info, beta, axis, expected) +{ + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + SoftmaxAttributes softmax_attr{}; + softmax_attr.axis(axis).beta(beta).is_log_softmax(false); + ITensorInfo* src_info = context.create_tensor_info(input_info); + ITensorInfo* dst_info = context.create_tensor_info(output_info); + const bool res = static_cast<bool>(GpuSoftmax::validate_op(sketch, src_info, dst_info, softmax_attr)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} + +template <typename T> +using DynamicFusionSoftmaxLayerFixture = DynamicFusionSoftmaxValidationFixture<CLTensor, CLAccessor, GpuSoftmax, T>; + +TEST_SUITE(FLOAT) +TEST_SUITE(FP32) + +FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionSoftmaxLayerFixture<float>, framework::DatasetMode::DISABLED, + combine( + datasets::SoftmaxLayerSmallShapes(), + make("DataType", DataType::F32), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }), + make("is_log", {false, true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + + +FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionSoftmaxLayerFixture<float>, framework::DatasetMode::DISABLED, + combine( + datasets::SoftmaxLayerLargeShapes(), + make("DataType", DataType::F32), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }), + make("is_log", {false, true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + + +FIXTURE_DATA_TEST_CASE(Run4D, DynamicFusionSoftmaxLayerFixture<float>, framework::DatasetMode::DISABLED, + combine( + datasets::SoftmaxLayer4DShapes(), + make("DataType", DataType::F32), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }), + make("is_log", {false, true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // FP32 +TEST_SUITE(FP16) + +FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionSoftmaxLayerFixture<half>, framework::DatasetMode::DISABLED, + combine( + datasets::SoftmaxLayerSmallShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }), + make("is_log", {false, true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + + +FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionSoftmaxLayerFixture<half>, framework::DatasetMode::DISABLED, + combine( + datasets::SoftmaxLayerLargeShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }), + make("is_log", {false, true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + + +FIXTURE_DATA_TEST_CASE(Run4D, DynamicFusionSoftmaxLayerFixture<half>, framework::DatasetMode::DISABLED, + combine( + datasets::SoftmaxLayer4DShapes(), + make("DataType", DataType::F16), + make("Beta", { 1.0f, 2.0f }), + make("Axis", { 0 }), + make("is_log", {false, true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() // FP16 +TEST_SUITE_END() // FLOAT + +TEST_SUITE_END() // SOFTMAX +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp b/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp new file mode 100644 index 0000000000..c7ab1e717c --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Sub.cpp @@ -0,0 +1,262 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuSub.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/DynamicFusionDataset.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/* Synced with tests/validation/CL/ArithmeticSubtraction.cpp from the standard interface. + * + * Difference | Why the difference + * No quantized tests | Not supported yet + * No in place tests | Not supported yet + * No activation tests | Not needed in dynamic fusion interface + * + */ +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(SUB) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( + framework::dataset::make("LhsInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U32), // Unsupported data type U32 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8_SIGNED), // Unsupported data type QASYMM8 + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching shapes + TensorInfo(TensorShape(32U, 1U, 1U), 1, DataType::F32), // Broadcasting allowed for lhs + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(15U, 23U, 3U), 1, DataType::F32), // Broadcast Y dimension is not allowed + TensorInfo(TensorShape( 3U, 8U, 9U), 1, DataType::S16), // Invalid data type combination + TensorInfo(TensorShape(32U, 13U, 2U, 2), 1, DataType::F32), // Batching is allowed + }), + framework::dataset::make("RhsInfo",{ TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::U32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QASYMM8_SIGNED), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S16), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::S32), + TensorInfo(TensorShape(48U, 11U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 1U, 1U), 1, DataType::F32), // Broadcasting allowed for rhs + TensorInfo(TensorShape(15U, 1U, 3U), 1, DataType::F32), + TensorInfo(TensorShape( 3U, 8U, 1U), 1, DataType::S16), + TensorInfo(TensorShape(32U, 13U, 2U, 2), 1, DataType::F32), + })), + framework::dataset::make("Expected", { true, false, false, false, false, false, false, false, true, true, false, false, true })), + input1_info, input2_info, expected) +{ + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Validate Elementwise Sub + auto lhs_info = context.create_tensor_info(input1_info); + auto rhs_info = context.create_tensor_info(input2_info); + + bool res = bool(GpuSub::validate_op(sketch, lhs_info, rhs_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using DynamicFusionCLSubFixture = + DynamicFusionGpuElementwiseBinaryOneOpValidationFixture<CLTensor, CLAccessor, GpuSub, T>; + +template <typename T> +using DynamicFusionCLSubBroadcastFixture = + DynamicFusionGpuElementwiseBinaryBroadcastOneOpValidationFixture<CLTensor, CLAccessor, GpuSub, T>; + +template <typename T> +using DynamicFusionCLSubTwoOpsFixture = + DynamicFusionGpuElementwiseBinaryTwoOpsValidationFixture<CLTensor, CLAccessor, GpuSub, T>; + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionCLSubFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLargeOneOp, + DynamicFusionCLSubFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::LargeShapes()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunSmallBroadcastOneOp, + DynamicFusionCLSubBroadcastFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::TemporaryLimitedSmallShapesBroadcast()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLargeBroadcastOneOp, + DynamicFusionCLSubBroadcastFixture<float>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::TemporaryLimitedLargeShapesBroadcast()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE( + RunSmallTwoOps, + DynamicFusionCLSubTwoOpsFixture<float>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::DynamicFusionElementwiseBinaryTwoOpsSmallShapes()), + framework::dataset::make("DataType", {DataType::F32})), + framework::dataset::make("InPlace", {false})), + framework::dataset::make("FuseTwoOps", {true}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // FP32 + +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionCLSubFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunSmallBroadcastOneOp, + DynamicFusionCLSubBroadcastFixture<half>, + framework::DatasetMode::ALL, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::TemporaryLimitedSmallShapesBroadcast()), + framework::dataset::make("DataType", {DataType::F16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(S32) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionCLSubFixture<int32_t>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::S32})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // S32 + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionCLSubFixture<int16_t>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::S16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, + DynamicFusionCLSubFixture<int16_t>, + framework::DatasetMode::NIGHTLY, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::LargeShapes()), + framework::dataset::make("DataType", {DataType::S16})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // S16 + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, + DynamicFusionCLSubFixture<uint8_t>, + framework::DatasetMode::PRECOMMIT, + combine(combine(combine(framework::dataset::make("ElementwiseOp", {ArithmeticOperation::SUB}), + datasets::SmallShapes()), + framework::dataset::make("DataType", {DataType::U8})), + framework::dataset::make("InPlace", {false}))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // U8 + +TEST_SUITE_END() // SUB +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp b/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp new file mode 100644 index 0000000000..2560f3aab1 --- /dev/null +++ b/tests/validation/dynamic_fusion/gpu/cl/Tanh.cpp @@ -0,0 +1,154 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuTanh.h" +#include "arm_compute/runtime/CL/CLTensor.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h" +#include "tests/validation/Validation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_f32(0.00001f); +constexpr AbsoluteTolerance<float> tolerance_f16(0.001f); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(DYNAMIC_FUSION) +TEST_SUITE(TANH) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), + TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::QASYMM8), // Unsupported data type + }), + framework::dataset::make("Expected", { true, true, false })), + input_info, expected) +{ + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Fuse tanh + const ITensorInfo* src_info = context.create_tensor_info(input_info); + + const bool res = static_cast<bool>(GpuTanh::validate_op(sketch, src_info)); + ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +template <typename T> +using DynamicFusionTanhOpFixture = DynamicFusionTanhValidationFixture<CLTensor, CLAccessor, GpuTanh, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionTanhOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionTanhOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + ARM_COMPUTE_TEST_INFO("Currently 5D+ tensors are unsupported for this operation."); + framework::ARM_COMPUTE_PRINT_INFO(); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionTanhOpFixture<half>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {true})), + framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f16); +} + +TEST_SUITE_END() // FP16 + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSmallOneOp, + DynamicFusionTanhOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunSmall5dOneOp, + DynamicFusionTanhOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(datasets::Small5dShapes(), framework::dataset::make("Fuse", {false})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + ARM_COMPUTE_TEST_INFO("Currently 5D+ tensors are unsupported for this operation."); + framework::ARM_COMPUTE_PRINT_INFO(); +} + +FIXTURE_DATA_TEST_CASE(RunSmallTwoOps, + DynamicFusionTanhOpFixture<float>, + framework::DatasetMode::ALL, + combine(combine(datasets::SmallShapes(), framework::dataset::make("Fuse", {true})), + framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +TEST_SUITE_END() // FP32 +TEST_SUITE_END() // Float + +TEST_SUITE_END() // TANH +TEST_SUITE_END() // DYNAMIC_FUSION +TEST_SUITE_END() // CL +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/ActivationLayerFixture.h b/tests/validation/fixtures/ActivationLayerFixture.h index 531b3abdf9..a24ba8913e 100644 --- a/tests/validation/fixtures/ActivationLayerFixture.h +++ b/tests/validation/fixtures/ActivationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_ACTIVATION_LAYER_FIXTURE -#define ARM_COMPUTE_TEST_ACTIVATION_LAYER_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_ACTIVATIONLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_ACTIVATIONLAYERFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -47,12 +47,7 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ActivationValidationGenericFixture : public framework::Fixture { public: - ActivationValidationGenericFixture() - : _target(parameters->get_ctx<TensorType>()) - { - } - template <typename...> void setup(TensorShape shape, bool in_place, ActivationLayerInfo::ActivationFunction function, float alpha_beta, DataType data_type, QuantizationInfo quantization_info) { ActivationLayerInfo info(function, alpha_beta, alpha_beta); @@ -120,13 +115,12 @@ protected: TensorType compute_target(const TensorShape &shape, ActivationLayerInfo info) { - auto ctx = parameters->get_ctx<TensorType>(); // Create tensors - TensorType src = create_tensor<TensorType>(shape, _data_type, 1, _input_quantization_info, DataLayout::NCHW, ctx); - TensorType dst = create_tensor<TensorType>(shape, _data_type, 1, _output_quantization_info, DataLayout::NCHW, ctx); + TensorType src = create_tensor<TensorType>(shape, _data_type, 1, _input_quantization_info, DataLayout::NCHW); + TensorType dst = create_tensor<TensorType>(shape, _data_type, 1, _output_quantization_info, DataLayout::NCHW); // Create and configure function - FunctionType act_layer(ctx); + FunctionType act_layer; TensorType *dst_ptr = _in_place ? nullptr : &dst; @@ -234,7 +228,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ActivationValidationFixture : public ActivationValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, bool in_place, ActivationLayerInfo::ActivationFunction function, float alpha_beta, DataType data_type) { ActivationValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, in_place, function, alpha_beta, data_type, QuantizationInfo()); @@ -245,7 +238,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ActivationValidationQuantizedFixture : public ActivationValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, bool in_place, ActivationLayerInfo::ActivationFunction function, float alpha_beta, DataType data_type, QuantizationInfo quantization_info) { ActivationValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, in_place, function, alpha_beta, data_type, quantization_info); @@ -255,4 +247,4 @@ public: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_ACTIVATION_LAYER_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_ACTIVATIONLAYERFIXTURE_H diff --git a/tests/validation/fixtures/AddMulAddFixture.h b/tests/validation/fixtures/AddMulAddFixture.h new file mode 100644 index 0000000000..d13fef2f02 --- /dev/null +++ b/tests/validation/fixtures/AddMulAddFixture.h @@ -0,0 +1,270 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef ACL_TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE_H + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/reference/ActivationLayer.h" +#include "tests/validation/reference/ArithmeticOperations.h" +#include "tests/validation/reference/DequantizationLayer.h" +#include "tests/validation/reference/PixelWiseMultiplication.h" +#include "tests/validation/reference/QuantizationLayer.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class AddMulAddGenericFixture : public framework::Fixture +{ +public: + void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info, bool interm_out) + { + compute_target(shape, data_type, act_info, interm_out); + } + +protected: + template <typename U> + void fill(U &&tensor, int i, DataType data_type) + { + switch(data_type) + { + case DataType::F32: + library->fill_tensor_uniform(tensor, i, -10.f, 10.f); + break; + case DataType::F16: + library->fill_tensor_uniform(tensor, i, -1.f, 1.f); + break; + default: + library->fill_tensor_uniform(tensor, i); + break; + } + } + + void compute_target(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info, bool interm_out) + { + TensorShape b_shape(shape.x()); + + // Create tensors + TensorType input1 = create_tensor<TensorType>(shape, data_type, 1, _input1_qinfo); + TensorType input2 = create_tensor<TensorType>(shape, data_type, 1, _input2_qinfo); + TensorType bn_mul = create_tensor<TensorType>(b_shape, data_type, 1, _bn_mul_qinfo); + TensorType bn_add = create_tensor<TensorType>(b_shape, data_type, 1, _bn_add_qinfo); + TensorType add_output = create_tensor<TensorType>(shape, data_type, 1, _add_output_qinfo); + TensorType final_output = create_tensor<TensorType>(shape, data_type, 1, _final_output_qinfo); + + // Create and configure function + FunctionType add_mul_add; + ARM_COMPUTE_ERROR_THROW_ON(add_mul_add.validate(input1.info(), input2.info(), bn_mul.info(), + bn_add.info(), interm_out ? add_output.info() : nullptr, final_output.info(), + ConvertPolicy::SATURATE, act_info)); + + add_mul_add.configure(&input1, &input2, &bn_mul, &bn_add, interm_out ? &add_output : nullptr, + &final_output, ConvertPolicy::SATURATE, act_info); + + // Allocate tensors + input1.allocator()->allocate(); + input2.allocator()->allocate(); + bn_mul.allocator()->allocate(); + bn_add.allocator()->allocate(); + + if(interm_out) + { + add_output.allocator()->allocate(); + } + + final_output.allocator()->allocate(); + + // Fill tensors + fill(AccessorType(input1), 0, data_type); + fill(AccessorType(input2), 1, data_type); + fill(AccessorType(bn_mul), 2, data_type); + fill(AccessorType(bn_add), 3, data_type); + + // // Compute function + add_mul_add.run(); + + _target = std::move(final_output); + + if(interm_out) + { + _interm_target = std::move(add_output); + } + } + + TensorType _target{}; + TensorType _interm_target{}; + SimpleTensor<T> _reference{}; + SimpleTensor<T> _interm_reference{}; + + QuantizationInfo _input1_qinfo{}; + QuantizationInfo _input2_qinfo{}; + QuantizationInfo _bn_mul_qinfo{}; + QuantizationInfo _bn_add_qinfo{}; + QuantizationInfo _add_output_qinfo{}; + QuantizationInfo _final_output_qinfo{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool interm_out> +class AddMulAddFloatValidationFixture : public AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + using Parent = AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>; + + void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo act_info) + { + Parent::setup(shape, data_type, act_info, interm_out); + compute_reference(shape, data_type, act_info); + } + + // Compute Reference is moved outside of the generic fixture because with the quantized data types, + // it becomes a very different implementation with intermediate tensors' data types being always float. + // This way the reference calculations are more readable and the size of the classes will be smaller + // due to unrepeated fill() and target() methods. + void compute_reference(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info) + { + TensorShape b_shape(shape.x()); + + // Create reference + SimpleTensor<T> input1{ shape, data_type }; + SimpleTensor<T> input2{ shape, data_type }; + SimpleTensor<T> bn_mul{ b_shape, data_type }; + SimpleTensor<T> bn_add{ b_shape, data_type }; + SimpleTensor<T> add_output{ shape, data_type, 1 }; + + SimpleTensor<T> bn_mul_out{ shape, data_type }; + SimpleTensor<T> bn_add_out{ shape, data_type }; + + // Fill reference + Parent::fill(input1, 0, data_type); + Parent::fill(input2, 1, data_type); + Parent::fill(bn_mul, 2, data_type); + Parent::fill(bn_add, 3, data_type); + + reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, input1, input2, add_output, ConvertPolicy::SATURATE); + bn_mul_out = reference::pixel_wise_multiplication<T, T, T>(add_output, bn_mul, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_UP, data_type); + reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, bn_mul_out, bn_add, bn_add_out, ConvertPolicy::SATURATE); + + if(interm_out) + { + Parent::_interm_reference = std::move(add_output); + } + + if(act_info.enabled() && act_info.activation() != ActivationLayerInfo::ActivationFunction::IDENTITY) + { + Parent::_reference = reference::activation_layer(bn_add_out, act_info); + } + else + { + Parent::_reference = std::move(bn_add_out); + } + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool interm_out> +class AddMulAddQuantizedValidationFixture : public AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + using Parent = AddMulAddGenericFixture<TensorType, AccessorType, FunctionType, T>; + + void setup(const TensorShape &shape, DataType data_type, ActivationLayerInfo act_info, + QuantizationInfo input1_qinfo, QuantizationInfo input2_qinfo, QuantizationInfo bn_mul_qinfo, + QuantizationInfo bn_add_qinfo, QuantizationInfo add_output_qinfo, QuantizationInfo final_output_qinfo) + { + // Quantization arguments moved to class attributes to prevent long function declerations + Parent::_input1_qinfo = input1_qinfo; + Parent::_input2_qinfo = input2_qinfo; + Parent::_bn_mul_qinfo = bn_mul_qinfo; + Parent::_bn_add_qinfo = bn_add_qinfo; + Parent::_add_output_qinfo = add_output_qinfo; + Parent::_final_output_qinfo = final_output_qinfo; + + Parent::setup(shape, data_type, act_info, interm_out); + compute_reference(shape, data_type, act_info); + } + + // Compute Reference is moved outside of the generic fixture because with the quantized data types, + // it becomes a very different implementation with intermediate tensors' data types being always float. + // This way the reference calculations are more readable and the size of the classes will be smaller + // due to unrepeated fill() and target() methods. + void compute_reference(const TensorShape &shape, DataType data_type, ActivationLayerInfo &act_info) + { + TensorShape b_shape(shape.x()); + + // Create reference + SimpleTensor<T> input1{ shape, data_type, 1, Parent::_input1_qinfo }; + SimpleTensor<T> input2{ shape, data_type, 1, Parent::_input2_qinfo }; + SimpleTensor<T> bn_mul{ b_shape, data_type, 1, Parent::_bn_mul_qinfo }; + SimpleTensor<T> bn_add{ b_shape, data_type, 1, Parent::_bn_add_qinfo }; + + // Fill input tensors + Parent::fill(input1, 0, data_type); + Parent::fill(input2, 1, data_type); + Parent::fill(bn_mul, 2, data_type); + Parent::fill(bn_add, 3, data_type); + + SimpleTensor<float> input1_dequantized = reference::dequantization_layer<float>(input1); + SimpleTensor<float> input2_dequantized = reference::dequantization_layer<float>(input2); + SimpleTensor<float> bn_mul_dequantized = reference::dequantization_layer<float>(bn_mul); + SimpleTensor<float> bn_add_dequantized = reference::dequantization_layer<float>(bn_add); + + SimpleTensor<float> add_output_dequantized{ shape, DataType::F32 }; + SimpleTensor<float> bn_add_out_dequantized{ shape, DataType::F32 }; + + reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, input1_dequantized, input2_dequantized, add_output_dequantized, ConvertPolicy::SATURATE); + SimpleTensor<float> bn_mul_out_dequantized = reference::pixel_wise_multiplication<float, float, float>(add_output_dequantized, bn_mul_dequantized, 1.f, ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_UP, DataType::F32); + reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, bn_mul_out_dequantized, bn_add_dequantized, bn_add_out_dequantized, ConvertPolicy::SATURATE); + + if(interm_out) + { + Parent::_interm_reference = reference::quantization_layer<float, T>(add_output_dequantized, data_type, Parent::_add_output_qinfo); + } + + if(act_info.enabled() && act_info.activation() != ActivationLayerInfo::ActivationFunction::IDENTITY) + { + SimpleTensor<T> ref = reference::quantization_layer<float, T>(bn_add_out_dequantized, data_type, Parent::_final_output_qinfo); + Parent::_reference = reference::activation_layer(ref, act_info); + } + else + { + Parent::_reference = reference::quantization_layer<float, T>(bn_add_out_dequantized, data_type, Parent::_final_output_qinfo); + } + } +}; +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_ADDMULADDFIXTURE_H diff --git a/tests/validation/fixtures/ArgMinMaxFixture.h b/tests/validation/fixtures/ArgMinMaxFixture.h index caa6bb8d9c..7a823568a8 100644 --- a/tests/validation/fixtures/ArgMinMaxFixture.h +++ b/tests/validation/fixtures/ArgMinMaxFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,15 +42,14 @@ namespace test { namespace validation { -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2> class ArgMinMaxValidationBaseFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape shape, DataType data_type, int axis, ReductionOperation op, QuantizationInfo q_info) + void setup(TensorShape shape, DataType input_type, DataType output_type, int axis, ReductionOperation op, QuantizationInfo q_info) { - _target = compute_target(shape, data_type, axis, op, q_info); - _reference = compute_reference(shape, data_type, axis, op, q_info); + _target = compute_target(shape, input_type, output_type, axis, op, q_info); + _reference = compute_reference(shape, input_type, output_type, axis, op, q_info); } protected: @@ -80,7 +79,7 @@ protected: case DataType::QASYMM8: { std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<uint32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, 0); break; @@ -88,7 +87,7 @@ protected: case DataType::QASYMM8_SIGNED: { std::pair<int, int> bounds = get_quantized_qasymm8_signed_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<int8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<int32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, 0); break; @@ -98,11 +97,11 @@ protected: } } - TensorType compute_target(TensorShape &src_shape, DataType data_type, int axis, ReductionOperation op, QuantizationInfo q_info) + TensorType compute_target(TensorShape &src_shape, DataType input_type, DataType output_type, int axis, ReductionOperation op, QuantizationInfo q_info) { // Create tensors - TensorType src = create_tensor<TensorType>(src_shape, data_type, 1, q_info); - TensorType dst; + TensorType src = create_tensor<TensorType>(src_shape, input_type, 1, q_info); + TensorType dst = create_tensor<TensorType>(compute_output_shape(src_shape, axis), output_type, 1, q_info); // Create and configure function FunctionType arg_min_max_layer; @@ -127,41 +126,43 @@ protected: return dst; } - SimpleTensor<int32_t> compute_reference(TensorShape &src_shape, DataType data_type, int axis, ReductionOperation op, QuantizationInfo q_info) + TensorShape compute_output_shape(const TensorShape &src_shape, int axis) + { + return arm_compute::misc::shape_calculator::compute_reduced_shape(src_shape, axis, false); + } + + SimpleTensor<T2> compute_reference(TensorShape &src_shape, DataType input_type, DataType output_type, int axis, ReductionOperation op, QuantizationInfo q_info) { // Create reference - SimpleTensor<T> src{ src_shape, data_type, 1, q_info }; + SimpleTensor<T1> src{ src_shape, input_type, 1, q_info }; // Fill reference fill(src); - TensorShape output_shape = arm_compute::misc::shape_calculator::compute_reduced_shape(src_shape, axis, false); - return reference::reduction_operation<T, int32_t>(src, output_shape, axis, op); + return reference::reduction_operation<T1, T2>(src, compute_output_shape(src_shape, axis), axis, op, output_type); } - TensorType _target{}; - SimpleTensor<int32_t> _reference{}; + TensorType _target{}; + SimpleTensor<T2> _reference{}; }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class ArgMinMaxValidationQuantizedFixture : public ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2> +class ArgMinMaxValidationQuantizedFixture : public ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type, int axis, ReductionOperation op, QuantizationInfo quantization_info) + void setup(const TensorShape &shape, DataType input_type, DataType output_type, int axis, ReductionOperation op, QuantizationInfo quantization_info) { - ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, op, quantization_info); + ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, input_type, output_type, axis, op, quantization_info); } }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class ArgMinMaxValidationFixture : public ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2> +class ArgMinMaxValidationFixture : public ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type, int axis, ReductionOperation op) + void setup(const TensorShape &shape, DataType input_type, DataType output_type, int axis, ReductionOperation op) { - ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, op, QuantizationInfo()); + ArgMinMaxValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, input_type, output_type, axis, op, QuantizationInfo()); } }; } // namespace validation diff --git a/tests/validation/fixtures/ArithmeticDivisionFixture.h b/tests/validation/fixtures/ArithmeticDivisionFixture.h index 2c2c01c31c..e11a386130 100644 --- a/tests/validation/fixtures/ArithmeticDivisionFixture.h +++ b/tests/validation/fixtures/ArithmeticDivisionFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionBroadcastValidationFixture : public framework::Fixture { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type) { _target = compute_target(shape0, shape1, data_type); @@ -117,7 +116,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionValidationFixture : public ArithmeticDivisionBroadcastValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ArithmeticDivisionBroadcastValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type); diff --git a/tests/validation/fixtures/ArithmeticOperationsFixture.h b/tests/validation/fixtures/ArithmeticOperationsFixture.h index 1dfc2ce579..0785af1151 100644 --- a/tests/validation/fixtures/ArithmeticOperationsFixture.h +++ b/tests/validation/fixtures/ArithmeticOperationsFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,16 +45,14 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticOperationGenericFixture : public framework::Fixture { public: - template <typename...> - void setup(reference::ArithmeticOperation op, const TensorShape &shape0, const TensorShape &shape1, - DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, ActivationLayerInfo act_info, bool in_place) + void setup(reference::ArithmeticOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, ActivationLayerInfo act_info, bool is_inplace) { - _op = op; - _act_info = act_info; - _in_place = in_place; - _target = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out); - _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out); + _op = op; + _act_info = act_info; + _is_inplace = is_inplace; + _target = compute_target(shape0, shape1, data_type, convert_policy, qinfo0, qinfo1, qinfo_out); + _reference = compute_reference(shape0, shape1, data_type, convert_policy, qinfo0, qinfo1, qinfo_out); } protected: @@ -64,31 +62,55 @@ protected: library->fill_tensor_uniform(tensor, i); } - TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) { // Create tensors - TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, qinfo0); - TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, qinfo1); - TensorType dst = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, qinfo_out); - TensorType *dst_to_use = _in_place ? &ref_src1 : &dst; + const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1); + TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type, 1, qinfo0); + TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type, 1, qinfo1); + TensorType dst = create_tensor<TensorType>(out_shape, data_type, 1, qinfo_out); + + // Check whether do in-place computation and whether inputs are broadcast compatible + TensorType *actual_dst = &dst; + if(_is_inplace) + { + bool src1_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape0, 0) && (qinfo0 == qinfo_out); + bool src2_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape1, 0) && (qinfo1 == qinfo_out); + bool do_in_place = out_shape.total_size() != 0 && (src1_is_inplace || src2_is_inplace); + ARM_COMPUTE_ASSERT(do_in_place); + + if(src1_is_inplace) + { + actual_dst = &ref_src1; + } + else + { + actual_dst = &ref_src2; + } + } // Create and configure function FunctionType arith_op; - arith_op.configure(&ref_src1, &ref_src2, dst_to_use, convert_policy, _act_info); + arith_op.configure(&ref_src1, &ref_src2, actual_dst, convert_policy, _act_info); ARM_COMPUTE_ASSERT(ref_src1.info()->is_resizable()); ARM_COMPUTE_ASSERT(ref_src2.info()->is_resizable()); - ARM_COMPUTE_ASSERT(dst_to_use->info()->is_resizable()); // Allocate tensors ref_src1.allocator()->allocate(); ref_src2.allocator()->allocate(); - dst_to_use->allocator()->allocate(); ARM_COMPUTE_ASSERT(!ref_src1.info()->is_resizable()); ARM_COMPUTE_ASSERT(!ref_src2.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!dst_to_use->info()->is_resizable()); + + // If don't do in-place computation, still need to allocate original dst + if(!_is_inplace) + { + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + dst.allocator()->allocate(); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + } // Fill tensors fill(AccessorType(ref_src1), 0); @@ -97,50 +119,40 @@ protected: // Compute function arith_op.run(); - if(_in_place) - { - return ref_src1; - } - return dst; + return std::move(*actual_dst); } - SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, - DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, + SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) { - // current in-place implementation only supports same metadata of input and output tensors. - // By ignoring output quantization information here, we can make test cases implementation much simpler. - QuantizationInfo output_qinfo = _in_place ? qinfo0 : qinfo_out; - // Create reference - SimpleTensor<T> ref_src1{ shape0, data_type0, 1, qinfo0 }; - SimpleTensor<T> ref_src2{ shape1, data_type1, 1, qinfo1 }; - SimpleTensor<T> ref_dst{ TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, output_qinfo }; + SimpleTensor<T> ref_src1{ shape0, data_type, 1, qinfo0 }; + SimpleTensor<T> ref_src2{ shape1, data_type, 1, qinfo1 }; + SimpleTensor<T> ref_dst{ TensorShape::broadcast_shape(shape0, shape1), data_type, 1, qinfo_out }; // Fill reference fill(ref_src1, 0); fill(ref_src2, 1); auto result = reference::arithmetic_operation<T>(_op, ref_src1, ref_src2, ref_dst, convert_policy); - return _act_info.enabled() ? reference::activation_layer(result, _act_info, output_qinfo) : result; + return _act_info.enabled() ? reference::activation_layer(result, _act_info, qinfo_out) : result; } TensorType _target{}; SimpleTensor<T> _reference{}; reference::ArithmeticOperation _op{ reference::ArithmeticOperation::ADD }; ActivationLayerInfo _act_info{}; - bool _in_place{}; + bool _is_inplace{}; }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, data_type0, data_type1, - output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), false); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), is_inplace); } }; @@ -148,11 +160,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticAdditionValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy) + void setup(const TensorShape &shape, DataType data_type, ConvertPolicy convert_policy, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type0, data_type1, - output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), false); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), is_inplace); } }; @@ -160,11 +171,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticAdditionBroadcastValidationFloatFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, data_type0, data_type1, - output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, false); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -172,11 +182,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticAdditionValidationFloatFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type0, data_type1, - output_data_type, convert_policy, QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, false); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -184,13 +193,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticAdditionValidationQuantizedFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + void setup(const TensorShape &shape, DataType data_type, ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type0, data_type1, - output_data_type, convert_policy, qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), false); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape, shape, data_type, convert_policy, + qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), is_inplace); } }; @@ -198,13 +205,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticAdditionValidationQuantizedBroadcastFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, + bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, - data_type0, data_type1, output_data_type, convert_policy, - qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), false); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::ADD, shape0, shape1, data_type, convert_policy, + qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), is_inplace); } }; @@ -212,12 +217,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticSubtractionBroadcastValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, bool in_place) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1, - data_type0, data_type1, output_data_type, convert_policy, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), in_place); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), is_inplace); } }; @@ -225,13 +228,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticSubtractionBroadcastValidationFloatFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info, - bool in_place) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info, + bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1, - data_type0, data_type1, output_data_type, convert_policy, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, in_place); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -239,12 +240,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticSubtractionValidationFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, bool in_place) + void setup(const TensorShape &shape, DataType data_type, ConvertPolicy convert_policy, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape, - data_type0, data_type1, output_data_type, convert_policy, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), in_place); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), is_inplace); } }; @@ -252,12 +251,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticSubtractionValidationFloatFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info, bool in_place) + void setup(const TensorShape &shape, DataType data_type, ConvertPolicy convert_policy, ActivationLayerInfo act_info, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape, - data_type0, data_type1, output_data_type, convert_policy, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, in_place); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape, data_type, convert_policy, + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -265,14 +262,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticSubtractionValidationQuantizedFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool in_place) + void setup(const TensorShape &shape, DataType data_type, ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape, - data_type0, data_type1, output_data_type, - convert_policy, qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), in_place); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape, shape, data_type, convert_policy, + qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), is_inplace); } }; @@ -280,13 +274,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticSubtractionValidationQuantizedBroadcastFixture : public ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool in_place) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, ConvertPolicy convert_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, + bool is_inplace) { - ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1, - data_type0, data_type1, output_data_type, convert_policy, - qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), in_place); + ArithmeticOperationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(reference::ArithmeticOperation::SUB, shape0, shape1, data_type, convert_policy, + qinfo0, qinfo1, qinfo_out, ActivationLayerInfo(), is_inplace); } }; } // namespace validation diff --git a/tests/validation/fixtures/BatchNormalizationLayerFixture.h b/tests/validation/fixtures/BatchNormalizationLayerFixture.h index 38b4d4375a..54a0ed9e09 100644 --- a/tests/validation/fixtures/BatchNormalizationLayerFixture.h +++ b/tests/validation/fixtures/BatchNormalizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BatchNormalizationLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape0, TensorShape shape1, float epsilon, bool use_beta, bool use_gamma, ActivationLayerInfo act_info, DataType dt, DataLayout data_layout) { _data_type = dt; diff --git a/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h b/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h index 1ffb8f38f3..161eeb0ef4 100644 --- a/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h +++ b/tests/validation/fixtures/BatchNormalizationLayerFusionFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename ConvolutionFuncti class BatchNormalizationLayerFusionValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape src_shape, TensorShape w_shape, TensorShape b_shape, TensorShape dst_shape, PadStrideInfo info, Size2D dilation, bool use_conv_b, bool use_beta, bool use_gamma, float epsilon, DataType dt, DataLayout data_layout) { diff --git a/tests/validation/fixtures/BatchToSpaceLayerFixture.h b/tests/validation/fixtures/BatchToSpaceLayerFixture.h index 6554c09de4..56a6109dbc 100644 --- a/tests/validation/fixtures/BatchToSpaceLayerFixture.h +++ b/tests/validation/fixtures/BatchToSpaceLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #ifndef ARM_COMPUTE_TEST_BATCH_TO_SPACE_LAYER_FIXTURE #define ARM_COMPUTE_TEST_BATCH_TO_SPACE_LAYER_FIXTURE +#include "arm_compute/core/Helpers.h" #include "tests/Globals.h" #include "tests/framework/Asserts.h" #include "tests/framework/Fixture.h" @@ -39,11 +40,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BatchToSpaceLayerValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape input_shape, TensorShape block_shape_shape, TensorShape output_shape, DataType data_type, DataLayout data_layout) + void setup(const TensorShape &input_shape, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, const TensorShape &output_shape, DataType data_type, DataLayout data_layout) { - _target = compute_target(input_shape, block_shape_shape, output_shape, data_type, data_layout); - _reference = compute_reference(input_shape, block_shape_shape, output_shape, data_type); + _target = compute_target(input_shape, block_shape, crop_info, output_shape, data_type, data_layout); + _reference = compute_reference(input_shape, block_shape, crop_info, output_shape, data_type); } protected: @@ -56,9 +56,10 @@ protected: DistributionType distribution{ T(-1.0f), T(1.0f) }; library->fill(tensor, distribution, i); } - TensorType compute_target(TensorShape input_shape, TensorShape block_shape_shape, TensorShape output_shape, + TensorType compute_target(TensorShape input_shape, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, TensorShape output_shape, DataType data_type, DataLayout data_layout) { + ARM_COMPUTE_ERROR_ON(block_shape.size() != 2U); // Only support batch to 2D space (x, y) for now if(data_layout == DataLayout::NHWC) { permute(input_shape, PermutationVector(2U, 0U, 1U)); @@ -66,64 +67,49 @@ protected: } // Create tensors - TensorType input = create_tensor<TensorType>(input_shape, data_type, 1, QuantizationInfo(), data_layout); - TensorType block_shape = create_tensor<TensorType>(block_shape_shape, DataType::S32); - TensorType output = create_tensor<TensorType>(output_shape, data_type, 1, QuantizationInfo(), data_layout); + TensorType input = create_tensor<TensorType>(input_shape, data_type, 1, QuantizationInfo(), data_layout); + TensorType output = create_tensor<TensorType>(output_shape, data_type, 1, QuantizationInfo(), data_layout); // Create and configure function FunctionType batch_to_space; - batch_to_space.configure(&input, &block_shape, &output); + batch_to_space.configure(&input, block_shape.at(0), block_shape.at(1), &output, crop_info); ARM_COMPUTE_ASSERT(input.info()->is_resizable()); - ARM_COMPUTE_ASSERT(block_shape.info()->is_resizable()); ARM_COMPUTE_ASSERT(output.info()->is_resizable()); // Allocate tensors input.allocator()->allocate(); - block_shape.allocator()->allocate(); output.allocator()->allocate(); ARM_COMPUTE_ASSERT(!input.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!block_shape.info()->is_resizable()); ARM_COMPUTE_ASSERT(!output.info()->is_resizable()); // Fill tensors fill(AccessorType(input), 0); - { - auto block_shape_data = AccessorType(block_shape); - const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); - for(unsigned int i = 0; i < block_shape_shape.x(); ++i) - { - static_cast<int32_t *>(block_shape_data.data())[i] = output_shape[i + idx_width] / input_shape[i + idx_width]; - } - } // Compute function batch_to_space.run(); return output; } - SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &block_shape_shape, - const TensorShape &output_shape, DataType data_type) + SimpleTensor<T> compute_reference(const TensorShape &input_shape, const std::vector<int32_t> &block_shape, + const CropInfo &crop_info, const TensorShape &output_shape, DataType data_type) { + ARM_COMPUTE_ERROR_ON(block_shape.size() != 2U); // Only support batch to 2D space (x, y) for now // Create reference - SimpleTensor<T> input{ input_shape, data_type }; - SimpleTensor<int32_t> block_shape{ block_shape_shape, DataType::S32 }; + SimpleTensor<T> input{ input_shape, data_type }; // Fill reference fill(input, 0); - for(unsigned int i = 0; i < block_shape_shape.x(); ++i) - { - block_shape[i] = output_shape[i] / input_shape[i]; - } // Compute reference - return reference::batch_to_space(input, block_shape, output_shape); + return reference::batch_to_space(input, block_shape, crop_info, output_shape); } TensorType _target{}; SimpleTensor<T> _reference{}; }; + } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/fixtures/BitwiseAndFixture.h b/tests/validation/fixtures/BitwiseAndFixture.h index 0550532e4b..745a34058e 100644 --- a/tests/validation/fixtures/BitwiseAndFixture.h +++ b/tests/validation/fixtures/BitwiseAndFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BitwiseAndValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); diff --git a/tests/validation/fixtures/BitwiseNotFixture.h b/tests/validation/fixtures/BitwiseNotFixture.h index 9cf1938381..bdfd255156 100644 --- a/tests/validation/fixtures/BitwiseNotFixture.h +++ b/tests/validation/fixtures/BitwiseNotFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BitwiseNotValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); diff --git a/tests/validation/fixtures/BitwiseOrFixture.h b/tests/validation/fixtures/BitwiseOrFixture.h index c4fc7d7c22..03560e0171 100644 --- a/tests/validation/fixtures/BitwiseOrFixture.h +++ b/tests/validation/fixtures/BitwiseOrFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BitwiseOrValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); diff --git a/tests/validation/fixtures/BitwiseXorFixture.h b/tests/validation/fixtures/BitwiseXorFixture.h index 479694c78e..4872b231a5 100644 --- a/tests/validation/fixtures/BitwiseXorFixture.h +++ b/tests/validation/fixtures/BitwiseXorFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BitwiseXorValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); diff --git a/tests/validation/fixtures/BoundingBoxTransformFixture.h b/tests/validation/fixtures/BoundingBoxTransformFixture.h index cd6ce99796..03edaeab16 100644 --- a/tests/validation/fixtures/BoundingBoxTransformFixture.h +++ b/tests/validation/fixtures/BoundingBoxTransformFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -102,7 +102,6 @@ class BoundingBoxTransformGenericFixture : public framework::Fixture public: using TDeltas = typename std::conditional<std::is_same<typename std::decay<T>::type, uint16_t>::value, uint8_t, T>::type; - template <typename...> void setup(TensorShape deltas_shape, const BoundingBoxTransformInfo &info, DataType data_type, QuantizationInfo deltas_qinfo) { const bool is_qasymm16 = data_type == DataType::QASYMM16; @@ -215,7 +214,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BoundingBoxTransformFixture : public BoundingBoxTransformGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape deltas_shape, const BoundingBoxTransformInfo &info, DataType data_type) { BoundingBoxTransformGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(deltas_shape, info, data_type, QuantizationInfo()); @@ -228,7 +226,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class BoundingBoxTransformQuantizedFixture : public BoundingBoxTransformGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape deltas_shape, const BoundingBoxTransformInfo &info, DataType data_type, QuantizationInfo deltas_qinfo) { BoundingBoxTransformGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(deltas_shape, info, data_type, deltas_qinfo); diff --git a/tests/validation/fixtures/CastFixture.h b/tests/validation/fixtures/CastFixture.h index 056f664261..e9d624e6f3 100644 --- a/tests/validation/fixtures/CastFixture.h +++ b/tests/validation/fixtures/CastFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -36,7 +36,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class CastValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType dt_in, DataType dt_out, ConvertPolicy policy) { _target = compute_target(shape, dt_in, dt_out, policy); @@ -86,6 +85,16 @@ protected: library->fill_tensor_uniform(tensor, i, static_cast<int32_t>(signed_min), static_cast<int32_t>(signed_max)); break; } + case DataType::U64: + { + library->fill_tensor_uniform(tensor, i, static_cast<uint64_t>(unsigned_min), static_cast<uint64_t>(unsigned_max)); + break; + } + case DataType::S64: + { + library->fill_tensor_uniform(tensor, i, static_cast<int64_t>(signed_min), static_cast<int64_t>(signed_max)); + break; + } default: ARM_COMPUTE_ERROR("NOT SUPPORTED!"); } diff --git a/tests/validation/fixtures/ChannelShuffleLayerFixture.h b/tests/validation/fixtures/ChannelShuffleLayerFixture.h index 9199d9296d..530dba3893 100644 --- a/tests/validation/fixtures/ChannelShuffleLayerFixture.h +++ b/tests/validation/fixtures/ChannelShuffleLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ChannelShuffleLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, unsigned int num_groups, DataType data_type, DataLayout data_layout) { _target = compute_target(shape, data_type, num_groups, data_layout); diff --git a/tests/validation/fixtures/Col2ImFixture.h b/tests/validation/fixtures/Col2ImFixture.h index 88d420aa3d..4d56d607b7 100644 --- a/tests/validation/fixtures/Col2ImFixture.h +++ b/tests/validation/fixtures/Col2ImFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,10 +45,9 @@ namespace validation using namespace arm_compute::misc::shape_calculator; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool batch_size_on_z> -class Col2ImValidationFixture : public framework::Fixture +class Col2ImOpValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, const unsigned int convolved_width, unsigned int convolved_height, unsigned int num_groups, DataType data_type) { const Size2D convolved_dims(convolved_width, convolved_height); @@ -74,7 +73,7 @@ protected: // Create and configure function FunctionType col2im_func; - col2im_func.configure(&src, &dst, convolved_dims, num_groups); + col2im_func.configure(src.info(), dst.info(), convolved_dims, num_groups); ARM_COMPUTE_ASSERT(src.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); @@ -89,8 +88,13 @@ protected: // Fill tensors fill(AccessorType(src), 0); + arm_compute::ITensorPack pack = + { + { arm_compute::TensorType::ACL_SRC, &src }, + { arm_compute::TensorType::ACL_DST, &dst } + }; // Compute function - col2im_func.run(); + col2im_func.run(pack); return dst; } diff --git a/tests/validation/fixtures/ComparisonFixture.h b/tests/validation/fixtures/ComparisonFixture.h index 6b38a376a4..f25d5abb73 100644 --- a/tests/validation/fixtures/ComparisonFixture.h +++ b/tests/validation/fixtures/ComparisonFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComparisonValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(ComparisonOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type, QuantizationInfo qinfo0, QuantizationInfo qinfo1) { _target = compute_target(op, shape0, shape1, data_type, qinfo0, qinfo1); @@ -117,7 +116,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComparisonBroadcastValidationFixture : public ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(ComparisonOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type) { ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(op, shape0, shape1, data_type, QuantizationInfo(), QuantizationInfo()); @@ -128,7 +126,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComparisonValidationFixture : public ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(ComparisonOperation op, const TensorShape &shape, DataType data_type) { ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(op, shape, shape, data_type, QuantizationInfo(), QuantizationInfo()); @@ -139,7 +136,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComparisonValidationQuantizedFixture : public ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(ComparisonOperation op, const TensorShape &shape, DataType data_type, QuantizationInfo qinfo0, QuantizationInfo qinfo1) { @@ -151,7 +147,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComparisonQuantizedBroadcastValidationFixture : public ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(ComparisonOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type, QuantizationInfo qinfo0, QuantizationInfo qinfo1) { ComparisonValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(op, shape0, shape1, data_type, qinfo0, qinfo1); diff --git a/tests/validation/fixtures/ComputeAllAnchorsFixture.h b/tests/validation/fixtures/ComputeAllAnchorsFixture.h index e0fe35ba6f..620f1b53fa 100644 --- a/tests/validation/fixtures/ComputeAllAnchorsFixture.h +++ b/tests/validation/fixtures/ComputeAllAnchorsFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComputeAllAnchorsGenericFixture : public framework::Fixture { public: - template <typename...> void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type, QuantizationInfo qinfo) { _target = compute_target(num_anchors, data_type, info, qinfo); @@ -107,7 +106,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComputeAllAnchorsFixture : public ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type) { ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(num_anchors, info, data_type, QuantizationInfo()); @@ -118,7 +116,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ComputeAllAnchorsQuantizedFixture : public ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(size_t num_anchors, const ComputeAnchorsInfo &info, DataType data_type, QuantizationInfo qinfo) { ComputeAllAnchorsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(num_anchors, info, data_type, qinfo); diff --git a/tests/validation/fixtures/ConcatenateLayerFixture.h b/tests/validation/fixtures/ConcatenateLayerFixture.h index bafb8b2392..3a021661ac 100644 --- a/tests/validation/fixtures/ConcatenateLayerFixture.h +++ b/tests/validation/fixtures/ConcatenateLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,6 @@ private: using SrcITensorType = typename std::conditional<CI, const ITensorType, ITensorType>::type; public: - template <typename...> void setup(TensorShape shape, DataType data_type, unsigned int axis) { // Create input shapes diff --git a/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h b/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h index ae844332c3..7ad14e1b40 100644 --- a/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h +++ b/tests/validation/fixtures/ConvertFullyConnectedWeightsFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ConvertFullyConnectedWeightsValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, unsigned int weights_w, DataLayout training_data_layout, DataType data_type) { const unsigned int height = input_shape.x() * input_shape.y() * input_shape.z(); @@ -61,7 +60,7 @@ protected: { case DataType::QASYMM8: { - std::uniform_int_distribution<uint8_t> distribution(0, 10); + std::uniform_int_distribution<uint32_t> distribution(0, 10); library->fill(tensor, distribution, i); break; } diff --git a/tests/validation/fixtures/ConvolutionLayerFixture.h b/tests/validation/fixtures/ConvolutionLayerFixture.h index 6dbf3d5731..0622e5e6f0 100644 --- a/tests/validation/fixtures/ConvolutionLayerFixture.h +++ b/tests/validation/fixtures/ConvolutionLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,12 +21,19 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_CONVOLUTION_LAYER_FIXTURE -#define ARM_COMPUTE_TEST_CONVOLUTION_LAYER_FIXTURE + +#ifndef ACL_TESTS_VALIDATION_FIXTURES_CONVOLUTIONLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_CONVOLUTIONLAYERFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" +#include "arm_compute/graph/Utils.h" +#ifdef ARM_COMPUTE_OPENCL_ENABLED +#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h" +#endif // ARM_COMPUTE_OPENCL_ENABLED #include "arm_compute/runtime/NEON/NEScheduler.h" +#include "src/core/NEON/kernels/arm_gemm/utils.hpp" +#include "src/graph/mutators/MutatorUtils.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/IAccessor.h" @@ -35,10 +42,12 @@ #include "tests/validation/Helpers.h" #include "tests/validation/reference/ActivationLayer.h" #include "tests/validation/reference/ConvolutionLayer.h" +#include "tests/validation/reference/PadLayer.h" #include "tests/validation/reference/Permute.h" #include "tests/validation/reference/Utils.h" #include <random> +#include <type_traits> namespace arm_compute { @@ -49,13 +58,30 @@ namespace validation namespace detail { template <typename ConvolutionFunction, typename TensorType> -void configure_conv_function(ConvolutionFunction &func, +#ifdef ARM_COMPUTE_OPENCL_ENABLED +std::enable_if_t<!std::is_same<ConvolutionFunction, CLGEMMConvolutionLayer>::value, void> +#else // ARM_COMPUTE_OPENCL_ENABLED +void +#endif // ARM_COMPUTE_OPENCL_ENABLED +configure_conv_function(ConvolutionFunction &func, + TensorType *src, const TensorType *weights, const TensorType *bias, TensorType *dst, + const PadStrideInfo &info, const WeightsInfo &weights_info, + const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups) +{ + func.configure(src, weights, bias, dst, info, weights_info, dilation, act_info, false /* enable_fast_math */, num_groups); +} + +#ifdef ARM_COMPUTE_OPENCL_ENABLED +template <typename ConvolutionFunction, typename TensorType> +std::enable_if_t<std::is_same<ConvolutionFunction, CLGEMMConvolutionLayer>::value, void> +configure_conv_function(ConvolutionFunction &func, TensorType *src, const TensorType *weights, const TensorType *bias, TensorType *dst, const PadStrideInfo &info, const WeightsInfo &weights_info, const Size2D &dilation, const ActivationLayerInfo &act_info, unsigned int num_groups) { func.configure(src, weights, bias, dst, info, weights_info, dilation, act_info, num_groups); } +#endif // ARM_COMPUTE_OPENCL_ENABLED } // namespace detail template <typename TensorType, typename AccessorType, typename FunctionType, typename T, typename TW> @@ -66,25 +92,67 @@ public: || std::is_same<typename std::decay<T>::type, int8_t>::value, int32_t, T >::type; + void setup_quantization(TensorShape input_shape, TensorShape weights_shape, QuantizationInfo &input_q_info, + QuantizationInfo &weights_q_info, DataType data_type) + { + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + _hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + _quantization_info = QuantizationInfo(scale_lhs, offset_lhs); + _weight_quantization_info = QuantizationInfo(scale_rhs, offset_rhs); + + QuantizationHint q_hint = suggest_conv_dst_q_info_and_bias(input_q_info, weights_q_info, + weights_shape.y() /* heights */, weights_shape.x() /* width */, input_shape.z() /* channels */, + data_type, 0.5f /* bias_fraction */); + + _dst_q_info = q_hint.q_info; + _min_bias = q_hint.bias_min; + _max_bias = q_hint.bias_max; + } + public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, DataType weights_data_type, DataLayout data_layout, QuantizationInfo quantization_info, QuantizationInfo weight_quantization_info, ActivationLayerInfo act_info, - bool mixed_layout = false) + bool mixed_layout = false, PaddingList pre_pad_layer = PaddingList({}), bool padded_weights = false) { + // This hash is used by random generators. There may be hash collisions but + // this is intentional as it's a very easy way to make the the current + // random generation process almost different for many test configurations, + // which were using the same set of values before. + _hash = input_shape[0] + input_shape[1] + input_shape[2] + input_shape[3] + + + weights_shape[0] + weights_shape[1] + weights_shape[2] + weights_shape[3] + + mixed_layout + (data_type == DataType::QASYMM8_SIGNED) + (data_layout == DataLayout::NHWC); + _mixed_layout = mixed_layout; _data_type = data_type; _weights_data_type = weights_data_type; - _is_quantized = is_data_type_quantized_asymmetric(data_type); + const bool is_quantized = is_data_type_quantized(weights_data_type); _is_bfloat16 = data_type == DataType::BFLOAT16; - _bias_data_type = _is_quantized ? DataType::S32 : (_is_bfloat16 ? DataType::F32 : data_type); + _bias_data_type = is_quantized ? DataType::S32 : (_is_bfloat16 ? DataType::F32 : data_type); _output_data_type = _is_bfloat16 ? DataType::F32 : data_type; _quantization_info = quantization_info; _weight_quantization_info = weight_quantization_info; _data_layout = data_layout; + _dst_q_info = quantization_info; + + if(is_quantized && !is_data_type_quantized_symmetric(weights_data_type) && (!act_info.enabled() || act_info.activation() == ActivationFunction::IDENTITY)) + { + setup_quantization(input_shape, weights_shape, _quantization_info, _weight_quantization_info, data_type); + _use_dynamic_output_quant = true; + } - _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, reshape_weights, dilation, act_info); - _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info); + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, reshape_weights, dilation, act_info, pre_pad_layer, padded_weights); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, dilation, act_info, pre_pad_layer); } protected: @@ -118,16 +186,34 @@ protected: { case DataType::QASYMM8: { - std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second); - library->fill(tensor, distribution, i); + if(_use_dynamic_output_quant) + { + std::uniform_int_distribution<int32_t> distribution(0, 255); + library->fill(tensor, distribution, i); + } + else + { + // Legacy initialization in case the output quantization info can't be reliably estimated + std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); + std::uniform_int_distribution<uint32_t> distribution(bounds.first, bounds.second); + library->fill(tensor, distribution, i); + } break; } case DataType::QASYMM8_SIGNED: { - std::pair<int, int> bounds = get_quantized_qasymm8_signed_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<int8_t> distribution(bounds.first, bounds.second); - library->fill(tensor, distribution, i); + if(_use_dynamic_output_quant) + { + std::uniform_int_distribution<int32_t> distribution(-128, 127); + library->fill(tensor, distribution, i); + } + else + { + // Legacy initialization in case the output quantization info can't be reliably estimated + std::pair<int, int> bounds = get_quantized_qasymm8_signed_bounds(tensor.quantization_info(), -1.0f, 1.0f); + std::uniform_int_distribution<int32_t> distribution(bounds.first, bounds.second); + library->fill(tensor, distribution, i); + } break; } case DataType::QSYMM8_PER_CHANNEL: @@ -146,13 +232,13 @@ protected: max_bound = bounds.second; } } - std::uniform_int_distribution<int8_t> distribution(min_bound, max_bound); + std::uniform_int_distribution<int32_t> distribution(min_bound, max_bound); library->fill(tensor, distribution, i); break; } case DataType::S32: { - std::uniform_int_distribution<int32_t> distribution(-100, 100); + std::uniform_int_distribution<int32_t> distribution(_min_bias, _max_bias); library->fill(tensor, distribution, i); break; } @@ -179,8 +265,9 @@ protected: } } + // given input is IN nchw format TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const PadStrideInfo &info, - bool reshape_weights, const Size2D &dilation, const ActivationLayerInfo act_info) + bool reshape_weights, const Size2D &dilation, const ActivationLayerInfo act_info, PaddingList pre_pad_layer = PaddingList({}), bool padded_weights = false) { ARM_COMPUTE_ERROR_ON((input_shape[2] % weights_shape[2]) != 0); @@ -191,6 +278,18 @@ protected: permute(input_shape, PermutationVector(2U, 0U, 1U)); permute(weights_shape, PermutationVector(2U, 0U, 1U)); permute(output_shape, PermutationVector(2U, 0U, 1U)); + + if(pre_pad_layer.size() > 0) + { + // make sure paddings exist for each c,h,w dimensions + for(unsigned int i = 0; i < 3 - pre_pad_layer.size(); ++i) + { + pre_pad_layer.push_back({ 0, 0 }); + } + + // rotate padding info from nchw to nhwc + std::rotate(pre_pad_layer.begin(), pre_pad_layer.begin() + 2, pre_pad_layer.begin() + 3); + } } const int idx_width = get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH); @@ -202,19 +301,47 @@ protected: // Create tensors TensorType src = create_tensor<TensorType>(input_shape, _data_type, 1, _quantization_info, _data_layout); TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, _weights_data_type, 1, _weight_quantization_info, _data_layout); - TensorType bias = create_tensor<TensorType>(bias_shape, _bias_data_type, 1, _quantization_info, _data_layout); - TensorType dst = create_tensor<TensorType>(output_shape, _output_data_type, 1, _quantization_info, _data_layout); + TensorType bias = create_tensor<TensorType>(bias_shape, _bias_data_type, 1, QuantizationInfo() /*bias is not a quantized type*/, _data_layout); + TensorType dst = create_tensor<TensorType>(output_shape, _output_data_type, 1, _dst_q_info, _data_layout); // Create and configure function FunctionType conv; - detail::configure_conv_function(conv, &src, &weights, &bias, &dst, info, weights_info, dilation, act_info, num_groups); + + const unsigned int height_index = arm_compute::graph::get_dimension_idx(_data_layout, DataLayoutDimension::HEIGHT); + const unsigned int width_index = arm_compute::graph::get_dimension_idx(_data_layout, DataLayoutDimension::WIDTH); + + const PaddingInfo pad_w = width_index < pre_pad_layer.size() ? pre_pad_layer[width_index] : PaddingInfo(0, 0); + const PaddingInfo pad_h = height_index < pre_pad_layer.size() ? pre_pad_layer[height_index] : PaddingInfo(0, 0); + + if(pre_pad_layer.size() > 0 && arm_compute::graph::is_padding_in_height_or_width(_data_layout, pre_pad_layer)) + { + // this is the logic implemented in NodeFusionMutator -> fuse_pad_with_convolution + const PadStrideInfo new_conv_info( + info.stride().first, + info.stride().second, + info.pad_left() + pad_w.first, + info.pad_right() + pad_w.second, + info.pad_top() + pad_h.first, + info.pad_bottom() + pad_h.second, + info.round()); + detail::configure_conv_function(conv, &src, &weights, &bias, &dst, new_conv_info, weights_info, dilation, act_info, num_groups); + } + else + { + detail::configure_conv_function(conv, &src, &weights, &bias, &dst, info, weights_info, dilation, act_info, num_groups); + } ARM_COMPUTE_ASSERT(src.info()->is_resizable()); ARM_COMPUTE_ASSERT(weights.info()->is_resizable()); ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); - - add_padding_x({ &src, &weights, &bias, &dst }, _data_layout); + // Test "add padding after configure" behavior. This behavior should not affect the correctness + add_padding_x({ &src, &bias, &dst }, _data_layout); + // Padding weights may affect code path in some backends + if (padded_weights) + { + add_padding_x({ &weights }, _data_layout); + } // Allocate tensors src.allocator()->allocate(); @@ -228,9 +355,9 @@ protected: ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); // Fill tensors - fill(AccessorType(src), 0); - fill(AccessorType(weights), 1); - fill(AccessorType(bias), 2); + fill(AccessorType(src), 0 + _hash); + fill(AccessorType(weights), 1 + _hash); + fill(AccessorType(bias), 2 + _hash); if(_mixed_layout) { @@ -246,7 +373,7 @@ protected: } SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, - const Size2D &dilation, const ActivationLayerInfo act_info) + const Size2D &dilation, const ActivationLayerInfo act_info, PaddingList pre_pad_layer = PaddingList({})) { ARM_COMPUTE_ERROR_ON((input_shape[2] % weights_shape[2]) != 0); @@ -262,9 +389,9 @@ protected: SimpleTensor<TW> weights{ weights_shape, weights_dt, 1, _weight_quantization_info }; SimpleTensor<TBias> bias{ bias_shape, bias_dt, 1, _quantization_info }; - fill(src, 0); - fill(weights, 1); - fill(bias, 2); + fill(src, 0 + _hash); + fill(weights, 1 + _hash); + fill(bias, 2 + _hash); // Fill with bfloat16 to perform the conversion and reduce the mismatches in the output if(_is_bfloat16) @@ -273,9 +400,14 @@ protected: regularize_values(static_cast<void *>(weights.data()), weights.num_elements()); } - return (act_info.enabled()) ? reference::activation_layer<T>(reference::convolution_layer<T>(src, weights, bias, output_shape, info, dilation, num_groups), + if(pre_pad_layer.size() > 0) + { + src = reference::pad_layer<T>(src, pre_pad_layer, PixelValue(0), PaddingMode::CONSTANT); + } + + return (act_info.enabled()) ? reference::activation_layer<T>(reference::convolution_layer<T>(src, weights, bias, output_shape, info, dilation, num_groups, _dst_q_info), act_info) : - reference::convolution_layer<T>(src, weights, bias, output_shape, info, dilation, num_groups); + reference::convolution_layer<T>(src, weights, bias, output_shape, info, dilation, num_groups, _dst_q_info); } TensorType _target{}; @@ -287,16 +419,19 @@ protected: DataLayout _data_layout{}; QuantizationInfo _quantization_info{}; QuantizationInfo _weight_quantization_info{}; - bool _is_quantized = false; + QuantizationInfo _dst_q_info{}; bool _is_bfloat16 = false; bool _mixed_layout = false; + bool _use_dynamic_output_quant{false}; + int32_t _hash{0}; + int32_t _min_bias{-100}; + int32_t _max_bias{100}; }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> class ConvolutionValidationFixture : public ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info) { @@ -307,10 +442,35 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> +class ConvolutionValidationPaddedWeightsFixture : public ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> +{ +public: + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, + DataLayout data_layout) + { + ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, + data_type, data_type, data_layout, + QuantizationInfo(), QuantizationInfo(), ActivationLayerInfo(), mixed_layout, PaddingList({}), true); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> +class ConvolutionValidationWithPaddingFixture : public ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> +{ +public: + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, + DataLayout data_layout, ActivationLayerInfo act_info, PaddingList pre_pad_layer = PaddingList({})) + { + ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, dilation, reshape_weights, + data_type, data_type, data_layout, + QuantizationInfo(), QuantizationInfo(), act_info, mixed_layout, pre_pad_layer); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> class ConvolutionValidationQuantizedFixture : public ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info, ActivationLayerInfo act_info) { @@ -323,7 +483,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ConvolutionValidationQuantizedPerChannelFixture : public ConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T, TW> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, bool reshape_weights, DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info, ActivationLayerInfo act_info, DataType weights_data_type) { @@ -339,7 +498,311 @@ public: quantization_info, QuantizationInfo(weights_scales), act_info); } }; + + +#ifdef ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS +inline TensorInfo prepare_weights(const TensorInfo tensor_info, const arm_compute::WeightFormat weight_format) +{ + const DataLayout data_layout = tensor_info.data_layout(); + ARM_COMPUTE_EXPECT(data_layout == DataLayout::NHWC, framework::LogLevel::ERRORS); + const DataType data_type = tensor_info.data_type(); + const TensorShape tensor_shape = tensor_info.tensor_shape(); + const int N = tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES)]; // N=O + const int H = tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)]; + const int W = tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)]; + const int C = tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)]; // C=I + + const int interleave_by = arm_compute::interleave_by(weight_format); + const int block_by = arm_compute::block_by(weight_format); + const int Ip = arm_gemm::roundup<unsigned int>(C, block_by); // C'=I' + const int Op = arm_gemm::roundup<unsigned int>(N, interleave_by); // O'=N' + + arm_compute::Strides strides_in_bytes = tensor_info.strides_in_bytes(); + strides_in_bytes.set(1, Ip * interleave_by * H * W * tensor_info.element_size()); + strides_in_bytes.set(2, Ip * Op * tensor_info.element_size()); + + const size_t offset_first_element_in_bytes = tensor_info.offset_first_element_in_bytes(); + + // Total size needs to include padded dimensions + const size_t total_size_in_bytes = Op * H * W * Ip * tensor_info.element_size(); + + const TensorShape TS(Ip, W, H, Op); + + TensorInfo new_tensor_info = tensor_info; + new_tensor_info.init(TS, 1 /*num_channels, deprecated*/, data_type, strides_in_bytes, + offset_first_element_in_bytes, total_size_in_bytes); + return new_tensor_info; +} + +template <typename ScalarType, typename AccessorType> +inline void rearrange_data(const AccessorType src, AccessorType dst, const arm_compute::WeightFormat weight_format) +{ + ARM_COMPUTE_EXPECT(arm_compute::is_fixed_format(weight_format), framework::LogLevel::ERRORS); + // Data Layout: OHWIo<interleave_by>i<block_by> + const int interleave_by = arm_compute::interleave_by(weight_format); + const int block_by = arm_compute::block_by(weight_format); + const TensorShape src_tensor_shape = src.shape(); + const DataLayout data_layout = src.data_layout(); + ARM_COMPUTE_EXPECT(data_layout == DataLayout::NHWC, framework::LogLevel::ERRORS); + const unsigned int O = src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES)]; // N=O + const unsigned int H = src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)]; + const unsigned int W = src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)]; + const unsigned int I = src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)]; // C=I + const unsigned int Ip = arm_gemm::roundup<unsigned int>(I, block_by); // C'=I' + const unsigned int Op = arm_gemm::roundup<unsigned int>(O, interleave_by); // N'=O' + + ARM_COMPUTE_EXPECT_EQUAL(Op * H * W * Ip, (unsigned)dst.num_elements(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(src.num_elements() <= dst.num_elements(), framework::LogLevel::ERRORS); + + const ScalarType *src_ptr = reinterpret_cast<const ScalarType *>(src.data()); + ScalarType *dst_ptr = reinterpret_cast<ScalarType *>(dst.data()); + for(unsigned i = 0; i < I; ++i) + for(unsigned w = 0; w < W; ++w) + for(unsigned h = 0; h < H; ++h) + for(unsigned o = 0; o < O; ++o) + { + ScalarType src_element; + switch(data_layout) + { + case DataLayout::NHWC: + { + src_element = src_ptr[o * H * W * I + h * W * I + w * I + i]; + } + break; + default: + { + ARM_COMPUTE_ERROR("Unsupported memory layout."); + } + } + const int x5 = std::floor(((float)o) / interleave_by); + const int x4 = h; + const int x3 = w; + const int x2 = std::floor((float)i / block_by); + const int x1 = o % interleave_by; + const int x0 = i % block_by; + unsigned dst_idx = x5 * H * W * Ip * interleave_by + + x4 * W * Ip * interleave_by + + x3 * Ip * interleave_by + + x2 * interleave_by * block_by + + x1 * block_by + + x0; + dst_ptr[dst_idx] = src_element; + } +} + +template <typename ConvolutionFunction, typename TensorClass, typename AccessorType, typename ScalarType, bool enable_fast_math> +class VariableWeightsFixtureBaseClass : public framework::Fixture +{ +public: + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataLayout data_layout, + const DataType data_type) + { + conv = std::make_unique<ConvolutionFunction>(); + // prepare data + _data_layout = data_layout; + // Fixed format kernels for variable weights can work only with NHWC format. + ARM_COMPUTE_EXPECT_EQUAL(_data_layout, DataLayout::NHWC, framework::LogLevel::ERRORS); + _data_type = data_type; + // run the code + compute_target(input_shape, weights_shape, bias_shape, output_shape, info, dilation); + compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, dilation); + } + void teardown() + { + _target.allocator()->free(); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + +private: + virtual void configure_and_execute_kernel(TensorInfo src_tensor_info, TensorInfo weight_tensor_info, TensorInfo bias_tensor_info, TensorInfo dst_tensor_info, const WeightsInfo weights_info, + const PadStrideInfo &conv_info, + const Size2D &dilation) = 0; + + void compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const PadStrideInfo &conv_info, + const Size2D &dilation) + { + // The dataset is always in NCHW format - we need to make C the + // innermost dimension because the fixed-format kernel work only + // with NHWC layout. + permute(input_shape, PermutationVector(2U, 0U, 1U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + permute(output_shape, PermutationVector(2U, 0U, 1U)); + const auto src_tensor_info = TensorInfo(input_shape, 1, _data_type, _data_layout); + const auto weight_tensor_info = TensorInfo(weights_shape, 1, _data_type, _data_layout); + const auto bias_tensor_info = TensorInfo(bias_shape, 1, _data_type, _data_layout); + auto dst_tensor_info = TensorInfo(output_shape, 1, _data_type, _data_layout); + + const int kernel_height = weights_shape[get_data_layout_dimension_index(_data_layout, DataLayoutDimension::HEIGHT)]; + const int kernel_width = weights_shape[get_data_layout_dimension_index(_data_layout, DataLayoutDimension::WIDTH)]; + const int num_kernels = weights_shape[get_data_layout_dimension_index(_data_layout, DataLayoutDimension::BATCHES)]; + + const WeightsInfo query_weights_info(/*reshape_weights*/ false, kernel_width, kernel_height, num_kernels, false, arm_compute::WeightFormat::ANY); + const bool kernel_found = bool(ConvolutionFunction::has_opt_impl(_computed_weight_format, &src_tensor_info, &weight_tensor_info, + &bias_tensor_info, &dst_tensor_info, conv_info, query_weights_info)); + // Make surethat the setup founds a fixed-format kernel as requested by the test case. + ARM_COMPUTE_EXPECT(kernel_found, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(arm_compute::is_fixed_format(_computed_weight_format), framework::LogLevel::ERRORS); + + const WeightsInfo weights_info(/*reshape_weights*/ false, kernel_width, kernel_height, num_kernels, false, _computed_weight_format); + configure_and_execute_kernel(src_tensor_info, weight_tensor_info, bias_tensor_info, dst_tensor_info, weights_info, conv_info, + dilation); + } + void compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, + const Size2D &dilation) + { + ARM_COMPUTE_UNUSED(input_shape, weights_shape, bias_shape, output_shape, info, + dilation); + + // Create reference + SimpleTensor<ScalarType> src{ input_shape, _data_type }; + SimpleTensor<ScalarType> weights{ weights_shape, _data_type }; + SimpleTensor<ScalarType> bias{ bias_shape, _data_type }; + fill(src, 0); + fill(bias, 1); + fill(weights, 3); + _reference = reference::convolution_layer<ScalarType>(src, weights, bias, output_shape, info, dilation, 1 /*num_groups*/); + } + DataLayout _data_layout{}; + DataType _data_type{}; + +protected: + std::unique_ptr<ConvolutionFunction> conv{}; + arm_compute::WeightFormat _computed_weight_format{ arm_compute::WeightFormat::UNSPECIFIED }; + TensorClass _target{}; + SimpleTensor<ScalarType> _reference{}; +}; + +template <typename ConvolutionFunction, typename TensorClass, typename AccessorType, typename ScalarType, bool enable_fast_math> +class VariableWeightsFixture : public VariableWeightsFixtureBaseClass<ConvolutionFunction, TensorClass, AccessorType, ScalarType, enable_fast_math> +{ + void configure_and_execute_kernel(TensorInfo src_tensor_info, TensorInfo weight_tensor_info, TensorInfo bias_tensor_info, TensorInfo dst_tensor_info, const WeightsInfo weights_info, + const PadStrideInfo &conv_info, + const Size2D &dilation) + { + this->conv->configure(&src_tensor_info, &weight_tensor_info, &bias_tensor_info, &dst_tensor_info, conv_info, weights_info, dilation, ActivationLayerInfo(), enable_fast_math); + + // Allocate input tensors + auto src = create_tensor<TensorClass>(src_tensor_info); + auto weights_original = create_tensor<TensorClass>(weight_tensor_info); + const TensorInfo new_tensor_info = prepare_weights(weight_tensor_info, this->_computed_weight_format); + auto weights_transformed = create_tensor<TensorClass>(new_tensor_info); + auto bias = create_tensor<TensorClass>(bias_tensor_info); + src.allocator()->allocate(); + weights_original.allocator()->allocate(); + weights_transformed.allocator()->allocate(); + bias.allocator()->allocate(); + // Allocate destination tensor + this->_target = create_tensor<TensorClass>(dst_tensor_info); + this->_target.allocator()->allocate(); + + // Prepare source and biases that are left unchanged. + this->fill(AccessorType(src), 0); + this->fill(AccessorType(bias), 1); + + // First run + this->fill(AccessorType(weights_original), 2); + rearrange_data<ScalarType, AccessorType>(AccessorType(weights_original), AccessorType(weights_transformed), this->_computed_weight_format); + ITensorPack run_pack{ { TensorType::ACL_SRC_0, &src }, { TensorType::ACL_SRC_1, &weights_transformed }, { TensorType::ACL_SRC_2, &bias }, { TensorType::ACL_DST, &(this->_target) } }; + this->conv->run(run_pack); + // Second run, with new weights + this->fill(AccessorType(weights_original), 3); + rearrange_data<ScalarType, AccessorType>(AccessorType(weights_original), AccessorType(weights_transformed), this->_computed_weight_format); + this->conv->run(run_pack); + src.allocator()->free(); + weights_original.allocator()->free(); + weights_transformed.allocator()->free(); + bias.allocator()->free(); + } +}; + +template <typename ConvolutionFunction, typename TensorClass, typename AccessorType, typename ScalarType, bool enable_fast_math> +class VariableWeightsFixtureNEInterface : public VariableWeightsFixtureBaseClass<ConvolutionFunction, TensorClass, AccessorType, ScalarType, enable_fast_math> +{ + void configure_and_execute_kernel(TensorInfo src_tensor_info, TensorInfo weight_tensor_info, TensorInfo bias_tensor_info, TensorInfo dst_tensor_info, const WeightsInfo weights_info, + const PadStrideInfo &conv_info, + const Size2D &dilation) + { + // Allocate input tensors + auto src = create_tensor<TensorClass>(src_tensor_info); + auto weights_original = create_tensor<TensorClass>(weight_tensor_info); + const TensorInfo new_tensor_info = prepare_weights(weight_tensor_info, this->_computed_weight_format); + auto weights_transformed = create_tensor<TensorClass>(new_tensor_info); + auto bias = create_tensor<TensorClass>(bias_tensor_info); + src.allocator()->allocate(); + weights_original.allocator()->allocate(); + weights_transformed.allocator()->allocate(); + bias.allocator()->allocate(); + // Allocate destination tensor + this->_target = create_tensor<TensorClass>(dst_tensor_info); + this->_target.allocator()->allocate(); + this->conv->configure(&src, &weights_transformed, &bias, &(this->_target), conv_info, weights_info, dilation, ActivationLayerInfo(), enable_fast_math); + // Prepare source and biases that are left unchanged. + this->fill(AccessorType(src), 0); + this->fill(AccessorType(bias), 1); + + // First run + this->fill(AccessorType(weights_original), 2); + rearrange_data<ScalarType, AccessorType>(AccessorType(weights_original), AccessorType(weights_transformed), this->_computed_weight_format); + this->conv->run(); + // Second run, with new weights + this->fill(AccessorType(weights_original), 3); + rearrange_data<ScalarType, AccessorType>(AccessorType(weights_original), AccessorType(weights_transformed), this->_computed_weight_format); + this->conv->run(); + src.allocator()->free(); + weights_original.allocator()->free(); + weights_transformed.allocator()->free(); + bias.allocator()->free(); + } +}; + +template <typename ConvolutionClass, bool enable_fast_math> +class HasOptImplFixture : public framework::Fixture +{ +public: + void setup(DataType data_type, arm_compute::WeightFormat query_weight_format) + { + auto conv = std::make_unique<ConvolutionClass>(); + const auto src_info = TensorInfo(TensorShape(56U, 56U, 64U), 1, data_type, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(64, 3U, 3U, 64U), 1, enable_fast_math ? DataType::BFLOAT16 : data_type, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(64U), 1, data_type, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(56U, 56U, 64U), 1, data_type, DataLayout::NHWC); + const auto conv_info = PadStrideInfo(1, 1, 1, 1, 1, 1, DimensionRoundingType::FLOOR); + const WeightsInfo weights_info(false, 3U, 3U, 64U, false, query_weight_format); + _kernel_found = bool(ConvolutionClass::has_opt_impl(_computed_weight_format, &src_info, &weight_info, + &bias_info, &dst_info, conv_info, weights_info, + Size2D(1U, 1U) /*dilation*/, ActivationLayerInfo() /*act_info*/, enable_fast_math)); + } + +protected: + bool _kernel_found{ false }; + arm_compute::WeightFormat _computed_weight_format{ arm_compute::WeightFormat::UNSPECIFIED }; +}; +#endif // ARM_COMPUTE_ENABLE_FIXED_FORMAT_KERNELS + } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_CONVOLUTION_LAYER_FIXTURE */ + +#endif // ACL_TESTS_VALIDATION_FIXTURES_CONVOLUTIONLAYERFIXTURE_H diff --git a/tests/validation/fixtures/CopyFixture.h b/tests/validation/fixtures/CopyFixture.h index eca75e7b69..f5e711a500 100644 --- a/tests/validation/fixtures/CopyFixture.h +++ b/tests/validation/fixtures/CopyFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class CopyFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape output_shape, DataType data_type) { _target = compute_target(input_shape, output_shape, data_type); diff --git a/tests/validation/fixtures/CropResizeFixture.h b/tests/validation/fixtures/CropResizeFixture.h index 5aa63c4ef8..30a3fd8569 100644 --- a/tests/validation/fixtures/CropResizeFixture.h +++ b/tests/validation/fixtures/CropResizeFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class CropResizeFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape src_shape, TensorShape boxes_shape, Coordinates2D crop_size, InterpolationPolicy method, float extrapolation_value, bool is_outside_bounds, DataType data_type) { diff --git a/tests/validation/fixtures/DeconvolutionLayerFixture.h b/tests/validation/fixtures/DeconvolutionLayerFixture.h index 14f071eed0..83170c413c 100644 --- a/tests/validation/fixtures/DeconvolutionLayerFixture.h +++ b/tests/validation/fixtures/DeconvolutionLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -49,7 +49,6 @@ public: using TBias = typename std::conditional < std::is_same<typename std::decay<T>::type, uint8_t>::value || std::is_same<typename std::decay<T>::type, int8_t>::value, int32_t, T >::type; public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, DataType data_type, DataType weights_data_type, DataLayout data_layout, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, QuantizationInfo weights_quantization_info, bool add_bias) @@ -75,14 +74,14 @@ protected: case DataType::QASYMM8: { std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<uint32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, i); break; } case DataType::QASYMM8_SIGNED: { std::pair<int, int> bounds = get_quantized_qasymm8_signed_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<int8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<int32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, i); break; } @@ -102,7 +101,7 @@ protected: max_bound = bounds.second; } } - std::uniform_int_distribution<int8_t> distribution(min_bound, max_bound); + std::uniform_int_distribution<int32_t> distribution(min_bound, max_bound); library->fill(tensor, distribution, i); break; } @@ -246,11 +245,9 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DeconvolutionValidationFixture : public DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int padx, unsigned int pady, unsigned int num_kernels, DataType data_type, DataLayout data_layout, bool add_bias) { - ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported"); const TensorShape weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels); const TensorShape bias_shape(num_kernels); const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL); @@ -267,11 +264,9 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DeconvolutionValidationAsymmFixture : public DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int pad_left, unsigned int pad_right, unsigned int pad_top, unsigned int pad_bottom, unsigned int num_kernels, DataType data_type, DataLayout data_layout, bool add_bias) { - ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported"); const TensorShape weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels); const TensorShape bias_shape(num_kernels); const PadStrideInfo info(sx, sy, pad_left, pad_right, pad_top, pad_bottom, DimensionRoundingType::CEIL); @@ -288,11 +283,9 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DeconvolutionValidationQuantizedFixture : public DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int padx, unsigned int pady, unsigned int num_kernels, DataType data_type, DataLayout data_layout, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, bool add_bias) { - ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported"); const TensorShape weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels); const TensorShape bias_shape(num_kernels); const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL); @@ -310,12 +303,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DeconvolutionValidationQuantizedPerChannelFixture : public DeconvolutionLayerFixtureBase<TensorType, AccessorType, FunctionType, T, TW> { public: - template <typename...> void setup(TensorShape input_shape, unsigned int sx, unsigned int sy, unsigned int padx, unsigned int pady, unsigned int num_kernels, DataType data_type, DataLayout data_layout, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, bool add_bias, DataType weights_data_type) { - ARM_COMPUTE_ERROR_ON_MSG(kernel_size_x != kernel_size_y, "Only square kernels supported"); const TensorShape weights_shape(kernel_size_x, kernel_size_y, input_shape.z(), num_kernels); const TensorShape bias_shape(num_kernels); const PadStrideInfo info(sx, sy, padx, pady, DimensionRoundingType::CEIL); diff --git a/tests/validation/fixtures/DepthConvertLayerFixture.h b/tests/validation/fixtures/DepthConvertLayerFixture.h index 130b583dc1..f55d20bf3e 100644 --- a/tests/validation/fixtures/DepthConvertLayerFixture.h +++ b/tests/validation/fixtures/DepthConvertLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DepthConvertLayerValidationBaseFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType dt_in, DataType dt_out, ConvertPolicy policy, uint32_t shift, QuantizationInfo quantization_info) { _shift = shift; @@ -61,7 +60,7 @@ protected: if(is_data_type_quantized(tensor.data_type())) { std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<uint32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, i); } @@ -130,7 +129,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DepthConvertLayerValidationFixture : public DepthConvertLayerValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> void setup(TensorShape shape, DataType dt_in, DataType dt_out, ConvertPolicy policy, uint32_t shift) { DepthConvertLayerValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, dt_in, dt_out, policy, @@ -142,7 +140,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DepthConvertLayerValidationQuantizedFixture : public DepthConvertLayerValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> void setup(TensorShape shape, DataType dt_in, DataType dt_out, ConvertPolicy policy, uint32_t shift, QuantizationInfo quantization_info) { DepthConvertLayerValidationBaseFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, dt_in, dt_out, policy, diff --git a/tests/validation/fixtures/DepthToSpaceLayerFixture.h b/tests/validation/fixtures/DepthToSpaceLayerFixture.h index b7f5552474..abe3d8b22f 100644 --- a/tests/validation/fixtures/DepthToSpaceLayerFixture.h +++ b/tests/validation/fixtures/DepthToSpaceLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -39,7 +39,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DepthToSpaceLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, int32_t block_shape, TensorShape output_shape, DataType data_type, DataLayout data_layout) { _target = compute_target(input_shape, block_shape, output_shape, data_type, data_layout); diff --git a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h index 2c943735ca..6e2e3a3846 100644 --- a/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h +++ b/tests/validation/fixtures/DepthwiseConvolutionLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_DEPTHWISE_CONVOLUTION_FIXTURE -#define ARM_COMPUTE_TEST_DEPTHWISE_CONVOLUTION_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DEPTHWISECONVOLUTIONLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DEPTHWISECONVOLUTIONLAYERFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -38,6 +38,7 @@ #include "utils/Utils.h" +#include <cstdint> #include <random> namespace arm_compute @@ -54,25 +55,61 @@ class DepthwiseConvolutionLayerValidationGenericFixture : public framework::Fixt public: using TBias = typename std::conditional < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T >::type; + void setup_quantization(TensorShape input_shape, TensorShape weights_shape, QuantizationInfo &input_q_info, + QuantizationInfo &weights_q_info, DataType data_type) + { + ARM_COMPUTE_UNUSED(input_shape); + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + _hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + _input_quantization_info = QuantizationInfo(scale_lhs, offset_lhs); + _weights_quantization_info = QuantizationInfo(scale_rhs, offset_rhs); + + QuantizationHint q_hint = suggest_conv_dst_q_info_and_bias(input_q_info, weights_q_info, + weights_shape.y() /* heights */, weights_shape.x() /* width */, 1 /* channels */, + data_type, 0.5f /* bias_fraction */); + + _output_quantization_info = q_hint.q_info; + _min_bias = q_hint.bias_min; + _max_bias = q_hint.bias_max; + } + public: - template <typename...> void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType input_data_type, DataType weights_data_type, QuantizationInfo input_quantization_info, QuantizationInfo weights_quantization_info, QuantizationInfo output_quantization_info, - DataLayout data_layout, ActivationLayerInfo act_info, bool mixed_layout = false) + DataLayout data_layout, ActivationLayerInfo act_info, bool mixed_layout = false, bool in_place = false, bool run_twice = false) { + ARM_COMPUTE_ERROR_ON(mixed_layout && in_place); + // This hash is used by random generators. There may be hash collisions but + // this is intentional as it's a very easy way to make the the current + // random generation process almost different for many test configurations, + // which were using the same set of values before. + _hash = in_shape[0] + in_shape[1] + in_shape[2] + in_shape[3] + + kernel_size.width + kernel_size.height + dilation.x() + + dilation.y() + pad_stride_info.pad_bottom() + pad_stride_info.pad_left() + pad_stride_info.pad_right() + pad_stride_info.pad_top(); + _mixed_layout = mixed_layout; _input_shape = in_shape; _input_data_type = input_data_type; _weights_data_type = weights_data_type; - _input_quantization_info = input_quantization_info; - _weights_quantization_info = weights_quantization_info; - _output_quantization_info = output_quantization_info; _data_layout = data_layout; _pad_stride_info = pad_stride_info; _act_info = act_info; _depth_multiplier = depth_multiplier; _dilation = dilation; + _in_place = in_place; + _run_twice = run_twice; _bias_data_type = is_data_type_quantized(_input_data_type) ? DataType::S32 : _input_data_type; @@ -85,6 +122,16 @@ public: _weights_shape.set(2, _output_shape.z()); _biases_shape = TensorShape(_weights_shape[2]); + + _input_quantization_info = input_quantization_info; + _weights_quantization_info = weights_quantization_info; + _output_quantization_info = output_quantization_info; + + if(is_data_type_quantized(_input_data_type) && !is_data_type_quantized_symmetric(weights_data_type) && (!act_info.enabled() || act_info.activation() == ActivationFunction::IDENTITY)) + { + setup_quantization(in_shape, _weights_shape, _input_quantization_info, _weights_quantization_info, _input_data_type); + _use_dynamic_output_quant = true; + } } void configure_target() @@ -101,13 +148,28 @@ public: } // Create tensors - _src = create_tensor<TensorType>(input_shape, _input_data_type, 1, _input_quantization_info, _data_layout); - _weights = create_tensor<TensorType>(weights_shape, _weights_data_type, 1, _weights_quantization_info, _data_layout); - _biases = create_tensor<TensorType>(_biases_shape, _bias_data_type, 1, _input_quantization_info, _data_layout); - _target = create_tensor<TensorType>(output_shape, _input_data_type, 1, _output_quantization_info, _data_layout); + _src = create_tensor<TensorType>(input_shape, _input_data_type, 1, _input_quantization_info, _data_layout); + _weights = create_tensor<TensorType>(weights_shape, _weights_data_type, 1, _weights_quantization_info, _data_layout); + if(_run_twice) { + _weights.info()->set_are_values_constant(false); + } + _biases = create_tensor<TensorType>(_biases_shape, _bias_data_type, 1, _input_quantization_info, _data_layout); + TensorType *target_to_use = nullptr; + if(!_in_place) + { + _target = create_tensor<TensorType>(output_shape, _input_data_type, 1, _output_quantization_info, _data_layout); + target_to_use = &_target; + } + + add_padding_x({ &_src, &_biases }, _data_layout); + add_padding_x({ &_weights }, _data_layout, true); + if(!_in_place) + { + add_padding_x({ &_target }, _data_layout); + } // Create Depthwise Convolution configure function - _dwc.configure(&_src, &_weights, &_biases, &_target, _pad_stride_info, _depth_multiplier, _act_info, _dilation); + _dwc.configure(&_src, &_weights, &_biases, target_to_use, _pad_stride_info, _depth_multiplier, _act_info, _dilation); ARM_COMPUTE_ASSERT(_src.info()->is_resizable()); ARM_COMPUTE_ASSERT(_weights.info()->is_resizable()); @@ -117,24 +179,35 @@ public: void allocate_and_run_target() { - // TODO: uncomment after COMPMID-4361 - // add_padding_x({ &_src, &_weights, &_biases, &_target }, _data_layout); - // Allocate tensors _src.allocator()->allocate(); _weights.allocator()->allocate(); _biases.allocator()->allocate(); - _target.allocator()->allocate(); ARM_COMPUTE_ASSERT(!_src.info()->is_resizable()); ARM_COMPUTE_ASSERT(!_weights.info()->is_resizable()); ARM_COMPUTE_ASSERT(!_biases.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!_target.info()->is_resizable()); + + if(!_in_place) + { + _target.allocator()->allocate(); + ARM_COMPUTE_ASSERT(!_target.info()->is_resizable()); + } // Fill tensors - fill(AccessorType(_src), 0); - fill(AccessorType(_weights), 1); - fill(AccessorType(_biases), 2); + fill(AccessorType(_src), 0 + _hash); + fill(AccessorType(_weights), 1 + _hash); + fill(AccessorType(_biases), 2 + _hash); + + // Run with variable input + if(_run_twice) { + _dwc.run(); + + // Fill tensors with a new seed + fill(AccessorType(_src), 3 + _hash); + fill(AccessorType(_weights), 4 + _hash); + fill(AccessorType(_biases), 5 + _hash); + } if(_mixed_layout) { @@ -153,9 +226,20 @@ public: SimpleTensor<TW> weights{ _weights_shape, _weights_data_type, 1, _weights_quantization_info }; SimpleTensor<TBias> biases{ _biases_shape, _bias_data_type, 1, _input_quantization_info }; - fill(src, 0); - fill(weights, 1); - fill(biases, 2); + fill(src, 0 + _hash); + fill(weights, 1 + _hash); + fill(biases, 2 + _hash); + + if(_run_twice) { + SimpleTensor<T> depth_out = reference::depthwise_convolution(src, weights, biases, _output_shape, _pad_stride_info, _depth_multiplier, _dilation, _output_quantization_info); + if(_act_info.enabled()) { + reference::activation_layer<T>(depth_out, _act_info); + } + + fill(src, 3 + _hash); + fill(weights, 4 + _hash); + fill(biases, 5 + _hash); + } SimpleTensor<T> depth_out = reference::depthwise_convolution(src, weights, biases, _output_shape, _pad_stride_info, _depth_multiplier, _dilation, _output_quantization_info); _reference = (_act_info.enabled()) ? reference::activation_layer<T>(depth_out, _act_info) : depth_out; @@ -164,6 +248,7 @@ public: protected: void mix_layout(FunctionType &layer, TensorType &src, TensorType &dst) { + ARM_COMPUTE_ERROR_ON(_in_place); // Test Multi DataLayout graph cases, when the data layout changes after configure src.info()->set_data_layout(_data_layout == DataLayout::NCHW ? DataLayout::NHWC : DataLayout::NCHW); dst.info()->set_data_layout(_data_layout == DataLayout::NCHW ? DataLayout::NHWC : DataLayout::NCHW); @@ -183,32 +268,77 @@ protected: { case DataType::QASYMM8: { - std::uniform_int_distribution<uint8_t> distribution(0, 10); - library->fill(tensor, distribution, i); + if(_use_dynamic_output_quant) + { + std::uniform_int_distribution<int32_t> distribution(0, 255); + library->fill(tensor, distribution, i); + } + else + { + // Legacy initialization in case the output quantization info can't be reliably estimated + std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); + std::uniform_int_distribution<uint32_t> distribution(bounds.first, bounds.second); + library->fill(tensor, distribution, i); + } break; } case DataType::QASYMM8_SIGNED: + { + if(_use_dynamic_output_quant) + { + std::uniform_int_distribution<int32_t> distribution(-128, 127); + library->fill(tensor, distribution, i); + } + else + { + // Legacy initialization in case the output quantization info can't be reliably estimated + std::pair<int, int> bounds = get_quantized_qasymm8_signed_bounds(tensor.quantization_info(), -1.0f, 1.0f); + std::uniform_int_distribution<int32_t> distribution(bounds.first, bounds.second); + library->fill(tensor, distribution, i); + } + break; + } case DataType::QSYMM8_PER_CHANNEL: { - std::uniform_int_distribution<int8_t> distribution(-10, 10); + int min_bound = 128; + int max_bound = -127; + for(size_t i = 0; i < _weights_quantization_info.scale().size(); i++) + { + std::pair<int, int> bounds = get_symm_quantized_per_channel_bounds(tensor.quantization_info(), -1.0f, 1.0f, i); + if(bounds.first < min_bound) + { + min_bound = bounds.first; + } + if(bounds.second > max_bound) + { + max_bound = bounds.second; + } + } + std::uniform_int_distribution<int32_t> distribution(min_bound, max_bound); library->fill(tensor, distribution, i); break; } - case DataType::F16: + case DataType::S32: { - arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; + std::uniform_int_distribution<int32_t> distribution(_min_bias, _max_bias); library->fill(tensor, distribution, i); break; } - case DataType::F32: + case DataType::BFLOAT16: { - std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + arm_compute::utils::uniform_real_distribution_16bit<bfloat16> distribution{ -1.0f, 1.0f }; library->fill(tensor, distribution, i); break; } - case DataType::S32: + case DataType::F16: { - std::uniform_int_distribution<int32_t> distribution(-100, 100); + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); library->fill(tensor, distribution, i); break; } @@ -241,19 +371,32 @@ protected: unsigned int _depth_multiplier{}; Size2D _dilation{}; bool _mixed_layout{ false }; + bool _in_place{ false }; + bool _run_twice{ false }; + bool _use_dynamic_output_quant{false}; + + int32_t _hash{0}; + // Random initialization limits + // Default values are previously handcrafted limits + // that sould be used when we don't use dynamic quantization + int32_t _min_bias{-100}; + int32_t _max_bias{100}; + int32_t _min_u8{0}; + int32_t _max_u8{50}; + int32_t _min_s8{-25}; + int32_t _max_s8{25}; }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false, bool in_place = false, bool run_twice = false> class DepthwiseConvolutionLayerValidationFixture : public DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info) { DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T>::setup(in_shape, kernel_size, pad_stride_info, dilation, depth_multiplier, data_type, data_type, QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), - data_layout, act_info, mixed_layout); + data_layout, act_info, mixed_layout, in_place, run_twice); } }; @@ -261,7 +404,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DepthwiseConvolutionLayerNativeValidationFixture : public DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(size_t width, size_t height, size_t channel, size_t batch, Size2D kernel_size, size_t depth_multiplier, Size2D dilation, Size2D stride, bool padding_valid, DataType data_type, DataLayout data_layout) { @@ -276,7 +418,7 @@ public: if(padding_valid) { - _conv_info = PadStrideInfo(); + _conv_info = PadStrideInfo(stride.width, stride.height); } else { @@ -301,6 +443,9 @@ public: _biases = create_tensor<TensorType>(_biases_shape, _data_type, 1, QuantizationInfo(), _data_layout); _target = create_tensor<TensorType>(TensorShape(), _data_type, 1, QuantizationInfo(), _data_layout); + add_padding_x({ &_src, &_biases, &_target }, _data_layout); + add_padding_x({ &_weights }, _data_layout, true); + // Create Depthwise Convolution configure function const ConvolutionInfo info { @@ -316,8 +461,6 @@ public: void allocate_and_run_target() { - add_padding_x({ &_src, &_weights, &_biases, &_target }, _data_layout); - // Allocate tensors _src.allocator()->allocate(); _weights.allocator()->allocate(); @@ -394,20 +537,21 @@ protected: unsigned int _depth_multiplier{}; }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool in_place = false> class DepthwiseConvolutionLayerNativeConfigurableValidationFixture : public DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(size_t width, size_t height, size_t channel, size_t batch, Size2D kernel_size, size_t depth_multiplier, Size2D dilation, Size2D stride, bool padding_valid, DataType data_type, - DataLayout data_layout, const ActivationLayerInfo &act_info, unsigned int n0) + DataLayout data_layout, const ActivationLayerInfo &act_info, unsigned int n0, bool export_to_cl_image) { - _dilation = dilation; - _depth_multiplier = depth_multiplier; - _data_type = data_type; - _data_layout = data_layout; - _act_info = act_info; - _n0 = n0; + _dilation = dilation; + _depth_multiplier = depth_multiplier; + _data_type = data_type; + _data_layout = data_layout; + _act_info = act_info; + _n0 = n0; + _export_to_cl_image = export_to_cl_image; + _in_place = in_place; _input_shape = TensorShape(width, height, channel, batch); _weights_shape = TensorShape(kernel_size.width, kernel_size.height, channel * _depth_multiplier); @@ -415,16 +559,29 @@ public: if(padding_valid) { - _conv_info = PadStrideInfo(); + _conv_info = calculate_same_pad(_input_shape, _weights_shape, PadStrideInfo(stride.width, stride.height), DataLayout::NCHW, _dilation); } else { - _conv_info = calculate_same_pad(_input_shape, _weights_shape, PadStrideInfo(stride.width, stride.height), DataLayout::NCHW, _dilation); + _conv_info = PadStrideInfo(stride.width, stride.height); } } void configure_target() { +#if defined(ARM_COMPUTE_OPENCL_ENABLED) + if(_export_to_cl_image) + { + _validate_output &= image2d_from_buffer_supported(CLKernelLibrary::get().get_device()); + _validate_output &= (get_cl_image_pitch_alignment(CLKernelLibrary::get().get_device()) != 0); + } +#endif // ARM_COMPUTE_OPENCL_ENABLED + + if(!_validate_output) + { + return; + } + TensorShape input_shape = _input_shape; TensorShape weights_shape = _weights_shape; @@ -435,19 +592,32 @@ public: } // Create tensors - _src = create_tensor<TensorType>(input_shape, _data_type, 1, QuantizationInfo(), _data_layout); - _weights = create_tensor<TensorType>(weights_shape, _data_type, 1, QuantizationInfo(), _data_layout); - _biases = create_tensor<TensorType>(_biases_shape, _data_type, 1, QuantizationInfo(), _data_layout); - _target = create_tensor<TensorType>(TensorShape(), _data_type, 1, QuantizationInfo(), _data_layout); + _src = create_tensor<TensorType>(input_shape, _data_type, 1, QuantizationInfo(), _data_layout); + _weights = create_tensor<TensorType>(weights_shape, _data_type, 1, QuantizationInfo(), _data_layout); + _biases = create_tensor<TensorType>(_biases_shape, _data_type, 1, QuantizationInfo(), _data_layout); + TensorType *target_to_use = nullptr; + if(!_in_place) + { + _target = create_tensor<TensorType>(TensorShape(), _data_type, 1, QuantizationInfo(), _data_layout); + target_to_use = &_target; + } + + DWCComputeKernelInfo dwc_info; + dwc_info.n0 = _n0; + dwc_info.m0 = _conv_info.stride().first == 1 && _dilation.x() == 1 ? 8 : 1; + dwc_info.export_input_to_cl_image = false; + dwc_info.export_weights_to_cl_image = _export_to_cl_image; - DWCWeightsKernelInfo dwc_weights_info; - dwc_weights_info.n0 = _n0; + const ConvolutionInfo conv_kernel_info + { + _conv_info, _depth_multiplier, _act_info, _dilation + }; - DWCKernelInfo dwc_info; - dwc_info.activation_info = _act_info; + add_padding_x({ &_src, &_biases, &_target }, _data_layout); + add_padding_x({ &_weights }, _data_layout, _export_to_cl_image); // Don't add left padding if cl image will be used // Create Depthwise Convolution configure function - _dwc.configure(&_src, &_weights, &_biases, &_target, dwc_weights_info, dwc_info, _conv_info, _depth_multiplier, _dilation); + _dwc.configure(&_src, &_weights, &_biases, target_to_use, dwc_info, conv_kernel_info); ARM_COMPUTE_ASSERT(_src.info()->is_resizable()); ARM_COMPUTE_ASSERT(_weights.info()->is_resizable()); @@ -457,18 +627,24 @@ public: void allocate_and_run_target() { - add_padding_x({ &_src, &_weights, &_biases, &_target }, _data_layout); + if(!_validate_output) + { + return; + } // Allocate tensors _src.allocator()->allocate(); _weights.allocator()->allocate(); _biases.allocator()->allocate(); - _target.allocator()->allocate(); ARM_COMPUTE_ASSERT(!_src.info()->is_resizable()); ARM_COMPUTE_ASSERT(!_weights.info()->is_resizable()); ARM_COMPUTE_ASSERT(!_biases.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!_target.info()->is_resizable()); + if(!_in_place) + { + _target.allocator()->allocate(); + ARM_COMPUTE_ASSERT(!_target.info()->is_resizable()); + } // Fill tensors fill(AccessorType(_src), 0); @@ -477,17 +653,28 @@ public: // Test Multi DataLayout graph cases, when the data layout changes after configure _src.info()->set_data_layout(_data_layout == DataLayout::NCHW ? DataLayout::NHWC : DataLayout::NCHW); - _target.info()->set_data_layout(_data_layout == DataLayout::NCHW ? DataLayout::NHWC : DataLayout::NCHW); + if(!_in_place) + { + _target.info()->set_data_layout(_data_layout == DataLayout::NCHW ? DataLayout::NHWC : DataLayout::NCHW); + } // Compute function _dwc.run(); // Reinstating original data layout for the test suite to properly check the values - _target.info()->set_data_layout(_data_layout); + if(!_in_place) + { + _target.info()->set_data_layout(_data_layout); + } } void compute_reference() { + if(!_validate_output) + { + return; + } + SimpleTensor<T> src{ _input_shape, _data_type }; SimpleTensor<T> weights{ _weights_shape, _data_type }; SimpleTensor<T> biases{ _biases_shape, _data_type }; @@ -542,27 +729,28 @@ protected: Size2D _dilation{}; unsigned int _depth_multiplier{}; unsigned int _n0{}; + bool _export_to_cl_image{}; + bool _validate_output{ true }; + bool _in_place{ false }; }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false, bool in_place = false> class DepthwiseConvolutionLayerValidationQuantizedFixture : public DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T> { public: - template <typename...> void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType data_type, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, DataLayout data_layout, ActivationLayerInfo act_info) { DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, T>::setup(in_shape, kernel_size, pad_stride_info, dilation, depth_multiplier, data_type, data_type, input_quantization_info, input_quantization_info, output_quantization_info, - data_layout, act_info, mixed_layout); + data_layout, act_info, mixed_layout, in_place); } }; -template <typename TensorType, typename AccessorType, typename FunctionType, typename T, typename TW> +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, typename TW, bool in_place = false> class DepthwiseConvolutionLayerValidationQuantizedPerChannelFixture : public DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, TW> { public: - template <typename...> void setup(TensorShape in_shape, Size2D kernel_size, PadStrideInfo pad_stride_info, Size2D dilation, unsigned int depth_multiplier, DataType input_data_type, DataType weights_data_type, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, DataLayout data_layout, ActivationLayerInfo act_info) { @@ -580,10 +768,10 @@ public: DepthwiseConvolutionLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T, TW>::setup(in_shape, kernel_size, pad_stride_info, dilation, depth_multiplier, input_data_type, weights_data_type, input_quantization_info, QuantizationInfo(weights_scales), output_quantization_info, - data_layout, act_info); + data_layout, act_info, false, in_place); } }; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_DEPTHWISE_CONVOLUTION_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_DEPTHWISECONVOLUTIONLAYERFIXTURE_H diff --git a/tests/validation/fixtures/DequantizationLayerFixture.h b/tests/validation/fixtures/DequantizationLayerFixture.h index af998bb740..4eb25a5bc5 100644 --- a/tests/validation/fixtures/DequantizationLayerFixture.h +++ b/tests/validation/fixtures/DequantizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DequantizationValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType src_data_type, DataType dst_datatype, DataLayout data_layout) { _quantization_info = generate_quantization_info(src_data_type, shape.z()); diff --git a/tests/validation/fixtures/DirectConvolution3DFixture.h b/tests/validation/fixtures/DirectConvolution3DFixture.h new file mode 100644 index 0000000000..e80ad2f54f --- /dev/null +++ b/tests/validation/fixtures/DirectConvolution3DFixture.h @@ -0,0 +1,189 @@ +/* + * Copyright (c) 2021, 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DIRECTCONVOLUTION3DFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DIRECTCONVOLUTION3DFIXTURE_H + +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "tests/framework/Asserts.h" // Required for ARM_COMPUTE_ASSERT +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/ActivationLayer.h" +#include "tests/validation/reference/Conv3D.h" + +#include <random> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::misc::shape_calculator; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolution3DValidationGenericFixture : public framework::Fixture +{ +public: + using TBias = typename std::conditional < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T >::type; + + void setup(const TensorShape &input_shape, int stride_x, int stride_y, int stride_z, int pad_x, int pad_y, int pad_z, unsigned int kernel_width, int kernel_height, int kernel_depth, + unsigned int num_kernels, bool has_bias, const ActivationLayerInfo &act_info, const DataType &data_type, const DataLayout &data_layout, + const QuantizationInfo &src_qinfo = QuantizationInfo(), const QuantizationInfo &weights_qinfo = QuantizationInfo(), const QuantizationInfo &dst_qinfo = QuantizationInfo()) + { + ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NDHWC); + + const TensorShape weights_shape(num_kernels, input_shape[0], kernel_width, kernel_height, kernel_depth); + const TensorShape bias_shape(num_kernels); + const DataType bias_data_type = is_data_type_quantized(data_type) ? DataType::S32 : data_type; + const Conv3dInfo conv3d_info(Size3D(stride_x, stride_y, stride_z), Padding3D(pad_x, pad_y, pad_z), act_info, Size3D(1U, 1U, 1U), DimensionRoundingType::FLOOR, false); + const TensorShape output_shape = compute_conv3d_shape(input_shape, weights_shape, conv3d_info); + + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, conv3d_info, has_bias, data_type, bias_data_type, data_layout, src_qinfo, weights_qinfo, dst_qinfo); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, conv3d_info, has_bias, data_type, bias_data_type, src_qinfo, weights_qinfo, dst_qinfo); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + + TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const Conv3dInfo &conv3d_info, + bool has_bias, const DataType &data_type, const DataType &bias_data_type, const DataLayout &data_layout, const QuantizationInfo &src_qinfo, + const QuantizationInfo &weights_qinfo, const QuantizationInfo &dst_qinfo) + { + // Create tensors + TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, src_qinfo, data_layout); + TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1, weights_qinfo, data_layout); + TensorType bias = has_bias ? create_tensor<TensorType>(bias_shape, bias_data_type, 1, QuantizationInfo()) : TensorType(); + TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, dst_qinfo, data_layout); + + // Create and configure function + FunctionType conv{}; + conv.configure(&src, &weights, has_bias ? &bias : nullptr, &dst, conv3d_info); + + ARM_COMPUTE_ASSERT(src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(weights.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + // Allocate tensors + src.allocator()->allocate(); + weights.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!weights.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(src), 0); + fill(AccessorType(weights), 1); + + if(has_bias) + { + ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); + bias.allocator()->allocate(); + ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); + fill(AccessorType(bias), 2); + } + + // Compute Direct Convolution 3D function + conv.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, + const Conv3dInfo &conv3d_info, bool has_bias, const DataType &data_type, const DataType &bias_data_type, const QuantizationInfo &src_qinfo, + const QuantizationInfo &weights_qinfo, const QuantizationInfo &dst_qinfo) + { + // Create reference + SimpleTensor<T> src{ input_shape, data_type, 1, src_qinfo }; + SimpleTensor<T> weights{ weights_shape, data_type, 1, weights_qinfo }; + SimpleTensor<TBias> bias{ bias_shape, bias_data_type }; + SimpleTensor<T> dst{ output_shape, data_type, 1, dst_qinfo }; + + // Fill reference + fill(src, 0); + fill(weights, 1); + + if(has_bias) + { + fill(bias, 2); + } + + return reference::activation_layer(reference::conv3d<T, TBias>(src, weights, bias, dst, conv3d_info), conv3d_info.act_info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolution3DValidationFixture : public DirectConvolution3DValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, int stride_x, int stride_y, int stride_z, int pad_x, int pad_y, int pad_z, unsigned int kernel_width, int kernel_height, int kernel_depth, + unsigned int num_kernels, bool has_bias, ActivationLayerInfo act_info, DataType data_type, DataLayout data_layout) + { + DirectConvolution3DValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, stride_z, pad_x, pad_y, pad_z, kernel_width, kernel_height, + kernel_depth, num_kernels, has_bias, act_info, data_type, data_layout); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolution3DValidationQuantizedFixture : public DirectConvolution3DValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, int stride_x, int stride_y, int stride_z, int pad_x, int pad_y, int pad_z, unsigned int kernel_width, int kernel_height, int kernel_depth, + unsigned int num_kernels, bool has_bias, ActivationLayerInfo act_info, DataType data_type, DataLayout data_layout, QuantizationInfo src_qinfo, QuantizationInfo weights_qinfo, + QuantizationInfo dst_qinfo) + { + DirectConvolution3DValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, stride_z, pad_x, pad_y, pad_z, kernel_width, kernel_height, + kernel_depth, num_kernels, has_bias, act_info, data_type, data_layout, src_qinfo, + weights_qinfo, dst_qinfo); + } +}; +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_DIRECTCONVOLUTION3DFIXTURE_H diff --git a/tests/validation/fixtures/DirectConvolutionLayerFixture.h b/tests/validation/fixtures/DirectConvolutionLayerFixture.h index 614aa20753..6f204642ca 100644 --- a/tests/validation/fixtures/DirectConvolutionLayerFixture.h +++ b/tests/validation/fixtures/DirectConvolutionLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,6 +21,10 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ + +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DIRECTCONVOLUTIONLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DIRECTCONVOLUTIONLAYERFIXTURE_H + #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -51,11 +55,52 @@ class DirectConvolutionValidationGenericFixture : public framework::Fixture public: using TBias = typename std::conditional < std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T >::type; - template <typename...> + void setup_quantization(const TensorShape &input_shape, const TensorShape &weights_shape, QuantizationInfo &input_q_info, + QuantizationInfo &weights_q_info, DataType data_type) + { + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + _hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + input_q_info = QuantizationInfo(scale_lhs, offset_lhs); + weights_q_info = QuantizationInfo(scale_rhs, offset_rhs); + + QuantizationHint q_hint = suggest_conv_dst_q_info_and_bias(input_q_info, weights_q_info, + weights_shape.y() /* heights */, weights_shape.x() /* width */, input_shape.z() /* channels */, + data_type, 0.5f /* bias_fraction */); + + _dst_q_info = q_hint.q_info; + _min_bias = q_hint.bias_min; + _max_bias = q_hint.bias_max; + + // Do not change here as these limits are the natural limits of the associated data types and + // are embeded in the computation of the dst quantization info. + _min_u8 = 0; + _max_u8 = 255; + _min_s8 = -128; + _max_s8 = 127; + } + void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info, DataLayout data_layout, bool mixed_layout = false) { - _quantization_info = quantization_info; + // This hash is used by random generators. There may be hash collisions but + // this is intentional as it's a very easy way to make the the current + // random generation process almost different for many test configurations, + // which were using the same set of values before. + _hash = input_shape[0] + input_shape[1] + input_shape[2] + input_shape[3] + + stride_x + stride_y + pad_x + pad_y + kernel_size + num_kernels + mixed_layout + + (data_layout == DataLayout::NHWC); + _data_type = data_type; _mixed_layout = mixed_layout; @@ -69,24 +114,48 @@ public: const TensorShape output_shape = compute_deep_convolution_shape(input_info, weights_info, info); - _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, quantization_info, act_info, data_layout); - _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, quantization_info, act_info); + QuantizationInfo input_q_info = quantization_info; + QuantizationInfo weights_q_info = quantization_info; + _dst_q_info = quantization_info; + + if(is_data_type_quantized(data_type) && (!act_info.enabled() || act_info.activation() == ActivationFunction::IDENTITY)) + { + setup_quantization(input_shape, weights_shape, input_q_info, weights_q_info, data_type); + } + + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, input_q_info, weights_q_info, act_info, data_layout); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, input_q_info, weights_q_info, act_info); } - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info, DataLayout data_layout) { ARM_COMPUTE_ERROR_ON(data_layout == DataLayout::UNKNOWN); ARM_COMPUTE_UNUSED(dilation); - _quantization_info = quantization_info; + // This hash is used by random generators. There may be hash collisions but + // this is intentional as it's a very easy way to make the the current + // random generation process almost different for many test configurations, + // which were using the same set of values before. + _hash = input_shape[0] + input_shape[1] + input_shape[2] + input_shape[3] + + weights_shape[0] + weights_shape[1] + weights_shape[2] + weights_shape[3] + dilation.x() + + dilation.y() + info.pad_bottom() + info.pad_left() + info.pad_right() + info.pad_top(); + _data_type = data_type; const DataType bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; - _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, quantization_info, act_info, data_layout); - _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, quantization_info, act_info); + QuantizationInfo input_q_info = quantization_info; + QuantizationInfo weights_q_info = quantization_info; + _dst_q_info = quantization_info; + + if(is_data_type_quantized(data_type) && (!act_info.enabled() || act_info.activation() == ActivationFunction::IDENTITY)) + { + setup_quantization(input_shape, weights_shape, input_q_info, weights_q_info, data_type); + } + + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, input_q_info, weights_q_info, act_info, data_layout); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, input_q_info, weights_q_info, act_info); } protected: @@ -112,14 +181,14 @@ protected: { case DataType::QASYMM8: { - std::uniform_int_distribution<uint8_t> distribution(0, 50); + std::uniform_int_distribution<uint32_t> distribution(_min_u8, _max_u8); library->fill(tensor, distribution, i); break; } case DataType::QASYMM8_SIGNED: { // Use small input range to avoid all the test results being saturated at the end. - std::uniform_int_distribution<int8_t> distribution(-25, 25); + std::uniform_int_distribution<int32_t> distribution(_min_s8, _max_s8); library->fill(tensor, distribution, i); break; } @@ -137,7 +206,7 @@ protected: } case DataType::S32: { - std::uniform_int_distribution<int32_t> distribution(-5, 5); + std::uniform_int_distribution<int32_t> distribution(_min_bias, _max_bias); library->fill(tensor, distribution, i); break; } @@ -147,7 +216,7 @@ protected: } TensorType compute_target(TensorShape input_shape, TensorShape weights_shape, const TensorShape &bias_shape, TensorShape output_shape, const PadStrideInfo &info, - DataType data_type, DataType bias_data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info, const DataLayout &data_layout) + DataType data_type, DataType bias_data_type, QuantizationInfo input_q_info, QuantizationInfo weights_q_info, ActivationLayerInfo act_info, const DataLayout &data_layout) { if(data_layout == DataLayout::NHWC) { @@ -157,10 +226,10 @@ protected: } // Create tensors - TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, quantization_info, data_layout); - TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1, quantization_info, data_layout); - TensorType bias = create_tensor<TensorType>(bias_shape, bias_data_type, 1, quantization_info); - TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, quantization_info, data_layout); + TensorType src = create_tensor<TensorType>(input_shape, data_type, 1, input_q_info, data_layout); + TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1, weights_q_info, data_layout); + TensorType bias = create_tensor<TensorType>(bias_shape, bias_data_type, 1, QuantizationInfo()); + TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, _dst_q_info, data_layout); add_padding_x({ &src, &bias, &dst }, data_layout); add_padding_x({ &weights }, data_layout, input_shape[0] % 4 == 0); // Don't add left padding if cl image will be used @@ -186,9 +255,9 @@ protected: ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); // Fill tensors - fill(AccessorType(src), 0); - fill(AccessorType(weights), 1); - fill(AccessorType(bias), 2); + fill(AccessorType(src), 0 + _hash); + fill(AccessorType(weights), 1 + _hash); + fill(AccessorType(bias), 2 + _hash); if(_mixed_layout) { @@ -204,33 +273,45 @@ protected: } SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, - DataType data_type, DataType bias_data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info) + DataType data_type, DataType bias_data_type, QuantizationInfo input_q_info, QuantizationInfo weights_q_info, ActivationLayerInfo act_info) { // Create reference - SimpleTensor<T> src{ input_shape, data_type, 1, quantization_info }; - SimpleTensor<T> weights{ weights_shape, data_type, 1, quantization_info }; - SimpleTensor<TBias> bias{ bias_shape, bias_data_type, 1, quantization_info }; + SimpleTensor<T> src{ input_shape, data_type, 1, input_q_info }; + SimpleTensor<T> weights{ weights_shape, data_type, 1, weights_q_info }; + SimpleTensor<TBias> bias{ bias_shape, bias_data_type, 1, QuantizationInfo() }; // Fill reference - fill(src, 0); - fill(weights, 1); - fill(bias, 2); - - SimpleTensor<T> dst = reference::convolution_layer<T>(src, weights, bias, output_shape, info); - return (act_info.enabled()) ? reference::activation_layer<T>(dst, act_info) : dst; + fill(src, 0 + _hash); + fill(weights, 1 + _hash); + fill(bias, 2 + _hash); + + SimpleTensor<T> dst = reference::convolution_layer<T>(src, weights, bias, output_shape, info, + Size2D(1U, 1U) /* dilation */, 1 /* num_groups */, _dst_q_info); + SimpleTensor<T> dst2 = (act_info.enabled()) ? reference::activation_layer<T>(dst, act_info) : dst; + return dst2; } TensorType _target{}; SimpleTensor<T> _reference{}; - QuantizationInfo _quantization_info{}; + QuantizationInfo _dst_q_info{}; DataType _data_type{}; bool _mixed_layout{ false }; + int32_t _hash{0}; + + // Random initialization limits + // Default values are previously handcrafted limits + // that sould be used when we don't use dynamic quantization + int32_t _min_bias{-5}; + int32_t _max_bias{5}; + int32_t _min_u8{0}; + int32_t _max_u8{50}; + int32_t _min_s8{-25}; + int32_t _max_s8{25}; }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> class DirectConvolutionValidationFixture : public DirectConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type, ActivationLayerInfo act_info, DataLayout data_layout) { @@ -243,7 +324,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DirectConvolutionValidationQuantizedFixture : public DirectConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info, DataLayout data_layout) { @@ -256,7 +336,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DirectConvolutionValidationWithTensorShapesQuantizedFixture : public DirectConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo act_info, DataLayout data_layout) { @@ -269,7 +348,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DirectConvolutionValidationWithTensorShapesFixture : public DirectConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, ActivationLayerInfo act_info) { @@ -281,3 +359,5 @@ public: } // namespace validation } // namespace test } // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_DIRECTCONVOLUTIONLAYERFIXTURE_H diff --git a/tests/validation/fixtures/DropoutLayerFixture.h b/tests/validation/fixtures/DropoutLayerFixture.h index 2a91911410..a84f2a6407 100644 --- a/tests/validation/fixtures/DropoutLayerFixture.h +++ b/tests/validation/fixtures/DropoutLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class DropoutLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, float ratio, bool forward, DataType data_type) { _target = compute_target(shape, ratio, forward, data_type); diff --git a/tests/validation/fixtures/ElementwiseOperationsFixture.h b/tests/validation/fixtures/ElementwiseOperationsFixture.h index 352720c03b..f36a1f75b7 100644 --- a/tests/validation/fixtures/ElementwiseOperationsFixture.h +++ b/tests/validation/fixtures/ElementwiseOperationsFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,11 +21,12 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_ELEMENTWISE_OPERATIONS_FIXTURE -#define ARM_COMPUTE_TEST_ELEMENTWISE_OPERATIONS_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_ELEMENTWISEOPERATIONSFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_ELEMENTWISEOPERATIONSFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" +#include "arm_compute/core/Validate.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/IAccessor.h" @@ -45,13 +46,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticOperationsGenericFixture : public framework::Fixture { public: - template <typename...> void setup(ArithmeticOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool use_dyanmic_shape = false) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace = false, bool use_dynamic_shape = false) { _op = op; - _use_dynamic_shape = use_dyanmic_shape; + _use_dynamic_shape = use_dynamic_shape; + _is_inplace = is_inplace; _target = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, qinfo0, qinfo1, qinfo_out); _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, qinfo0, qinfo1, qinfo_out); @@ -85,9 +86,29 @@ protected: QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) { // Create tensors - TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, qinfo0); - TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, qinfo1); - TensorType dst = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, qinfo_out); + const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1); + TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, qinfo0); + TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, qinfo1); + TensorType dst = create_tensor<TensorType>(out_shape, output_data_type, 1, qinfo_out); + + // Check whether do in-place computation and whether inputs are broadcast compatible + TensorType *actual_dst = &dst; + if(_is_inplace) + { + bool src1_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape0, 0) && (qinfo0 == qinfo_out) && (data_type0 == output_data_type); + bool src2_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape1, 0) && (qinfo1 == qinfo_out) && (data_type1 == output_data_type); + bool do_in_place = out_shape.total_size() != 0 && (src1_is_inplace || src2_is_inplace); + ARM_COMPUTE_ASSERT(do_in_place); + + if(src1_is_inplace) + { + actual_dst = &ref_src1; + } + else + { + actual_dst = &ref_src2; + } + } // if _use_dynamic_shape is true, this fixture will test scenario for dynamic shapes. // - At configure time, all input tensors are marked as dynamic using set_tensor_dynamic() @@ -101,7 +122,7 @@ protected: // Create and configure function FunctionType elem_op; - elem_op.configure(&ref_src1, &ref_src2, &dst); + elem_op.configure(&ref_src1, &ref_src2, actual_dst); if(_use_dynamic_shape) { @@ -111,16 +132,21 @@ protected: ARM_COMPUTE_ASSERT(ref_src1.info()->is_resizable()); ARM_COMPUTE_ASSERT(ref_src2.info()->is_resizable()); - ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); // Allocate tensors ref_src1.allocator()->allocate(); ref_src2.allocator()->allocate(); - dst.allocator()->allocate(); + + // If don't do in-place computation, still need to allocate original dst + if(!_is_inplace) + { + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + dst.allocator()->allocate(); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + } ARM_COMPUTE_ASSERT(!ref_src1.info()->is_resizable()); ARM_COMPUTE_ASSERT(!ref_src2.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); // Fill tensors fill(AccessorType(ref_src1), 0); @@ -129,7 +155,7 @@ protected: // Compute function elem_op.run(); - return dst; + return std::move(*actual_dst); } SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, @@ -152,6 +178,7 @@ protected: SimpleTensor<T> _reference{}; ArithmeticOperation _op{ ArithmeticOperation::ADD }; bool _use_dynamic_shape{ false }; + bool _is_inplace{ false }; }; // Arithmetic operation fused with activation function @@ -159,15 +186,15 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticOperationsFuseActivationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(ArithmeticOperation op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, ActivationLayerInfo act_info) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, ActivationLayerInfo act_info, bool is_inplace = true) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(op, shape0, shape1, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); - _act_info = act_info; + qinfo0, qinfo1, qinfo_out, is_inplace); + _act_info = act_info; + _is_inplace = is_inplace; } protected: @@ -175,26 +202,51 @@ protected: QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) { // Create tensors - TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, qinfo0); - TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, qinfo1); - TensorType dst = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, qinfo_out); + const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1); + TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, qinfo0); + TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, qinfo1); + TensorType dst = create_tensor<TensorType>(out_shape, output_data_type, 1, qinfo_out); + + // Check whether do in-place computation and whether inputs are broadcast compatible + TensorType *actual_dst = &dst; + if(_is_inplace) + { + bool src1_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape0, 0) && (qinfo0 == qinfo_out) && (data_type0 == output_data_type); + bool src2_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape1, 0) && (qinfo1 == qinfo_out) && (data_type1 == output_data_type); + bool do_in_place = out_shape.total_size() != 0 && (src1_is_inplace || src2_is_inplace); + ARM_COMPUTE_ASSERT(do_in_place); + + if(src1_is_inplace) + { + actual_dst = &ref_src1; + } + else + { + actual_dst = &ref_src2; + } + } // Create and configure function FunctionType elem_op; - elem_op.configure(&ref_src1, &ref_src2, &dst, _act_info); + elem_op.configure(&ref_src1, &ref_src2, actual_dst, _act_info); ARM_COMPUTE_ASSERT(ref_src1.info()->is_resizable()); ARM_COMPUTE_ASSERT(ref_src2.info()->is_resizable()); - ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); // Allocate tensors ref_src1.allocator()->allocate(); ref_src2.allocator()->allocate(); - dst.allocator()->allocate(); + + // If don't do in-place computation, still need to allocate original dst + if(!_is_inplace) + { + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + dst.allocator()->allocate(); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + } ARM_COMPUTE_ASSERT(!ref_src1.info()->is_resizable()); ARM_COMPUTE_ASSERT(!ref_src2.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); // Fill tensors fill(AccessorType(ref_src1), 0); @@ -203,7 +255,7 @@ protected: // Compute function elem_op.run(); - return dst; + return std::move(*actual_dst); } SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, @@ -216,18 +268,18 @@ protected: } ActivationLayerInfo _act_info{}; + bool _is_inplace{ false }; }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class ArithmeticDivisionBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -235,12 +287,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -248,12 +299,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionBroadcastDynamicShapeValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), true); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace, true /* use_dynamic_shape */); } }; @@ -261,12 +311,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionDynamicShapeValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), true); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace, true /* use_dynamic_shape */); } }; @@ -274,12 +323,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionBroadcastValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -287,12 +335,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -300,12 +347,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionValidationIntegerFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -313,14 +359,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ArithmeticDivisionValidationQuantizedFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::DIV, shape, shape, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -328,12 +373,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMaxBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MAX, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -341,12 +385,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMaxValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MAX, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -354,12 +397,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMaxBroadcastValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MAX, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -367,12 +409,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMaxValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MAX, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -380,14 +421,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMaxValidationQuantizedFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MAX, shape, shape, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -395,14 +435,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMaxQuantizedBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MAX, shape0, shape1, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -410,12 +449,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMinBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MIN, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -423,12 +461,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMinValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MIN, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -436,12 +473,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMinBroadcastValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MIN, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -449,12 +485,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMinValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MIN, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -462,14 +497,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMinValidationQuantizedFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MIN, shape, shape, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -477,14 +511,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseMinQuantizedBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::MIN, shape0, shape1, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -492,12 +525,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseSquaredDiffBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::SQUARED_DIFF, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -505,12 +537,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseSquaredDiffValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::SQUARED_DIFF, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -518,12 +549,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseSquaredDiffBroadcastValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::SQUARED_DIFF, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -531,12 +561,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseSquaredDiffValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::SQUARED_DIFF, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -544,14 +573,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseSquaredDiffValidationQuantizedFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::SQUARED_DIFF, shape, shape, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -559,14 +587,13 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwiseSquaredDiffQuantizedBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, - QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) + QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::SQUARED_DIFF, shape0, shape1, data_type0, data_type1, output_data_type, - qinfo0, qinfo1, qinfo_out); + qinfo0, qinfo1, qinfo_out, is_inplace); } }; @@ -574,7 +601,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PReluLayerBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::PRELU, shape0, shape1, @@ -587,7 +613,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PReluLayerValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::PRELU, shape, shape, @@ -600,7 +625,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PReluLayerValidationQuantizedFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) @@ -615,7 +639,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PReluLayerQuantizedBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out) @@ -630,12 +653,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwisePowerBroadcastValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::POWER, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -643,12 +665,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwisePowerValidationFixture : public ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, bool is_inplace) { ArithmeticOperationsGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::POWER, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo()); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), is_inplace); } }; @@ -656,12 +677,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwisePowerBroadcastValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::POWER, shape0, shape1, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; @@ -669,16 +689,15 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementwisePowerValidationFloatFixture : public ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ActivationLayerInfo act_info, bool is_inplace) { ArithmeticOperationsFuseActivationFixture<TensorType, AccessorType, FunctionType, T>::setup(ArithmeticOperation::POWER, shape, shape, data_type0, data_type1, output_data_type, - QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info); + QuantizationInfo(), QuantizationInfo(), QuantizationInfo(), act_info, is_inplace); } }; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_ARITHMETIC_OPERATIONS_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_ELEMENTWISEOPERATIONSFIXTURE_H diff --git a/tests/validation/fixtures/ElementwiseUnaryFixture.h b/tests/validation/fixtures/ElementwiseUnaryFixture.h index 7221226fd1..15344288db 100644 --- a/tests/validation/fixtures/ElementwiseUnaryFixture.h +++ b/tests/validation/fixtures/ElementwiseUnaryFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,8 +24,10 @@ #ifndef ARM_COMPUTE_TEST_ELEMENTWISE_UNARY_FIXTURE #define ARM_COMPUTE_TEST_ELEMENTWISE_UNARY_FIXTURE +#include "arm_compute/core/QuantizationInfo.h" #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" +#include "arm_compute/core/Utils.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/IAccessor.h" @@ -33,6 +35,11 @@ #include "tests/framework/Fixture.h" #include "tests/validation/reference/ElementwiseUnary.h" +#include <tuple> +#include <limits> +#include <type_traits> +#include <vector> + namespace arm_compute { namespace test @@ -43,12 +50,12 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ElementWiseUnaryValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape input_shape, DataType input_data_type, bool in_place, ElementWiseUnary op, bool use_dynamic_shape = false) + void setup(TensorShape input_shape, DataType input_data_type, bool in_place, ElementWiseUnary op, + bool use_dynamic_shape = false, QuantizationInfo qinfo = QuantizationInfo(), QuantizationInfo qinfo_out = QuantizationInfo()) { _op = op; - _target = compute_target(input_shape, input_data_type, in_place); - _reference = compute_reference(input_shape, input_data_type); + _target = compute_target(input_shape, input_data_type, in_place, qinfo, qinfo_out); + _reference = compute_reference(input_shape, input_data_type, qinfo, qinfo_out); _use_dynamic_shape = use_dynamic_shape; } @@ -63,60 +70,131 @@ protected: { case ElementWiseUnary::EXP: { - FloatDistributionType distribution{ FloatType(-1.0f), FloatType(1.0f) }; - library->fill(tensor, distribution, i); + switch(data_type) + { + case DataType::F32: + { + FloatDistributionType distribution{ FloatType(-86.63f), FloatType(88.36f) }; + library->fill(tensor, distribution, i); + break; + } + + case DataType::F16: + { + FloatDistributionType distribution{ FloatType(-9.00f), FloatType(10.73f) }; + library->fill(tensor, distribution, i); + break; + } + + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + library->fill_tensor_uniform(tensor, i); + break; + + default: + ARM_COMPUTE_ERROR("Not implemented"); + } + break; } case ElementWiseUnary::RSQRT: + case ElementWiseUnary::LOG: { - FloatDistributionType distribution{ FloatType(1.0f), FloatType(2.0f) }; - library->fill(tensor, distribution, i); - break; - } - case ElementWiseUnary::ABS: - case ElementWiseUnary::NEG: - { + // For floating-point data type, the chosen input range is all positive numbers + // (i.e. positive and negative zeros are excluded). switch(data_type) { + case DataType::F32: + { + FloatDistributionType distribution{ std::numeric_limits<float>::min(), std::numeric_limits<float>::max() }; + library->fill(tensor, distribution, i); + break; + } + case DataType::F16: { - arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -2.0f, 2.0f }; + FloatDistributionType distribution{ FloatType(0.00006103515625f), FloatType(65504.0f) }; library->fill(tensor, distribution, i); break; } + + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + library->fill_tensor_uniform(tensor, i); + break; + + default: + ARM_COMPUTE_ERROR("Not implemented"); + } + + break; + } + case ElementWiseUnary::SIN: + { + switch(data_type) + { case DataType::F32: + case DataType::F16: { - FloatDistributionType distribution{ FloatType(-2.0f), FloatType(2.0f) }; + FloatDistributionType distribution{ FloatType(-100.0f), FloatType(100.0f) }; library->fill(tensor, distribution, i); break; } + case DataType::S32: { - std::uniform_int_distribution<int32_t> distribution(-100, 100); + std::uniform_int_distribution<int32_t> distribution(std::numeric_limits<int32_t>::lowest(), std::numeric_limits<int32_t>::max()); library->fill(tensor, distribution, i); break; } + + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + library->fill_tensor_uniform(tensor, i); + break; + default: - ARM_COMPUTE_ERROR("DataType for Elementwise Negation Not implemented"); + ARM_COMPUTE_ERROR("Not implemented"); } + break; } - case ElementWiseUnary::LOG: - { - FloatDistributionType distribution{ FloatType(0.0000001f), FloatType(100.0f) }; - library->fill(tensor, distribution, i); - break; - } - case ElementWiseUnary::SIN: - { - FloatDistributionType distribution{ FloatType(-100.00f), FloatType(100.00f) }; - library->fill(tensor, distribution, i); - break; - } + case ElementWiseUnary::ABS: + case ElementWiseUnary::NEG: case ElementWiseUnary::ROUND: { - FloatDistributionType distribution{ FloatType(100.0f), FloatType(-100.0f) }; - library->fill(tensor, distribution, i); + switch(data_type) + { + case DataType::F32: + { + FloatDistributionType distribution{ std::numeric_limits<float>::lowest() / 2, std::numeric_limits<float>::max() / 2 }; + library->fill(tensor, distribution, i); + break; + } + + case DataType::F16: + { + FloatDistributionType distribution{ FloatType(-65504.0f), FloatType(65504.0f) }; + library->fill(tensor, distribution, i); + break; + } + + case DataType::S32: + { + std::uniform_int_distribution<int32_t> distribution(std::numeric_limits<int32_t>::lowest(), std::numeric_limits<int32_t>::max()); + library->fill(tensor, distribution, i); + break; + } + + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + library->fill_tensor_uniform(tensor, i); + break; + + default: + ARM_COMPUTE_ERROR("Not implemented"); + } + break; } default: @@ -124,12 +202,11 @@ protected: } } - TensorType compute_target(const TensorShape &shape, DataType data_type, bool in_place) + TensorType compute_target(const TensorShape &shape, DataType data_type, bool in_place, QuantizationInfo qinfo, QuantizationInfo qinfo_out) { // Create tensors - TensorType src = create_tensor<TensorType>(shape, data_type); - TensorType dst = create_tensor<TensorType>(shape, data_type); - + TensorType src = create_tensor<TensorType>(shape, data_type, 1, qinfo); + TensorType dst = create_tensor<TensorType>(shape, data_type, 1, qinfo_out); TensorType *actual_dst = in_place ? &src : &dst; // if _use_dynamic_shape is true, this fixture will test scenario for dynamic shapes. @@ -176,28 +253,39 @@ protected: } } - SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type) + SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, QuantizationInfo qinfo, QuantizationInfo qinfo_out) { // Create reference - SimpleTensor<T> src{ shape, data_type }; + SimpleTensor<T> src{ shape, data_type, 1, qinfo }; + SimpleTensor<T> dst{ shape, data_type, 1, qinfo_out }; // Fill reference fill(src, 0, data_type); - return reference::elementwise_unary<T>(src, _op); + return reference::elementwise_unary<T>(src, dst, _op); } TensorType _target{}; SimpleTensor<T> _reference{}; ElementWiseUnary _op{}; bool _use_dynamic_shape{ false }; + QuantizationInfo _input_qinfo{}; + QuantizationInfo _output_qinfo{}; +}; +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class RsqrtQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo qinfo, QuantizationInfo qinfo_out) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::RSQRT, false, qinfo, qinfo_out); + } }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class RsqrtValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::RSQRT); @@ -208,7 +296,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class RsqrtDynamicShapeValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::RSQRT, true); @@ -219,7 +306,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ExpValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::EXP); @@ -227,10 +313,19 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ExpQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::EXP, false, iq, oq); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class NegValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::NEG); @@ -238,10 +333,19 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class NegQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::NEG, false, iq, oq); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class NegValidationInPlaceFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type, bool in_place) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, in_place, ElementWiseUnary::NEG); @@ -249,10 +353,19 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class NegQuantizedValidationInPlaceFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, bool in_place, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, in_place, ElementWiseUnary::NEG, false, iq, oq); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class LogValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::LOG); @@ -260,10 +373,19 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class LogQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::LOG, false, iq, oq); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class AbsValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::ABS); @@ -271,10 +393,19 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class AbsQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::ABS, false, iq, oq); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class SinValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::SIN); @@ -282,15 +413,34 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class SinQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::SIN, false, iq, oq); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class RoundValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(const TensorShape &shape, DataType data_type) { ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::ROUND); } }; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class RoundQuantizedValidationFixture : public ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape, DataType data_type, QuantizationInfo iq, QuantizationInfo oq) + { + ElementWiseUnaryValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, false, ElementWiseUnary::ROUND, false, iq, oq); + } +}; } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/fixtures/FFTFixture.h b/tests/validation/fixtures/FFTFixture.h index a70335b6f1..024227b22a 100644 --- a/tests/validation/fixtures/FFTFixture.h +++ b/tests/validation/fixtures/FFTFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FFTValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); @@ -134,7 +133,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FFTConvolutionValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info, bool mixed_layout = false) { @@ -202,6 +200,8 @@ protected: TensorType bias = create_tensor<TensorType>(bias_shape, _data_type, 1, QuantizationInfo(), _data_layout); TensorType dst = create_tensor<TensorType>(output_shape, _data_type, 1, QuantizationInfo(), _data_layout); + add_padding_x({ &src, &weights, &bias, &dst }, _data_layout); + // Create and configure function FunctionType conv; conv.configure(&src, &weights, &bias, &dst, info, act_info, _data_type == DataType::F16); @@ -211,8 +211,6 @@ protected: ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); - add_padding_x({ &src, &weights, &bias, &dst }, _data_layout); - // Allocate tensors src.allocator()->allocate(); weights.allocator()->allocate(); @@ -271,7 +269,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FFTConvolutionValidationFixture : public FFTConvolutionValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, DataLayout data_layout, ActivationLayerInfo act_info) { diff --git a/tests/validation/fixtures/FillFixture.h b/tests/validation/fixtures/FillFixture.h index 706c13565d..0239a68903 100644 --- a/tests/validation/fixtures/FillFixture.h +++ b/tests/validation/fixtures/FillFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 Arm Limited. + * Copyright (c) 2019, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,7 +42,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FillFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, DataType data_type) { _target = compute_target(input_shape, data_type); diff --git a/tests/validation/fixtures/FlattenLayerFixture.h b/tests/validation/fixtures/FlattenLayerFixture.h index b6c0dde28a..e72487c7cf 100644 --- a/tests/validation/fixtures/FlattenLayerFixture.h +++ b/tests/validation/fixtures/FlattenLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FlattenLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { TensorShape shape_flatten; diff --git a/tests/validation/fixtures/FloorFixture.h b/tests/validation/fixtures/FloorFixture.h index d6e19444db..7d38666f47 100644 --- a/tests/validation/fixtures/FloorFixture.h +++ b/tests/validation/fixtures/FloorFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FloorValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); diff --git a/tests/validation/fixtures/FullyConnectedLayerFixture.h b/tests/validation/fixtures/FullyConnectedLayerFixture.h index 7d767642f3..344187868f 100644 --- a/tests/validation/fixtures/FullyConnectedLayerFixture.h +++ b/tests/validation/fixtures/FullyConnectedLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_FIXTURE -#define ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_FULLYCONNECTEDLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_FULLYCONNECTEDLAYERFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -34,6 +34,7 @@ #include "tests/framework/Asserts.h" #include "tests/framework/Fixture.h" #include "tests/validation/Helpers.h" +#include "tests/validation/Validation.h" #include "tests/validation/reference/ActivationLayer.h" #include "tests/validation/reference/FullyConnectedLayer.h" #include "tests/validation/reference/Utils.h" @@ -54,7 +55,40 @@ public: using TBias = typename std::conditional < (std::is_same<TDecay, uint8_t>::value || std::is_same<TDecay, int8_t>::value), int32_t, T >::type; public: - template <typename...> + void setup_quantization(TensorShape weights_shape, TensorShape output_shape, QuantizationInfo &input_q_info, QuantizationInfo &weights_q_info, DataType data_type) + { + _hash = weights_shape[0] + weights_shape[1] + output_shape[0] + output_shape[1]; + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + _hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + input_q_info = QuantizationInfo(scale_lhs, offset_lhs); + weights_q_info = QuantizationInfo(scale_rhs, offset_rhs); + + + const int k = weights_shape.x(); + QuantizationHint q_hint = suggest_mac_dst_q_info_and_bias(input_q_info, weights_q_info, k, data_type, 0.1f /* bias_fraction */, 4 /* number of standard deviations*/); + + _dst_q_info = q_hint.q_info; + _min_bias = q_hint.bias_min; + _max_bias = q_hint.bias_max; + + // Do not change here as these limits are the natural limits of the associated data types and + // are embedded in the computation of the dst quantization info. + _min_u8 = 0; + _max_u8 = 255; + _min_s8 = -128; + _max_s8 = 127; + } + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo activation_info, bool mixed_layout = false) { @@ -64,7 +98,20 @@ public: _mixed_layout = mixed_layout; _data_type = data_type; _bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; - _quantization_info = quantization_info; + + // Note : Quantization Info parameter from setup function is only used when quant datatype and activation function is not enabled or is identity. + if(is_data_type_quantized(data_type) && (!activation_info.enabled() || activation_info.activation() == ActivationFunction::IDENTITY)) + { + // Initialises quantization info with appropriate scale and offset for given input shapes. + setup_quantization(weights_shape, output_shape,_input_q_info, _weight_q_info, data_type); + } + else + { + _input_q_info = quantization_info; + _weight_q_info = quantization_info; + _dst_q_info = quantization_info; + } + _activation_info = activation_info; _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, transpose_weights, reshape_weights); @@ -92,17 +139,17 @@ protected: { if(_data_type == DataType::QASYMM8) { - std::uniform_int_distribution<uint8_t> distribution(0, 30); + std::uniform_int_distribution<uint32_t> distribution(_min_u8, _max_u8); library->fill(tensor, distribution, i); } else if(_data_type == DataType::QASYMM8_SIGNED) { - std::uniform_int_distribution<int8_t> distribution(-15, 15); + std::uniform_int_distribution<int32_t> distribution(_min_s8, _max_s8); library->fill(tensor, distribution, i); } else if(_data_type == DataType::S32) { - std::uniform_int_distribution<int32_t> distribution(-50, 50); + std::uniform_int_distribution<int32_t> distribution(_min_bias, _max_bias); library->fill(tensor, distribution, i); } else if(_data_type == DataType::F16) @@ -144,10 +191,10 @@ protected: } // Create tensors - TensorType src = create_tensor<TensorType>(input_shape, _data_type, 1, _quantization_info); - TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, _data_type, 1, _quantization_info); - TensorType bias = create_tensor<TensorType>(bias_shape, _bias_data_type, 1, _quantization_info); - TensorType dst = create_tensor<TensorType>(output_shape, _data_type, 1, _quantization_info); + TensorType src = create_tensor<TensorType>(input_shape, _data_type, 1, _input_q_info); + TensorType weights = create_tensor<TensorType>(reshaped_weights_shape, _data_type, 1, _weight_q_info); + TensorType bias = create_tensor<TensorType>(bias_shape, _bias_data_type, 1); + TensorType dst = create_tensor<TensorType>(output_shape, _data_type, 1, _dst_q_info); // Create Fully Connected layer info FullyConnectedLayerInfo fc_info; @@ -178,8 +225,8 @@ protected: ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); // Fill tensors - fill(AccessorType(src), 0); - fill(AccessorType(bias), 2); + fill(AccessorType(src), 0 + _hash); + fill(AccessorType(bias), 2 + _hash); if(!reshape_weights || !transpose_weights) { @@ -187,7 +234,7 @@ protected: RawTensor tmp(tmp_shape, _data_type, 1); // Fill with original shape - fill(tmp, 1); + fill(tmp, 1 + _hash); // Transpose elementwise tmp = transpose(tmp); @@ -204,7 +251,7 @@ protected: } else { - fill(AccessorType(weights), 1); + fill(AccessorType(weights), 1 + _hash); } if(_mixed_layout) @@ -223,16 +270,16 @@ protected: SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape) { // Create reference - SimpleTensor<T> src{ input_shape, _data_type, 1, _quantization_info }; - SimpleTensor<T> weights{ weights_shape, _data_type, 1, _quantization_info }; - SimpleTensor<TBias> bias{ bias_shape, _bias_data_type, 1, _quantization_info }; + SimpleTensor<T> src{ input_shape, _data_type, 1, _input_q_info }; + SimpleTensor<T> weights{ weights_shape, _data_type, 1, _weight_q_info }; + SimpleTensor<TBias> bias{ bias_shape, _bias_data_type, 1, QuantizationInfo() }; // Fill reference - fill(src, 0); - fill(weights, 1); - fill(bias, 2); + fill(src, 0 + _hash); + fill(weights, 1 + _hash); + fill(bias, 2 + _hash); - return reference::activation_layer(reference::fully_connected_layer<T>(src, weights, bias, output_shape), _activation_info, _quantization_info); + return reference::activation_layer(reference::fully_connected_layer<T>(src, weights, bias, output_shape, _dst_q_info), _activation_info, _dst_q_info); } TensorType _target{}; @@ -240,15 +287,28 @@ protected: DataType _data_type{}; DataType _bias_data_type{}; bool _mixed_layout{ false }; - QuantizationInfo _quantization_info{}; + QuantizationInfo _input_q_info{}; + QuantizationInfo _weight_q_info{}; + QuantizationInfo _dst_q_info{}; ActivationLayerInfo _activation_info{}; + + // Random initialization limits + // Default values are previously handcrafted limits + // that sould be used when we don't use dynamic quantization + int32_t _min_bias{-50}; + int32_t _max_bias{50}; + + int32_t _min_u8{0}; + int32_t _max_u8{30}; + int32_t _min_s8{-15}; + int32_t _max_s8{15}; + int _hash{0}; }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> class FullyConnectedLayerValidationFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type, ActivationLayerInfo activation_info) { @@ -262,7 +322,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class FullyConnectedLayerValidationQuantizedFixture : public FullyConnectedLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, bool transpose_weights, bool reshape_weights, DataType data_type, QuantizationInfo quantization_info, ActivationLayerInfo activation_info) { @@ -273,7 +332,7 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class FullyConnectedWithDynamicWeightsFixture : public framework::Fixture +class FullyConnectedWithDynamicTensorsFixture : public framework::Fixture { private: template <typename U> @@ -289,6 +348,21 @@ private: std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); library->fill(tensor, distribution, i); } + else if(_data_type == DataType::QASYMM8) + { + std::uniform_int_distribution<uint32_t> distribution(_min_u8, _max_u8); + library->fill(tensor, distribution, i); + } + else if(_data_type == DataType::QASYMM8_SIGNED) + { + std::uniform_int_distribution<int32_t> distribution(_min_s8, _max_s8); + library->fill(tensor, distribution, i); + } + else if(_data_type == DataType::S32) + { + std::uniform_int_distribution<int32_t> distribution(_min_bias, _max_bias); + library->fill(tensor, distribution, i); + } else { library->fill_tensor_uniform(tensor, i); @@ -316,82 +390,198 @@ private: } } - void validate_with_tolerance(TensorType &target, SimpleTensor<T> &ref) + void validate_with_tolerance(TensorType &target, SimpleTensor<float> &ref) { - if(_data_type == DataType::F32) - { - constexpr RelativeTolerance<float> rel_tolerance_f32(0.05f); - constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.0001f); - validate(AccessorType(target), ref, rel_tolerance_f32, 0, abs_tolerance_f32); - } - else - { - validate(AccessorType(target), ref); - } + constexpr RelativeTolerance<float> rel_tolerance_f32(0.01f); + constexpr AbsoluteTolerance<float> abs_tolerance_f32(0.001f); + validate(AccessorType(target), ref, rel_tolerance_f32, 0, abs_tolerance_f32); + } + + void validate_with_tolerance(TensorType &target, SimpleTensor<half_float::half> &ref) + { + constexpr AbsoluteTolerance<float> abs_tolerance_f16(0.3f); + const RelativeTolerance<half_float::half> rel_tolerance_f16(half_float::half(0.2f)); + constexpr float tolerance_num_f16 = 0.07f; + + validate(AccessorType(target), ref, rel_tolerance_f16, tolerance_num_f16, abs_tolerance_f16); + } + + void validate_with_tolerance(TensorType &target, SimpleTensor<uint8_t> &ref) + { + constexpr AbsoluteTolerance<uint32_t> tolerance_qasymm8(1); + validate(AccessorType(target), ref, tolerance_qasymm8); + } + + void validate_with_tolerance(TensorType &target, SimpleTensor<int8_t> &ref) + { + constexpr AbsoluteTolerance<int32_t> tolerance_qasymm8_signed(1); + validate(AccessorType(target), ref, tolerance_qasymm8_signed); + } + + void setup_quantization(TensorShape weights_shape, TensorShape output_shape, QuantizationInfo &input_q_info, QuantizationInfo &weights_q_info, DataType data_type) + { + _hash = weights_shape[0] + weights_shape[1] + output_shape[0] + output_shape[1]; + + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + _hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + input_q_info = QuantizationInfo(scale_lhs, offset_lhs); + weights_q_info = QuantizationInfo(scale_rhs, offset_rhs); + + const int k = weights_shape.x(); + QuantizationHint q_hint = suggest_mac_dst_q_info_and_bias(input_q_info, weights_q_info, k, data_type, 0.1f /* bias_fraction */, 4 /* number of standard deviations*/); + + _dst_q_info = q_hint.q_info; + _min_bias = q_hint.bias_min; + _max_bias = q_hint.bias_max; + + // Do not change here as these limits are the natural limits of the associated data types and + // are embedded in the computation of the dst quantization info. + _min_u8 = 0; + _max_u8 = 255; + _min_s8 = -128; + _max_s8 = 127; } public: - template <typename...> + using TDecay = typename std::decay<T>::type; + using TBias = typename std::conditional < (std::is_same<TDecay, uint8_t>::value || std::is_same<TDecay, int8_t>::value), int32_t, T >::type; + void setup(TensorShape src_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape dst_shape, - DataType data_type, ActivationLayerInfo activation_info) + DataType data_type, ActivationLayerInfo activation_info, bool constant_weights, bool constant_bias, bool weights_reshaped, bool remove_bias = false) { _data_type = data_type; - // Setup tensor meta-data - TensorInfo src_info(src_shape, 1, data_type); - _src.allocator()->init(src_info); + const bool is_quantized = is_data_type_quantized(data_type); + const DataType bias_data_type = (is_quantized) ? DataType::S32 : data_type; - TensorShape tr_weights_shape{ weights_shape[1], weights_shape[0] }; - TensorInfo wei_info(tr_weights_shape, 1, data_type); - _weights.allocator()->init(wei_info); + if (is_quantized && (!activation_info.enabled() || activation_info.activation() == ActivationFunction::IDENTITY)) + { + setup_quantization(weights_shape, dst_shape, _src_q_info, _weights_q_info, data_type); + } + else + { + _src_q_info = QuantizationInfo(0.1f, 10); + _dst_q_info = QuantizationInfo(0.3f, 20); + _weights_q_info = QuantizationInfo(0.2f, 5); + } - TensorInfo bias_info(bias_shape, 1, data_type); - _bias.allocator()->init(bias_info); + // Configure TensorInfo Objects + const TensorInfo src_info(src_shape, 1, data_type, _src_q_info); + const TensorInfo dst_info(dst_shape, 1, data_type, _dst_q_info); + TensorInfo bias_info(bias_shape, 1, bias_data_type); + TensorInfo wei_info(weights_shape, 1, data_type, _weights_q_info); - TensorInfo dst_info(dst_shape, 1, data_type); + if(!constant_weights && weights_reshaped) + { + const TensorShape tr_weights_shape{ weights_shape[1], weights_shape[0] }; + wei_info.set_tensor_shape(tr_weights_shape); + } + wei_info.set_are_values_constant(constant_weights); + bias_info.set_are_values_constant(constant_bias); + + // Initialise Tensors + _src.allocator()->init(src_info); + _weights.allocator()->init(wei_info); + if(!remove_bias) + _bias.allocator()->init(bias_info); _dst.allocator()->init(dst_info); // Configure FC layer and mark the weights as non constant FullyConnectedLayerInfo fc_info; - fc_info.activation_info = activation_info; - fc_info.are_weights_reshaped = true; - fc_info.transpose_weights = false; - fc_info.constant_weights = false; + fc_info.activation_info = activation_info; + if(!constant_weights) + { + fc_info.are_weights_reshaped = weights_reshaped; + fc_info.transpose_weights = !weights_reshaped; + } FunctionType fc; - fc.configure(&_src, &_weights, &_bias, &_dst, fc_info); + fc.configure(&_src, &_weights, (remove_bias) ? nullptr : &_bias, &_dst, fc_info); // Allocate all the tensors _src.allocator()->allocate(); _weights.allocator()->allocate(); - _bias.allocator()->allocate(); + if(!remove_bias) + _bias.allocator()->allocate(); _dst.allocator()->allocate(); // Run multiple iterations with different inputs constexpr int num_iterations = 5; int randomizer_offset = 0; + + // Create reference tensors + SimpleTensor<T> src{ src_shape, data_type, 1, _src_q_info }; + SimpleTensor<T> weights{ weights_shape, data_type, 1, _weights_q_info }; + SimpleTensor<TBias> bias{ bias_shape, bias_data_type }; + + // Fill weights and/or bias if they remain constant + if(constant_weights) + { + fill(AccessorType(_weights), 1 + _hash); + fill(weights, 1 + _hash); + } + if(constant_bias && !remove_bias) + { + fill(AccessorType(_bias), 2 + _hash); + fill(bias, 2 + _hash); + } + // To remove bias, fill with 0 + if(remove_bias && is_quantized) + { + library->fill_tensor_value(bias, 0); + } + else if(remove_bias) + { + library->fill_tensor_value(bias, (float)0.0); + } + for(int i = 0; i < num_iterations; ++i) { // Run target { fill(AccessorType(_src), randomizer_offset); - fill_transposed_weights(_weights, weights_shape, randomizer_offset + 1); - fill(AccessorType(_bias), randomizer_offset + 2); + if(!constant_weights) + { + if(weights_reshaped) + { + fill_transposed_weights(_weights, weights_shape, randomizer_offset + 1 + _hash); + } + else + { + fill(AccessorType(_weights), randomizer_offset + 1 +_hash); + } + } + if(!constant_bias && !remove_bias) + { + fill(AccessorType(_bias), randomizer_offset + 2 + _hash); + } fc.run(); } // Run reference and compare { - SimpleTensor<T> src{ src_shape, data_type }; - SimpleTensor<T> weights{ weights_shape, data_type }; - SimpleTensor<T> bias{ bias_shape, data_type }; - // Fill reference fill(src, randomizer_offset); - fill(weights, randomizer_offset + 1); - fill(bias, randomizer_offset + 2); + if(!constant_weights) + { + fill(weights, randomizer_offset + 1 + _hash); + } + if(!constant_bias && !remove_bias) + { + fill(bias, randomizer_offset + 2 + _hash); + } - auto dst = reference::activation_layer(reference::fully_connected_layer<T>(src, weights, bias, dst_shape), activation_info); + auto dst = reference::activation_layer(reference::fully_connected_layer<T>(src, weights, bias, dst_shape, _dst_q_info), activation_info, _dst_q_info); // Validate validate_with_tolerance(_dst, dst); @@ -404,8 +594,60 @@ public: private: TensorType _src{}, _weights{}, _bias{}, _dst{}; DataType _data_type{ DataType::UNKNOWN }; + + QuantizationInfo _src_q_info{}; + QuantizationInfo _weights_q_info{}; + QuantizationInfo _dst_q_info{}; + + // Random initialization limits + // Default values are previously handcrafted limits + // that sould be used when we don't use dynamic quantization + int32_t _min_bias{-50}; + int32_t _max_bias{50}; + + int32_t _min_u8{0}; + int32_t _max_u8{30}; + int32_t _min_s8{-15}; + int32_t _max_s8{15}; + int _hash{0}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class FullyConnectedWithDynamicWeightsFixture : public FullyConnectedWithDynamicTensorsFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape src_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape dst_shape, + DataType data_type, ActivationLayerInfo activation_info, bool weights_reshaped) + { + FullyConnectedWithDynamicTensorsFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, weights_shape, bias_shape, + dst_shape, data_type, activation_info, false, true, weights_reshaped, false); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class FullyConnectedDynamicNoBiasFixture : public FullyConnectedWithDynamicTensorsFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape src_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape dst_shape, + DataType data_type, ActivationLayerInfo activation_info, bool weights_reshaped) + { + FullyConnectedWithDynamicTensorsFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, weights_shape, bias_shape, + dst_shape, data_type, activation_info, false, true, weights_reshaped, true); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class FullyConnectedWithDynamicBiasFixture : public FullyConnectedWithDynamicTensorsFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape src_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape dst_shape, + DataType data_type, ActivationLayerInfo activation_info) + { + FullyConnectedWithDynamicTensorsFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, weights_shape, bias_shape, + dst_shape, data_type, activation_info, true, false, false, false); + } }; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_FULLY_CONNECTED_LAYER_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_FULLYCONNECTEDLAYERFIXTURE_H diff --git a/tests/validation/fixtures/FuseBatchNormalizationFixture.h b/tests/validation/fixtures/FuseBatchNormalizationFixture.h index 6fbabeee56..a05e4169a7 100644 --- a/tests/validation/fixtures/FuseBatchNormalizationFixture.h +++ b/tests/validation/fixtures/FuseBatchNormalizationFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, int class FuseBatchNormalizationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape_w, DataType data_type, DataLayout data_layout, bool in_place, bool with_bias, bool with_gamma, bool with_beta) { std::tie(_target_w, _target_b) = compute_target(shape_w, data_type, data_layout, in_place, with_bias, with_gamma, with_beta); diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h index c118da66ae..94bedc83e1 100644 --- a/tests/validation/fixtures/GEMMFixture.h +++ b/tests/validation/fixtures/GEMMFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_GEMM_FIXTURE -#define ARM_COMPUTE_TEST_GEMM_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_GEMMFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_GEMMFIXTURE_H #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/TensorShape.h" @@ -34,6 +34,7 @@ #include "tests/framework/Fixture.h" #include "tests/validation/Helpers.h" #include "tests/validation/reference/ActivationLayer.h" +#include "tests/validation/reference/ElementwiseOperations.h" #include "tests/validation/reference/GEMM.h" #include <random> @@ -44,16 +45,15 @@ namespace test { namespace validation { -template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false> -class GEMMValidationFixture : public framework::Fixture +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool pretranspose_a = false, bool pretranspose_b = false, bool run_twice = false> +class GEMMGenericValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape output_shape, float alpha, float beta, bool pretranspose, DataType data_type) + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape output_shape, float alpha, float beta, bool pretranspose, DataType data_type, bool accumulate=false) { ARM_COMPUTE_UNUSED(pretranspose); - _target = compute_target(shape_a, shape_b, shape_c, output_shape, alpha, beta, data_type); - _reference = compute_reference(shape_a, shape_b, output_shape, alpha, beta, data_type); + _target = compute_target(shape_a, shape_b, shape_c, output_shape, alpha, beta, data_type, accumulate); + _reference = compute_reference(shape_a, shape_b, output_shape, alpha, beta, data_type, accumulate); } protected: @@ -80,7 +80,7 @@ protected: } TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, const TensorShape &output_shape, float alpha, float beta, - DataType data_type) + DataType data_type, bool accumulate=false) { // Create tensors TensorType a = create_tensor<TensorType>(shape_a, data_type, 1); @@ -98,8 +98,8 @@ protected: (disable_c) ? nullptr : &c, &dst, alpha, beta, - GEMMInfo(false, false, false, (reinterpret_output_as_3d ? output_shape[2] : 0), reinterpret_input_as_3d, false, GEMMLowpOutputStageInfo(), false, (reinterpret_input_as_3d - || reinterpret_output_as_3d))); + GEMMInfo(false, false, false, (reinterpret_output_as_3d ? output_shape[2] : 0), reinterpret_input_as_3d, false, GEMMLowpOutputStageInfo(), false, false, (reinterpret_input_as_3d + || reinterpret_output_as_3d), arm_compute::ActivationLayerInfo(), false /* fixed_format */, arm_compute::WeightFormat::UNSPECIFIED, false /* pretranspose_B */, accumulate)); ARM_COMPUTE_ASSERT(a.info()->is_resizable()); ARM_COMPUTE_ASSERT(b.info()->is_resizable()); ARM_COMPUTE_ASSERT(c.info()->is_resizable()); @@ -121,10 +121,25 @@ protected: // Fill tensors fill(AccessorType(a), 0); fill(AccessorType(b), 1); + if (accumulate) + { + fill(AccessorType(dst), 6); + } if(!disable_c) { fill(AccessorType(c), 2); } + // Run with variable inputs. + if(run_twice) + { + gemm.run(); + fill(AccessorType(a), 3); // Fill tensors with new seed after run + fill(AccessorType(b), 4); + if(!disable_c) + { + fill(AccessorType(c), 5); + } + } // Compute GEMM function gemm.run(); @@ -133,10 +148,9 @@ protected: } SimpleTensor<T> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, float alpha, float beta, - DataType data_type) + DataType data_type, bool accumulate=false) { TensorShape shape_a_to_use = shape_a; - if(reinterpret_input_as_3d) { // Collapse the second and third dimension if the input is 3D @@ -147,6 +161,7 @@ protected: SimpleTensor<T> a{ shape_a_to_use, data_type, 1 }; SimpleTensor<T> b{ shape_b, data_type, 1 }; SimpleTensor<T> c{ output_shape, data_type, 1 }; + SimpleTensor<T> dst{ output_shape, data_type, 1 }; // Fill reference fill(a, 0); @@ -159,27 +174,96 @@ protected: const int m = reinterpret_output_as_3d ? output_shape[1] * output_shape[2] : output_shape[1]; const int batch_size = reinterpret_output_as_3d ? output_shape[3] : output_shape[2]; - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(c.data() + i * n, c.data(), n * sizeof(T)); } } + /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if pretranspose_A is set to true, then A is assumed to be (B x K x M), + therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K) + in order to be able to call reference implementation that works with (B x M x K) input. + Similarly, if pretranspose_B is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */ + + // Define transposed shapes + TensorShape a_transposed_shape(a.shape().y(), a.shape().x()); + TensorShape b_transposed_shape(b.shape().y(), b.shape().x()); + + // Define transposed tensors + SimpleTensor<T> a_transposed{ a_transposed_shape, data_type }; + SimpleTensor<T> b_transposed{ b_transposed_shape, data_type }; + + // pretranspose a if necessary + if(pretranspose_a) + { + transpose_matrix<T>(a, a_transposed); + } + + // pretranspose b if necessary + if(pretranspose_b) + { + transpose_matrix<T>(b, b_transposed); + } + + // Run with variable inputs. + if(run_twice) + { + reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta); + fill((pretranspose_a) ? a_transposed : a, 3); + fill((pretranspose_b) ? b_transposed : b, 4); + fill(c, 5); + } + + // Do in place summation + if (accumulate) + { + fill(dst, 6); + } + // Setting beta to 0 will effectively disable C for the // computation of the reference: alpha * A * B + 0 * C - return reference::gemm<T>(a, b, c, alpha, disable_c ? 0.f : beta); + // Use transposed tensors if boolean enabled else use original tensors + if (accumulate) + { + reference::gemm_accumulate<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta, dst); + return dst; + } + else + { + return reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, alpha, disable_c ? 0.f : beta); + } } TensorType _target{}; SimpleTensor<T> _reference{}; }; +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool pretranspose_a = false, bool pretranspose_b = false, bool run_twice = false> +class GEMMValidationFixture : protected GEMMGenericValidationFixture<TensorType, AccessorType, FunctionType, T, disable_c, reinterpret_input_as_3d, reinterpret_output_as_3d, pretranspose_a, pretranspose_b, run_twice> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape output_shape, float alpha, float beta, bool pretranspose, DataType data_type) + { + GEMMGenericValidationFixture<TensorType, AccessorType, FunctionType, T, disable_c, reinterpret_input_as_3d, reinterpret_output_as_3d, pretranspose_a, pretranspose_b, run_twice>::setup(shape_a, shape_b, shape_c, output_shape, alpha, beta, pretranspose, data_type, false /*accumulate*/); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool disable_c = false, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool pretranspose_a = false, bool pretranspose_b = false, bool run_twice = false> +class GEMMAccumulateValidationFixture : protected GEMMGenericValidationFixture<TensorType, AccessorType, FunctionType, T, disable_c, reinterpret_input_as_3d, reinterpret_output_as_3d, pretranspose_a, pretranspose_b, run_twice> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape output_shape, float alpha, float beta, bool pretranspose, DataType data_type) + { + bool accumulate = true; + GEMMGenericValidationFixture<TensorType, AccessorType, FunctionType, T, disable_c, reinterpret_input_as_3d, reinterpret_output_as_3d, pretranspose_a, pretranspose_b, run_twice>::setup(shape_a, shape_b, shape_c, output_shape, alpha, beta, pretranspose, data_type, accumulate); + } +}; + template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> class GEMMMatrixMultiplyValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) { @@ -255,8 +339,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -285,7 +368,7 @@ protected: if(broadcast_bias) { - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -303,7 +386,6 @@ template <typename TensorType, typename AccessorType, typename T, typename GEMMO class GEMMMatrixMultiply3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) { @@ -378,8 +460,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -408,7 +489,7 @@ protected: fill(rhs, 1); fill(bias, 2); - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -425,7 +506,6 @@ template <typename TensorType, typename AccessorType, typename T, typename Resha class GEMMMatrixMultiplyInterleavedTransposedValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) { @@ -533,8 +613,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -563,7 +642,7 @@ protected: if(broadcast_bias) { - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -581,7 +660,6 @@ template <typename TensorType, typename AccessorType, typename T, typename Resha class GEMMMatrixMultiplyInterleavedTransposed3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, float alpha, float beta, unsigned int v0, unsigned int h0, bool broadcast_bias, bool fp16_mixed_precision, const ActivationLayerInfo &act_info, DataType data_type, GPUTarget gpu_arch) { @@ -688,8 +766,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -718,7 +795,7 @@ protected: fill(rhs, 1); fill(bias, 2); - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -735,7 +812,6 @@ template <typename TensorType, typename AccessorType, typename T, typename Resha class GEMMMatrixMultiplyReshapedValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs, bool interleave_rhs, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool broadcast_bias, bool lhs_transpose, const ActivationLayerInfo &act_info) { @@ -863,8 +939,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -893,7 +968,7 @@ protected: if(broadcast_bias) { - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -919,7 +994,6 @@ template <typename TensorType, typename AccessorType, typename T, typename Resha class GEMMMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs, bool interleave_rhs, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool lhs_transpose, const ActivationLayerInfo &act_info) { @@ -1044,8 +1118,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -1074,7 +1147,7 @@ protected: fill(rhs, 1); fill(bias, 2); - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -1099,7 +1172,6 @@ template <typename TensorType, typename AccessorType, typename T, typename Resha class GEMMMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, bool export_to_cl_image, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) { @@ -1216,8 +1288,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -1246,7 +1317,7 @@ protected: if(broadcast_bias) { - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -1265,7 +1336,6 @@ template <typename TensorType, typename AccessorType, typename T, typename Resha class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, bool export_to_cl_image, bool has_pad_y, DataType data_type, float alpha, float beta, const ActivationLayerInfo &act_info) { @@ -1387,8 +1457,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -1417,7 +1486,7 @@ protected: fill(rhs, 1); fill(bias, 2); - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -1435,7 +1504,6 @@ template <typename TensorType, typename AccessorType, typename T, typename GEMMO class GEMMMatrixMultiplyNativeValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) { @@ -1524,8 +1592,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -1554,7 +1621,7 @@ protected: if(broadcast_bias) { - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -1572,7 +1639,6 @@ template <typename TensorType, typename AccessorType, typename T, typename GEMMO class GEMMMatrixMultiplyNative3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, DataType data_type, float alpha, float beta, const ActivationLayerInfo &act_info) { @@ -1660,8 +1726,7 @@ protected: ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); + { ACL_DST, &dst } }); gemm.run(gemm_pack); return dst; @@ -1690,7 +1755,7 @@ protected: fill(rhs, 1); fill(bias, 2); - // In case of broadcast, we need simply copy the first into the following "M" ones + // In case of broadcast, we need to simply copy the first into the following "M" ones for(int i = 1; i < m * batch_size; i++) { memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); @@ -1703,7 +1768,170 @@ protected: SimpleTensor<T> _reference{}; }; +template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> +class GEMMMatrixMultiplyReshapedOnlyRhsMMULValidationFixture : public framework::Fixture +{ +public: + void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, bool export_to_cl_image, DataType data_type, float alpha, + float beta, bool broadcast_bias, + const ActivationLayerInfo &act_info) + { + GEMMLHSMatrixInfo lhs_info; + lhs_info.m0 = m0; + lhs_info.k0 = k0; + + GEMMRHSMatrixInfo rhs_info; + rhs_info.n0 = n0; + rhs_info.k0 = k0; + rhs_info.interleave = true; + rhs_info.transpose = false; + rhs_info.h0 = 4; + rhs_info.export_to_cl_image = export_to_cl_image; + + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + const TensorShape bias_shape(n, + broadcast_bias ? 1 : m, + broadcast_bias ? 1 : batch_size); + + _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info); + _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); + using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; + + DistributionType distribution{ T(-1.0f), T(1.0f) }; + library->fill(tensor, distribution, i); + + // Fill border with infinity in order to check the presence of NaN values (i.e. inf * 0) + DistributionType distribution_inf{ T(std::numeric_limits<float>::infinity()), T(std::numeric_limits<float>::infinity()) }; + library->fill_borders_with_garbage(tensor, distribution_inf, i); + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, + DataType data_type, float alpha, float beta, bool broadcast_bias, const ActivationLayerInfo &act_info) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); + TensorType bias = create_tensor<TensorType>(bias_shape, data_type, 1); + TensorType rhs_reshaped; + TensorType dst; + + const unsigned int M = lhs_shape[1]; + const unsigned int N = rhs_shape[0]; + const unsigned int K = lhs_shape[0]; + GEMMKernelInfo kernel_info; + kernel_info.m = M; + kernel_info.n = N; + kernel_info.k = K; + kernel_info.depth_output_gemm3d = 0; + kernel_info.reinterpret_input_as_3d = false; + kernel_info.broadcast_bias = broadcast_bias; + kernel_info.activation_info = act_info; + + // Create and configure function + ReshapeRHSOperatorType reshape_rhs; + GEMMOperatorType gemm; + + validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); + if(!validate_result) + { + return nullptr; + } + + reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); + + validate_result = bool(gemm.validate(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info)); + if(!validate_result) + { + return nullptr; + } + + gemm.configure(lhs.info(), rhs_reshaped.info(), bias.info(), dst.info(), alpha, beta, lhs_info, rhs_info, kernel_info); + + ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + rhs_reshaped.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + fill(AccessorType(bias), 2); + + // Compute GEMM + ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; + reshape_rhs.run(reshape_rhs_pack); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, + { ACL_SRC_1, &rhs_reshaped }, + { ACL_SRC_2, &bias }, + { ACL_DST, &dst } }); + gemm.run(gemm_pack); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type, float alpha, float beta, bool broadcast_bias, + const ActivationLayerInfo &act_info) + { + if(!validate_result) + return SimpleTensor<T>(); + + TensorShape dst_shape = lhs_shape; + dst_shape[0] = rhs_shape[0]; + dst_shape[1] = lhs_shape[1]; + + // Create reference + SimpleTensor<T> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<T> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<T> bias{ dst_shape, data_type, 1 }; + + const int n = rhs_shape[0]; + const int m = lhs_shape[1]; + const int batch_size = lhs_shape[2]; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + fill(bias, 2); + + if(broadcast_bias) + { + // In case of broadcast, we need to simply copy the first into the following "M" ones + for(int i = 1; i < m * batch_size; i++) + { + memcpy(bias.data() + i * n, bias.data(), n * sizeof(T)); + } + } + + return reference::activation_layer(reference::gemm<T>(lhs, rhs, bias, alpha, beta), act_info); + } + + bool validate_result = true; + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_GEMM_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_GEMMFIXTURE_H diff --git a/tests/validation/fixtures/GEMMInterleave4x4Fixture.h b/tests/validation/fixtures/GEMMInterleave4x4Fixture.h index 44dc0dddbc..59fc460869 100644 --- a/tests/validation/fixtures/GEMMInterleave4x4Fixture.h +++ b/tests/validation/fixtures/GEMMInterleave4x4Fixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class GEMMInterleave4x4ValidationFixture : public framework::Fixture { public: - template <typename...> void setup(size_t x, size_t y, DataType data_type) { _data_type = data_type; @@ -88,7 +87,7 @@ protected: // Create and configure function FunctionType f; - f.configure(&a, &b); + f.configure(a.info(), b.info()); ARM_COMPUTE_ASSERT(a.info()->is_resizable()); ARM_COMPUTE_ASSERT(b.info()->is_resizable()); @@ -104,8 +103,9 @@ protected: fill(AccessorType(a), 0); fill(AccessorType(b), 0); - // Compute GEMM function - f.run(); + // Compute GEMM interleave kernel + ITensorPack tensors{ { ACL_SRC, &a }, { ACL_DST, &b } }; + f.run(tensors); return b; } diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h index 5cf210bab4..aa4eedb75d 100644 --- a/tests/validation/fixtures/GEMMLowpFixture.h +++ b/tests/validation/fixtures/GEMMLowpFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,22 +21,20 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE -#define ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_GEMMLOWPFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_GEMMLOWPFIXTURE_H -#include "arm_compute/core/KernelDescriptors.h" -#include "arm_compute/core/TensorShape.h" -#include "arm_compute/core/Types.h" #include "arm_compute/core/utils/quantization/AsymmHelpers.h" -#include "tests/AssetsLibrary.h" -#include "tests/Globals.h" -#include "tests/IAccessor.h" -#include "tests/framework/Asserts.h" -#include "tests/framework/Fixture.h" +#include "src/core/utils/quantization/AsymmHelpers.h" #include "tests/validation/Helpers.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/Validation.h" #include "tests/validation/reference/GEMMLowp.h" +#include "tests/validation/reference/ArithmeticOperations.h" +#include "tests/validation/reference/DequantizationLayer.h" -#include <random> +#include <cstdint> +#include <vector> namespace arm_compute { @@ -49,84 +47,88 @@ namespace template <typename U> void fill(U &&tensor, int i) { - switch(tensor.data_type()) - { - case DataType::QSYMM8_PER_CHANNEL: - { - int min_bound = 128; - int max_bound = -127; - for(size_t j = 0; j < tensor.quantization_info().scale().size(); j++) - { - std::pair<int, int> bounds = get_symm_quantized_per_channel_bounds(tensor.quantization_info(), -1.0f, 1.0f, i); - if(bounds.first < min_bound) - { - min_bound = bounds.first; - } - if(bounds.second > max_bound) - { - max_bound = bounds.second; - } - } - std::uniform_int_distribution<int8_t> distribution(min_bound, max_bound); - library->fill(tensor, distribution, i); - break; - } - case DataType::QASYMM8: - { - std::uniform_int_distribution<uint8_t> distribution(1, 254); - library->fill(tensor, distribution, i); - break; - } - case DataType::F16: - { - arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; - library->fill(tensor, distribution, i); - break; - } - case DataType::F32: - { - std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); - library->fill(tensor, distribution, i); - break; - } - default: - library->fill_tensor_uniform(tensor, i); - } + library->fill_tensor_uniform(tensor, i); } -template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d, bool reinterpret_output_as_3d, typename OutputType, bool is_fused = false> -TensorType compute_gemmlowp_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, - GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo(), DataType data_type_a = DataType::QASYMM8, DataType data_type_b = DataType::QASYMM8, - QuantizationInfo b_qinfo = QuantizationInfo()) +template <typename U> +void fill_quantized(U &&tensor, int i) { - // Create tensors - DataType data_type_output = output_stage.type == GEMMLowpOutputStageType::NONE ? DataType::S32 : data_type_a; - - TensorType a = create_tensor<TensorType>(shape_a, data_type_a, 1); - TensorType b = create_tensor<TensorType>(shape_b, data_type_b, 1); // gemm output before output stage mismatch if i pass data_layout_output here. to be investigated - TensorType output = create_tensor<TensorType>(shape_output, data_type_output, 1); - - a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset)); + ARM_COMPUTE_ASSERT(is_data_type_quantized(tensor.data_type())); + library->fill_tensor_uniform(tensor, i); +} - if(data_type_b == DataType::QSYMM8_PER_CHANNEL) +template <typename U> +void fill(U &&tensor, int i, int32_t min, int32_t max) +{ + if (tensor.data_type() == DataType::S32) { + std::uniform_int_distribution<int32_t> distribution(min, max); + library->fill(tensor, distribution, i); + } + else if(tensor.data_type() == DataType::F32) { - b.info()->set_quantization_info(b_qinfo); + std::uniform_real_distribution<float> distribution((float)min, (float)max); + library->fill(tensor, distribution, i); } else { - b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset)); + ARM_COMPUTE_ERROR("NOT SUPPORTED!"); + } +} + +/** Information about how to fill tensors */ +struct TensorFillInfo +{ + // Bias fill range. Default values are arbitrary + int32_t min_bias {-20000}; + int32_t max_bias {20000}; + + // Output fill range. Default values are arbitrary + int32_t min_output {-20000}; + int32_t max_output {20000}; + + // Optional extra hash to randomize tensor filling + int32_t hash {0}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d, bool reinterpret_output_as_3d, typename OutputType, bool is_fused = false, bool run_twice = false> +TensorType compute_gemmlowp_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, + const QuantizationInfo& output_qinfo, DataType data_type_a = DataType::QASYMM8, DataType data_type_b = DataType::QASYMM8, + GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo(), bool reshape_b_only_on_first_run = false, const TensorFillInfo& finfo = TensorFillInfo(), + bool accumulate = false, bool dynamic_qinfo = false, DataType data_type_output = DataType::UNKNOWN) +{ + ARM_COMPUTE_ASSERT(is_data_type_quantized_asymmetric(data_type_a)); + ARM_COMPUTE_ASSERT(data_type_a == data_type_b); + // If unknown, set to sensible defaults + if (data_type_output == DataType::UNKNOWN) { + data_type_output = output_stage.type == GEMMLowpOutputStageType::NONE ? DataType::S32 : data_type_a; } + + // Create tensors + TensorType a = create_tensor<TensorType>(shape_a, data_type_a, 1, dynamic_qinfo ? QuantizationInfo(1.0,0,true) : a_qinfo); + TensorType b = create_tensor<TensorType>(shape_b, data_type_b, 1, dynamic_qinfo ? QuantizationInfo(1.0,0,true) : b_qinfo); // gemm output before output stage mismatch if i pass data_layout_output here. to be investigated + TensorType output = create_tensor<TensorType>(shape_output, data_type_output, 1, output_qinfo /* output_qinfo will be ignored when output stage type is None */); + TensorType bias; if(is_fused) { TensorShape bias_shape(shape_b[0]); - bias = create_tensor<TensorType>(bias_shape, DataType::S32, 1); + bias = create_tensor<TensorType>(bias_shape,data_type_output == DataType::F32 ? DataType::F32 : DataType::S32, 1); } // Create and configure function // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output FunctionType gemmlowp; - gemmlowp.configure(&a, &b, is_fused ? &bias : nullptr, &output, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_output[2] : 0), reinterpret_input_as_3d, false, output_stage)); + gemmlowp.configure(&a, &b, is_fused ? &bias : nullptr, &output, GEMMInfo(false, false, reshape_b_only_on_first_run, (reinterpret_output_as_3d ? shape_output[2] : 0), reinterpret_input_as_3d, false, + output_stage, false /*fp_mixed_precision*/, false /*fast_math*/, false /*broadcast_bias*/, + arm_compute::ActivationLayerInfo(), false /* fixed_format */, arm_compute::WeightFormat::UNSPECIFIED, + false /* pretranspose_B */, accumulate)); + + // If the QuantizationInfo is dynamic, it needs to be settable after configure (note that we also force it to be dynamic) + if (dynamic_qinfo) + { + a.info()->set_quantization_info(QuantizationInfo(a_qinfo.scale(), a_qinfo.offset(), true)); + b.info()->set_quantization_info(QuantizationInfo(b_qinfo.scale(), b_qinfo.offset(), true)); + } ARM_COMPUTE_ASSERT(a.info()->is_resizable()); ARM_COMPUTE_ASSERT(b.info()->is_resizable()); @@ -144,25 +146,46 @@ TensorType compute_gemmlowp_target(const TensorShape &shape_a, const TensorShape ARM_COMPUTE_ASSERT(!output.info()->is_resizable()); // Fill tensors - fill(AccessorType(a), 0); - fill(AccessorType(b), 1); + fill_quantized(AccessorType(a), 0 + finfo.hash); + fill_quantized(AccessorType(b), 1 + finfo.hash); + + if (accumulate) + { + ARM_COMPUTE_ASSERT(accumulate != run_twice); + fill(AccessorType(output), 6 + finfo.hash, finfo.min_output, finfo.max_output); + } if(is_fused) { ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); bias.allocator()->allocate(); ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); - fill(AccessorType(bias), 2); + fill(AccessorType(bias), 2 + finfo.hash, finfo.min_bias, finfo.max_bias); } + + // Run with variable inputs. + if(run_twice) + { + gemmlowp.run(); + fill_quantized(AccessorType(a), 3 + finfo.hash); // Fill tensors with new seed after run + fill_quantized(AccessorType(b), 4 + finfo.hash); + if(is_fused) + { + fill(AccessorType(bias), 5 + finfo.hash, finfo.min_bias, finfo.max_bias); + } + } + // Compute GEMM function gemmlowp.run(); return output; } -template <bool reinterpret_input_as_3d, typename TI = uint8_t, typename TW = uint8_t> -SimpleTensor<int32_t> compute_gemmlowp_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, - DataType data_type_a = DataType::QASYMM8, DataType data_type_b = DataType::QASYMM8, QuantizationInfo b_qinfo = QuantizationInfo()) +template <bool reinterpret_input_as_3d, typename TI = uint8_t, typename TW = uint8_t, bool pretranspose_A = false, bool pretranspose_B = false, bool run_twice = false> +SimpleTensor<int32_t> compute_gemmlowp_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, + DataType data_type_a = DataType::QASYMM8, DataType data_type_b = DataType::QASYMM8, const TensorFillInfo& finfo = TensorFillInfo()) { + ARM_COMPUTE_ASSERT(is_data_type_quantized_asymmetric(data_type_a)); + ARM_COMPUTE_ASSERT(data_type_a == data_type_b); TensorShape shape_a_to_use = shape_a; if(reinterpret_input_as_3d) { @@ -171,101 +194,269 @@ SimpleTensor<int32_t> compute_gemmlowp_reference(const TensorShape &shape_a, con } // Create reference - SimpleTensor<TI> a{ shape_a_to_use, data_type_a, 1 }; - SimpleTensor<TW> b{ shape_b, data_type_b, 1, data_type_b == DataType::QSYMM8_PER_CHANNEL ? b_qinfo : QuantizationInfo(1.0f / 255, b_offset) }; + SimpleTensor<TI> a{ shape_a_to_use, data_type_a, 1, a_qinfo }; + SimpleTensor<TW> b{ shape_b, data_type_b, 1, b_qinfo }; + + TensorShape shape_a_to_use_transposed{ shape_a_to_use }; + TensorShape shape_b_transposed{ shape_b }; + + shape_a_to_use_transposed.set(0, shape_a_to_use[1]); + shape_a_to_use_transposed.set(1, shape_a_to_use[0]); + shape_b_transposed.set(0, shape_b[1]); + shape_b_transposed.set(1, shape_b[0]); + + SimpleTensor<TI> a_transposed{ shape_a_to_use_transposed, data_type_a, 1, a_qinfo }; + SimpleTensor<TW> b_transposed{ shape_b_transposed, data_type_b, 1, b_qinfo }; // Fill reference - fill(a, 0); - fill(b, 1); - return reference::gemmlowp_matrix_multiply_core<int32_t, TI, TW>(a, b, shape_output, a_offset, b_offset); -} + fill_quantized(a, 0 + finfo.hash); + fill_quantized(b, 1 + finfo.hash); + + // Transpose reference if required + /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if pretranspose_A is set to true, then A is assumed to be (B x K x M), + therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K) + in order to be able to call reference implementation that works with (B x M x K) input. + Similarly, if pretranspose_B is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */ + if(pretranspose_A) + { + transpose_matrix<TI>(a, a_transposed); + } + + if(pretranspose_B) + { + transpose_matrix<TW>(b, b_transposed); + } + + // Run with variable inputs. + const int32_t a_offset = a_qinfo.uniform().offset; + const int32_t b_offset = b_qinfo.uniform().offset; + + if(run_twice) + { + reference::gemmlowp_matrix_multiply_core<int32_t, TI, TW>((pretranspose_A ? a_transposed : a), (pretranspose_B ? b_transposed : b), shape_output, a_offset, b_offset); + fill_quantized((pretranspose_A) ? a_transposed : a, 3 + finfo.hash); + fill_quantized((pretranspose_B) ? b_transposed : b, 4 + finfo.hash); + } + + return reference::gemmlowp_matrix_multiply_core<int32_t, TI, TW>((pretranspose_A ? a_transposed : a), (pretranspose_B ? b_transposed : b), shape_output, a_offset, b_offset); } +} // namespace -template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false> -class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool run_twice = false> +class GEMMLowpGenericMatrixMultiplyCoreValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset) + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, bool accumulate=false, bool dynamic_qinfo = false) { - _target = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset); - _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset); + const auto a_qinfo = QuantizationInfo(1.0f / 255, a_offset); + const auto b_qinfo = QuantizationInfo(1.0f / 255, b_offset); + TensorFillInfo finfo; + _target = compute_target(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate, dynamic_qinfo); + _reference = compute_reference(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate); } protected: - TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const TensorFillInfo& finfo, const bool accumulate, const bool dynamic_qinfo) { - return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, int32_t>(shape_a, shape_b, shape_output, a_offset, b_offset); + const auto output_qinfo = QuantizationInfo(); // No output stage + return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, int32_t, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, output_qinfo, DataType::QASYMM8, DataType::QASYMM8, GEMMLowpOutputStageInfo(), false, finfo, accumulate, dynamic_qinfo); } - SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) + SimpleTensor<int32_t> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const TensorFillInfo& finfo, bool accumulate) { - return compute_gemmlowp_reference<reinterpret_input_as_3d>(shape_a, shape_b, shape_output, a_offset, b_offset); + SimpleTensor<int32_t> ref_output = compute_gemmlowp_reference<reinterpret_input_as_3d, uint8_t, uint8_t, false, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, + DataType::QASYMM8, DataType::QASYMM8, finfo); + + if (accumulate) + { + SimpleTensor<int32_t> output{ shape_output, DataType::S32, 1 }; + fill(output, 6 + finfo.hash, finfo.min_output, finfo.max_output); + reference::arithmetic_operation<int32_t>(reference::ArithmeticOperation::ADD, output, ref_output, output, ConvertPolicy::SATURATE); + return output; + } + + return ref_output; } TensorType _target{}; SimpleTensor<int32_t> _reference{}; }; -template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, typename TI = uint8_t, typename TW = uint8_t> -class GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture : public framework::Fixture +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool run_twice = false> +class GEMMLowpMatrixMultiplyCoreValidationFixture : protected GEMMLowpGenericMatrixMultiplyCoreValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, run_twice> { public: - template <typename...> - void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage, DataType data_type_b) + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset) { - ARM_COMPUTE_ASSERT(output_stage.type != GEMMLowpOutputStageType::NONE); - DataType data_type_a = data_type_b == DataType::QASYMM8_SIGNED ? DataType::QASYMM8_SIGNED : DataType::QASYMM8; + GEMMLowpGenericMatrixMultiplyCoreValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, run_twice>::setup(shape_a, shape_b, shape_output, a_offset, b_offset, false /* accumulate */); + } +}; - if(data_type_b == DataType::QSYMM8_PER_CHANNEL) - { - output_stage.is_quantized_per_channel = true; - const size_t num_channels = shape_b[0]; - std::vector<float> scales(num_channels); - std::uniform_real_distribution<float> distribution(0.f, 1.f); - library->fill(scales, distribution, 0); - output_stage.gemmlowp_multipliers.resize(num_channels); - output_stage.gemmlowp_shifts.resize(num_channels); - for(size_t i = 0; i < num_channels; ++i) - { - quantization::calculate_quantized_multiplier(scales[i], &output_stage.gemmlowp_multipliers[i], &output_stage.gemmlowp_shifts[i]); - } +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool run_twice = false> +class GEMMLowpMatrixMultiplyAccumulateValidationFixture : protected GEMMLowpGenericMatrixMultiplyCoreValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, run_twice> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset) + { + GEMMLowpGenericMatrixMultiplyCoreValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, run_twice>::setup(shape_a, shape_b, shape_output, a_offset, b_offset, true /* accumulate */); + } +}; - _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, 0, output_stage, data_type_a, data_type_b, QuantizationInfo(scales)); - _target = compute_target(shape_a, shape_b, shape_output, a_offset, 0, output_stage, data_type_a, data_type_b, QuantizationInfo(scales)); - } - else - { - _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage, data_type_a, data_type_b, QuantizationInfo()); - _target = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage, data_type_a, data_type_b, QuantizationInfo()); - } +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool run_twice = false> +class GEMMLowpMatrixMultiplyCoreDynamicQuantizationFixture : protected GEMMLowpGenericMatrixMultiplyCoreValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, run_twice> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset) + { + GEMMLowpGenericMatrixMultiplyCoreValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, run_twice>::setup(shape_a, shape_b, shape_output, a_offset, b_offset, false /* accumulate */, true /* dynamic_qinfo */); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, typename TI = uint8_t, typename TW = uint8_t, bool run_twice = false> +class GEMMLowpGenericMatrixMultiplyCoreFusedOffsetOutputValidationFixture : public framework::Fixture +{ +public: + /** Dynamically initialize the quantization info with saturation awareness + */ + template <typename T> + static void setup_quantization(DataType data_type, const TensorShape& shape_a, const TensorShape& shape_b, QuantizationInfo& a_qinfo, QuantizationInfo& b_qinfo, QuantizationInfo& output_qinfo, TensorFillInfo& finfo) + { + // This hash is used by random generators. There may be hash collisions but + // this is intentional as it's a very easy way to make the the current + // random generation process almost different for many test configurations, + // which were using the same set of values before. + finfo.hash = shape_a[0] + shape_a[1] + shape_b[0] + shape_b[1]; + + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + finfo.hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + a_qinfo = QuantizationInfo(scale_lhs, offset_lhs); + b_qinfo = QuantizationInfo(scale_rhs, offset_rhs); + + // reinterpret_input_as_3d or reinterpret_output_as_3d can be ignored, as the underlying gemm / matmul computation + // is equivalent to a standard 2D one with m-n-k dimensions + const int m = shape_a.y(); + const int n = shape_b.x(); + const int k = shape_a.x(); + + const float bias_fraction = 0.5f; // We enabled is_fused in compute_gemmlowp_target below, thus bias is included + + QuantizationHint q_hint = suggest_matmul_dst_q_info_and_bias(a_qinfo, b_qinfo, m, n, k, data_type, bias_fraction); + output_qinfo = q_hint.q_info; + finfo.min_bias = q_hint.bias_min; + finfo.max_bias = q_hint.bias_max; + + // Both target and reference implementations use negated offsets, i.e. + // float_val = (int_val + offset) * scale + // instead of + // float_val = (int_val - offset) * scale + // as usual. Therefore, after calculating the output quantization above, we + // negate the offsets of inputs' offsets. + a_qinfo = QuantizationInfo(scale_lhs, -offset_lhs); + b_qinfo = QuantizationInfo(scale_rhs, -offset_rhs); + } + + /** Initialize output stage info from quantization info */ + static Status init_gemmlowp_output_stage_info( + DataType data_type, + const QuantizationInfo& a_qinfo, + const QuantizationInfo& b_qinfo, + const QuantizationInfo& output_qinfo, + GEMMLowpOutputStageType type, + GEMMLowpOutputStageInfo &gemmlowp_output_stage_info) + { + ARM_COMPUTE_RETURN_ERROR_ON(!is_data_type_quantized_asymmetric(data_type)); + + const UniformQuantizationInfo aq_unif = a_qinfo.uniform(); + const UniformQuantizationInfo bq_unif = b_qinfo.uniform(); + const UniformQuantizationInfo oq_unif = output_qinfo.uniform(); + + float multiplier = (aq_unif.scale * bq_unif.scale) / oq_unif.scale; + int32_t int_multiplier; + int32_t shift; + + ARM_COMPUTE_RETURN_ON_ERROR( + quantization::calculate_quantized_multiplier(multiplier, &int_multiplier, &shift)); + + int32_t type_min = 0; + int32_t type_max = 0; + std::tie(type_min, type_max) = quantization::get_quantized_asymmetric_output_min_max(output_qinfo, ActivationLayerInfo(), data_type); + + gemmlowp_output_stage_info.gemmlowp_real_multiplier = multiplier; + gemmlowp_output_stage_info.gemmlowp_multiplier = int_multiplier; + gemmlowp_output_stage_info.gemmlowp_multipliers = { int_multiplier }; + gemmlowp_output_stage_info.gemmlowp_shift = shift; + gemmlowp_output_stage_info.gemmlowp_shifts = { shift }; + gemmlowp_output_stage_info.gemmlowp_offset = oq_unif.offset; + gemmlowp_output_stage_info.type = type; + gemmlowp_output_stage_info.gemmlowp_min_bound = type_min; + gemmlowp_output_stage_info.gemmlowp_max_bound = type_max; + + return Status{}; + } + + /** Currently this fixture only tests the following data type configurations: + * + * 1. a and b are of the same data type + * 2. The data type is quantized asymmetric + * + */ + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, GEMMLowpOutputStageType output_stage_type, DataType data_type, + bool reshape_b_only_on_first_run) + { + ARM_COMPUTE_ASSERT(output_stage_type != GEMMLowpOutputStageType::NONE); + ARM_COMPUTE_ASSERT(is_data_type_quantized_asymmetric(data_type)); + + // Randomized dynamic quantization: randomize quantization info in a way that ensures no result saturation + // most of the time + QuantizationInfo a_qinfo; + QuantizationInfo b_qinfo; + QuantizationInfo output_qinfo; + TensorFillInfo finfo; + setup_quantization<TI>(data_type, shape_a, shape_b, a_qinfo, b_qinfo, output_qinfo, finfo); + + GEMMLowpOutputStageInfo output_stage; + init_gemmlowp_output_stage_info(data_type, a_qinfo, b_qinfo, output_qinfo, output_stage_type, output_stage); + + _reference = compute_reference(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, data_type, data_type, output_stage, finfo); + _target = compute_target(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, output_qinfo, data_type, data_type, output_stage, reshape_b_only_on_first_run, finfo); } protected: - TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage, - DataType data_type_a, DataType data_type_b, QuantizationInfo b_qinfo) + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const QuantizationInfo& output_qinfo, + DataType data_type_a, DataType data_type_b, const GEMMLowpOutputStageInfo& output_stage, bool reshape_b_only_on_first_run = false, const TensorFillInfo& finfo = TensorFillInfo()) { - return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, qasymm8_t, true>(shape_a, shape_b, shape_output, a_offset, b_offset, - output_stage, data_type_a, data_type_b, b_qinfo); + return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, qasymm8_t, true, run_twice>(shape_a, shape_b, shape_output, a_qinfo, + b_qinfo, output_qinfo, data_type_a, data_type_b, output_stage, reshape_b_only_on_first_run, finfo); } - SimpleTensor<TI> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, - GEMMLowpOutputStageInfo output_stage, DataType data_type_a, DataType data_type_b, QuantizationInfo b_qinfo) + SimpleTensor<TI> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, + DataType data_type_a, DataType data_type_b, const GEMMLowpOutputStageInfo& output_stage, const TensorFillInfo& finfo = TensorFillInfo()) { - SimpleTensor<int32_t> output = compute_gemmlowp_reference<reinterpret_input_as_3d, TI, TW>(shape_a, shape_b, shape_output, a_offset, b_offset, data_type_a, data_type_b, b_qinfo); + SimpleTensor<int32_t> output = compute_gemmlowp_reference<reinterpret_input_as_3d, TI, TW, false, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, data_type_a, data_type_b, finfo); TensorShape bias_shape(shape_b[0]); SimpleTensor<int32_t> bias{ bias_shape, DataType::S32, 1 }; - fill(bias, 2); + (run_twice) ? fill(bias, 5 + finfo.hash, finfo.min_bias, finfo.max_bias) : fill(bias, 2 + finfo.hash, finfo.min_bias, finfo.max_bias); // Fill bias with same seed as last run of gemmlowp_target switch(output_stage.type) { case GEMMLowpOutputStageType::QUANTIZE_DOWN: - return reference::gemmlowp_quantize_down_scale<int32_t, TW>(output, bias, + return reference::gemmlowp_quantize_down_scale<int32_t, TI>(output, bias, output_stage.gemmlowp_offset, output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); break; case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT: - return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, TW>(output, bias, + return reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, TI>(output, bias, output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); break; default: @@ -277,11 +468,78 @@ protected: SimpleTensor<TI> _reference{}; }; +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, bool run_twice = false> +class GEMMLowpDequantizedMatrixMultiplyValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, bool accumulate) + { + const bool dynamic_qinfo = false; + const auto a_qinfo = QuantizationInfo(1.0f / 255, a_offset); + const auto b_qinfo = QuantizationInfo(5.0f / 255, b_offset); + TensorFillInfo finfo; + _target = compute_target(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate, dynamic_qinfo); + _reference = compute_reference(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, finfo, accumulate, dynamic_qinfo); + } + +protected: + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const TensorFillInfo& finfo, const bool accumulate, const bool dynamic_qinfo) + { + const auto output_qinfo = QuantizationInfo(); + return compute_gemmlowp_target<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, int32_t, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, output_qinfo, DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, GEMMLowpOutputStageInfo(), false, finfo, accumulate, dynamic_qinfo, DataType::F32); + } + + SimpleTensor<float> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, const QuantizationInfo& a_qinfo, const QuantizationInfo& b_qinfo, const TensorFillInfo& finfo, bool accumulate, const bool dynamic_qinfo) + { + QuantizationInfo s32_ref_output_quant_info = QuantizationInfo(a_qinfo.uniform().scale * b_qinfo.uniform().scale, 0, dynamic_qinfo); + + SimpleTensor<int32_t> s32_ref_output = compute_gemmlowp_reference<reinterpret_input_as_3d, int8_t, int8_t, false, false, run_twice>(shape_a, shape_b, shape_output, a_qinfo, b_qinfo, + DataType::QASYMM8_SIGNED, DataType::QASYMM8_SIGNED, finfo); + s32_ref_output.quantization_info(s32_ref_output_quant_info); + + SimpleTensor<float> f32_ref_output(s32_ref_output.shape(), DataType::F32); + f32_ref_output = reference::dequantization_layer<float, int32_t>(s32_ref_output); + + if (accumulate) + { + SimpleTensor<float> output{ shape_output, DataType::F32, 1 }; + fill(output, 6 + finfo.hash, finfo.min_output, finfo.max_output); + reference::arithmetic_operation<float>(reference::ArithmeticOperation::ADD, output, f32_ref_output, output, ConvertPolicy::SATURATE); + return output; + } + + return f32_ref_output; + } + + TensorType _target{}; + SimpleTensor<float> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, typename TI = uint8_t, typename TW = uint8_t, bool run_twice = false> +class GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture : public GEMMLowpGenericMatrixMultiplyCoreFusedOffsetOutputValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, TI, TW, run_twice> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, GEMMLowpOutputStageType output_stage_type, DataType data_type, bool reshape_b_only_on_first_run) + { + GEMMLowpGenericMatrixMultiplyCoreFusedOffsetOutputValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, TI, TW, run_twice>::setup(shape_a, shape_b, + shape_output, output_stage_type, data_type, reshape_b_only_on_first_run); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, bool reinterpret_input_as_3d = false, bool reinterpret_output_as_3d = false, typename TI = uint8_t, typename TW = uint8_t, bool run_twice = false> +class GEMMLowpBatchedMatrixMultiplyCoreFusedOffsetOutputFixture : public GEMMLowpGenericMatrixMultiplyCoreFusedOffsetOutputValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, TI, TW, run_twice> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, GEMMLowpOutputStageType output_stage_type, DataType data_type, bool reshape_b_only_on_first_run) + { + GEMMLowpGenericMatrixMultiplyCoreFusedOffsetOutputValidationFixture<TensorType, AccessorType, FunctionType, reinterpret_input_as_3d, reinterpret_output_as_3d, TI, TW, run_twice>::setup(shape_a, shape_b, shape_output, output_stage_type, data_type, reshape_b_only_on_first_run); + } +}; + template <typename TensorType, typename AccessorType, typename FunctionType> class GEMMLowpQuantizeDownInt32ToUint8ScaleValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) { _target = compute_target(shape, result_offset, result_mult_int, result_shift, min, max, add_bias); @@ -383,7 +641,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType> class GEMMLowpQuantizeDownInt32ToInt8ScaleValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, int32_t result_offset, int32_t result_mult_int, int32_t result_shift, int32_t min, int32_t max, bool add_bias) { _target = compute_target(shape, result_offset, result_mult_int, result_shift, min, max, add_bias); @@ -485,7 +742,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType> class GEMMLowpQuantizeDownInt32ToInt8ScaleByFixedPointValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias) { _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias); @@ -580,7 +836,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType> class GEMMLowpQuantizeDownInt32ToUint8ScaleByFixedPointValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t result_offset_after_shift, int32_t min, int32_t max, bool add_bias) { _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, result_offset_after_shift, min, max, add_bias); @@ -675,7 +930,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class GEMMLowpQuantizeDownInt32ScaleByFloatValidationFixture : public framework::Fixture { public: - template <typename...> void setup(DataType data_type, TensorShape shape, float result_real_multiplier, int32_t result_offset, int32_t min, int32_t max, bool add_bias) { _target = compute_target(data_type, shape, result_real_multiplier, result_offset, min, max, add_bias); @@ -778,7 +1032,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType> class GEMMLowpQuantizeDownInt32ToInt16ScaleByFixedPointValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, int32_t result_fixedpoint_multiplier, int32_t result_shift, int32_t min, int32_t max, bool add_bias) { _target = compute_target(shape, result_fixedpoint_multiplier, result_shift, min, max, add_bias); @@ -873,7 +1126,6 @@ template <typename TensorType, typename AccessorType, typename ReshapeLHSOperato class GEMMLowpMatrixMultiplyReshapedValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs, bool interleave_rhs, DataType data_type) { @@ -944,7 +1196,7 @@ protected: GEMMFunctionType gemm; reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); - gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K)); + gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), dst.info(), lhs_info, rhs_info, GEMMReshapeInfo(M, N, K)); ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); @@ -973,7 +1225,8 @@ protected: reshape_lhs.run(reshape_lhs_pack); ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; reshape_rhs.run(reshape_rhs_pack); - gemm.run(); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); return dst; } @@ -1023,7 +1276,6 @@ template <typename TensorType, typename AccessorType, typename ReshapeLHSOperato class GEMMLowpMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int v0, unsigned int h0, bool interleave_lhs, bool interleave_rhs, DataType data_type) { @@ -1098,7 +1350,7 @@ protected: GEMMFunctionType gemm; reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); - gemm.configure(&lhs_reshaped, &rhs_reshaped, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h)); + gemm.configure(lhs_reshaped.info(), rhs_reshaped.info(), dst.info(), lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h)); ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); @@ -1127,7 +1379,8 @@ protected: reshape_lhs.run(reshape_lhs_pack); ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; reshape_rhs.run(reshape_rhs_pack); - gemm.run(); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, { ACL_SRC_1, &rhs_reshaped }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); return dst; } @@ -1179,7 +1432,6 @@ template <typename TensorType, typename AccessorType, typename ReshapeRHSOperato class GEMMLowpMatrixMultiplyReshapedOnlyRHSValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, DataType data_type) { @@ -1251,7 +1503,7 @@ protected: ReshapeRHSOperatorType reshape_rhs; GEMMFunctionType gemm; reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); - gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info); + gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info); ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); @@ -1276,7 +1528,8 @@ protected: // Compute GEMM ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; reshape_rhs.run(reshape_rhs_pack); - gemm.run(); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs_reshaped }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); return dst; } @@ -1317,11 +1570,372 @@ protected: SimpleTensor<int32_t> _reference{}; }; +template <typename T, typename TensorType, typename AccessorType, typename ReshapeRHSOperatorType, typename GEMMFunctionType, typename ReduceOperation, typename CastOperation> +class GEMMLowpMatrixMultiplyReshapedOnlyRHSMMULOutputStageValidationFixture : public framework::Fixture +{ +public: + void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, + unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, bool broadcast_bias, DataType data_type) + { + GEMMLowpOutputStageInfo output_stage; + output_stage.type = GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT; + output_stage.output_data_type = data_type; + output_stage.gemmlowp_multipliers = std::vector<int32_t> { 1 }; + output_stage.gemmlowp_shifts = std::vector<int32_t> { 1 }; + output_stage.gemmlowp_multipliers[0] = 1; + output_stage.gemmlowp_shifts[0] = 1; + output_stage.gemmlowp_offset = 0; + constexpr float scale = 0.001f; + quantization::calculate_quantized_multiplier(scale, &output_stage.gemmlowp_multipliers[0], &output_stage.gemmlowp_shifts[0]); + output_stage.gemmlowp_min_bound = -100; + output_stage.gemmlowp_max_bound = 100; + + GEMMLHSMatrixInfo lhs_info; + lhs_info.m0 = m0; + lhs_info.k0 = k0; + + GEMMRHSMatrixInfo rhs_info; + rhs_info.n0 = n0; + rhs_info.k0 = k0; + rhs_info.h0 = h0; + rhs_info.interleave = interleave_rhs; + rhs_info.transpose = transpose_rhs; + + int a_offset = 1; + int b_offset = 1; + + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + const TensorShape bias_shape(n, + broadcast_bias ? 1 : m, + broadcast_bias ? 1 : batch_size); + + _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, output_stage, a_offset, b_offset); + if(gemm_validated == true) + { + _reference = compute_reference(lhs_shape, rhs_shape, bias_shape, data_type, output_stage, a_offset, b_offset); + } + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::QASYMM8: + { + // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path + std::uniform_int_distribution<> distribution(1, 254); + library->fill(tensor, distribution, i); + } + break; + case DataType::QASYMM8_SIGNED: + { + std::uniform_int_distribution<> distribution(-127, 126); + library->fill(tensor, distribution, i); + } + break; + case DataType::S32: + { + std::uniform_int_distribution<> distribution(-10000, 10000); + library->fill(tensor, distribution, i); + } + break; + default: + ARM_COMPUTE_ERROR("Unsupported data type"); + } + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, DataType data_type, GEMMLowpOutputStageInfo output_stage, const int a_offset, const int b_offset) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1, QuantizationInfo(1.0f / 255, a_offset)); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1, QuantizationInfo(1.0f / 255, b_offset)); + TensorType bias = create_tensor<TensorType>(bias_shape, DataType::S32, 1); + TensorType dst; + TensorType rhs_reshaped; + + const unsigned int M = lhs_shape[1]; + const unsigned int N = rhs_shape[0]; + const unsigned int K = lhs_shape[0]; + + // Tensors for precomputing sum of lhs rows / rhs columns + TensorType vec_sum_rows = create_tensor<TensorType>(TensorShape(M, 1, lhs_shape[2]), DataType::S32, 1); + TensorType vec_sum_cols = create_tensor<TensorType>(TensorShape(N, 1, rhs_shape[2]), DataType::S32, 1); + + GEMMKernelInfo gemm_info; + gemm_info.m = M; + gemm_info.n = N; + gemm_info.k = K; + gemm_info.lhs_info = lhs_info; + gemm_info.rhs_info = rhs_info; + gemm_info.output_stage = output_stage; + gemm_info.a_offset = a_offset; + gemm_info.b_offset = b_offset; + // The output tensor will be auto-initialized within the function + + // Create and configure function + ReshapeRHSOperatorType reshape_rhs; + GEMMFunctionType gemm; + reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); + + // If GEMM is not validated, do not try to run. The validation will check + // if the technology supports this extension. If not, the test will be skipped. + // If it supports, the test will fail anyway because target and reference + // will not match. + gemm_validated = bool(gemm.validate(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, vec_sum_cols.info(), vec_sum_rows.info(), bias.info())); + if(gemm_validated == true) + { + gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, vec_sum_cols.info(), vec_sum_rows.info(), bias.info()); + + ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(bias.info()->is_resizable()); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + rhs_reshaped.allocator()->allocate(); + bias.allocator()->allocate(); + vec_sum_cols.allocator()->allocate(); + vec_sum_rows.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!vec_sum_cols.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!vec_sum_rows.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + fill(AccessorType(bias), 2); + + TensorType lhs_32 = create_tensor<TensorType>(lhs_shape, DataType::S32, 1); + TensorType rhs_32 = create_tensor<TensorType>(rhs_shape, DataType::S32, 1); + CastOperation cast_lhs; + CastOperation cast_rhs; + cast_lhs.configure(&lhs, &lhs_32, ConvertPolicy::SATURATE); + cast_rhs.configure(&rhs, &rhs_32, ConvertPolicy::SATURATE); + lhs_32.allocator()->allocate(); + rhs_32.allocator()->allocate(); + cast_lhs.run(); + cast_rhs.run(); + + ReduceOperation lhs_sum_rows; + ReduceOperation rhs_sum_cols; + + lhs_sum_rows.configure(&lhs_32, &vec_sum_rows, 0, ReductionOperation::SUM, false); + rhs_sum_cols.configure(&rhs_32, &vec_sum_cols, 1, ReductionOperation::SUM); + + lhs_sum_rows.run(); + rhs_sum_cols.run(); + + // Compute GEMM + ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; + reshape_rhs.run(reshape_rhs_pack); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs_reshaped }, { ACL_SRC_2, &bias }, { ACL_DST, &dst }, { ACL_VEC_COL_SUM, &vec_sum_cols }, { ACL_VEC_ROW_SUM, &vec_sum_rows } }); + gemm.run(gemm_pack); + } + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const TensorShape &bias_shape, DataType data_type, GEMMLowpOutputStageInfo output_stage, + const int a_offset, const int b_offset) + { + TensorShape dst_shape = lhs_shape; + dst_shape[0] = rhs_shape[0]; + dst_shape[1] = lhs_shape[1]; + + // Create reference + SimpleTensor<T> lhs{ lhs_shape, data_type, 1, QuantizationInfo(1.0f / 255, a_offset) }; + SimpleTensor<T> rhs{ rhs_shape, data_type, 1, QuantizationInfo(1.0f / 255, b_offset) }; + SimpleTensor<int32_t> bias{ bias_shape, DataType::S32, 1 }; + SimpleTensor<int32_t> dst{ dst_shape, DataType::S32, 1 }; + SimpleTensor<T> dst_final{ dst_shape, data_type, 1 }; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + fill(bias, 2); + + dst = reference::gemmlowp_matrix_multiply_core<int32_t, T>(lhs, rhs, dst_shape, a_offset, b_offset); + dst_final = reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, T>(dst, bias, + output_stage.gemmlowp_multipliers, output_stage.gemmlowp_shifts, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); + return dst_final; + } + + bool gemm_validated = true; + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename ReshapeRHSOperatorType, typename GEMMFunctionType> +class GEMMLowpMatrixMultiplyReshapedOnlyRHSMMULValidationFixture : public framework::Fixture +{ +public: + void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, + unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, DataType data_type) + { + GEMMLHSMatrixInfo lhs_info; + lhs_info.m0 = m0; + lhs_info.k0 = k0; + + GEMMRHSMatrixInfo rhs_info; + rhs_info.n0 = n0; + rhs_info.k0 = k0; + rhs_info.h0 = h0; + rhs_info.interleave = interleave_rhs; + rhs_info.transpose = transpose_rhs; + + // Set the tensor shapes for LHS and RHS matrices + const TensorShape lhs_shape(k, m, batch_size); + const TensorShape rhs_shape(n, k, batch_size); + + _target = compute_target(lhs_shape, rhs_shape, lhs_info, rhs_info, data_type); + if(gemm_validated == true) + { + _reference = compute_reference(lhs_shape, rhs_shape, data_type); + } + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::QASYMM8: + { + // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path + std::uniform_int_distribution<> distribution(1, 254); + library->fill(tensor, distribution, i); + } + break; + case DataType::QASYMM8_SIGNED: + { + std::uniform_int_distribution<> distribution(-127, 126); + library->fill(tensor, distribution, i); + } + break; + default: + ARM_COMPUTE_ERROR("Unsupported data type"); + } + } + + TensorType compute_target(const TensorShape &lhs_shape, const TensorShape &rhs_shape, const GEMMLHSMatrixInfo &lhs_info, + const GEMMRHSMatrixInfo &rhs_info, DataType data_type) + { + // Create tensors + TensorType lhs = create_tensor<TensorType>(lhs_shape, data_type, 1); + TensorType rhs = create_tensor<TensorType>(rhs_shape, data_type, 1); + TensorType rhs_reshaped; + TensorType dst; + + const unsigned int M = lhs_shape[1]; + const unsigned int N = rhs_shape[0]; + const unsigned int K = lhs_shape[0]; + + GEMMKernelInfo gemm_info; + gemm_info.m = M; + gemm_info.n = N; + gemm_info.k = K; + gemm_info.lhs_info = lhs_info; + gemm_info.rhs_info = rhs_info; + // The output tensor will be auto-initialized within the function + + // Create and configure function + ReshapeRHSOperatorType reshape_rhs; + GEMMFunctionType gemm; + reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); + + // If GEMM is not validated, do not try to run. The validation will check + // if the technology supports this extension. If not, the test will be skipped. + // If it supports, the test will fail anyway because target and reference + // will not match. + gemm_validated = bool(gemm.validate(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, nullptr, nullptr, nullptr)); + if(gemm_validated == true) + { + gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info, nullptr, nullptr, nullptr); + + ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); + + // Allocate tensors + lhs.allocator()->allocate(); + rhs.allocator()->allocate(); + rhs_reshaped.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(lhs), 0); + fill(AccessorType(rhs), 1); + + // Compute GEMM + ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; + reshape_rhs.run(reshape_rhs_pack); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs_reshaped }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); + } + + return dst; + } + + SimpleTensor<int32_t> compute_reference(const TensorShape &lhs_shape, const TensorShape &rhs_shape, DataType data_type) + { + TensorShape dst_shape = lhs_shape; + dst_shape[0] = rhs_shape[0]; + dst_shape[1] = lhs_shape[1]; + + if(data_type == DataType::QASYMM8) + { + // Create reference + SimpleTensor<uint8_t> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<uint8_t> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<int32_t> dst{ dst_shape, DataType::S32, 1 }; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + + return reference::gemmlowp_matrix_multiply_core<int32_t, uint8_t>(lhs, rhs, dst_shape, 0, 0); + } + else + { + // Create reference + SimpleTensor<int8_t> lhs{ lhs_shape, data_type, 1 }; + SimpleTensor<int8_t> rhs{ rhs_shape, data_type, 1 }; + SimpleTensor<int32_t> dst{ dst_shape, DataType::S32, 1 }; + + // Fill reference + fill(lhs, 0); + fill(rhs, 1); + + return reference::gemmlowp_matrix_multiply_core<int32_t, int8_t>(lhs, rhs, dst_shape, 0, 0); + } + } + + bool gemm_validated = true; + TensorType _target{}; + SimpleTensor<int32_t> _reference{}; +}; + template <typename TensorType, typename AccessorType, typename ReshapeRHSOperatorType, typename GEMMFunctionType> class GEMMLowpMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0, unsigned int h0, bool interleave_rhs, bool transpose_rhs, DataType data_type) { @@ -1397,7 +2011,7 @@ protected: ReshapeRHSOperatorType reshape_rhs; GEMMFunctionType gemm; reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); - gemm.configure(&lhs, &rhs_reshaped, &dst, gemm_info); + gemm.configure(lhs.info(), rhs_reshaped.info(), dst.info(), gemm_info); ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); @@ -1422,7 +2036,8 @@ protected: // Compute GEMM ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; reshape_rhs.run(reshape_rhs_pack); - gemm.run(); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs_reshaped }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); return dst; } @@ -1469,7 +2084,6 @@ template <typename TensorType, typename AccessorType, typename GEMMFunctionType> class GEMMLowpMatrixMultiplyNativeValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0) { GEMMLHSMatrixInfo lhs_info; @@ -1512,7 +2126,7 @@ protected: // Create and configure function GEMMFunctionType gemm; - gemm.configure(&lhs, &rhs, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K)); + gemm.configure(lhs.info(), rhs.info(), dst.info(), lhs_info, rhs_info, GEMMReshapeInfo(M, N, K)); ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); @@ -1533,7 +2147,8 @@ protected: fill(AccessorType(rhs), 1); // Compute GEMM - gemm.run(); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); return dst; } @@ -1563,7 +2178,6 @@ template <typename TensorType, typename AccessorType, typename GEMMFunctionType> class GEMMLowpMatrixMultiplyNative3DValidationFixture : public framework::Fixture { public: - template <typename...> void setup(unsigned int m_w, unsigned int m_h, unsigned int n, unsigned int k, unsigned int batch_size, unsigned int m0, unsigned int n0, unsigned int k0) { GEMMLHSMatrixInfo lhs_info; @@ -1609,7 +2223,7 @@ protected: // Create and configure function GEMMFunctionType gemm; - gemm.configure(&lhs, &rhs, &dst, lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h)); + gemm.configure(lhs.info(), rhs.info(), dst.info(), lhs_info, rhs_info, GEMMReshapeInfo(M, N, K, 1, 1, m_h)); ARM_COMPUTE_ASSERT(lhs.info()->is_resizable()); ARM_COMPUTE_ASSERT(rhs.info()->is_resizable()); @@ -1630,7 +2244,8 @@ protected: fill(AccessorType(rhs), 1); // Compute GEMM - gemm.run(); + ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, { ACL_SRC_1, &rhs }, { ACL_DST, &dst } }); + gemm.run(gemm_pack); return dst; } @@ -1660,4 +2275,4 @@ protected: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_GEMMLOWPFIXTURE_H diff --git a/tests/validation/fixtures/GEMMReshapeLHSMatrixFixture.h b/tests/validation/fixtures/GEMMReshapeLHSMatrixFixture.h index a9d6c9b6aa..d88029f93e 100644 --- a/tests/validation/fixtures/GEMMReshapeLHSMatrixFixture.h +++ b/tests/validation/fixtures/GEMMReshapeLHSMatrixFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,6 @@ template <typename TensorType, typename AccessorType, typename OperatorType, typ class GEMMReshapeLHSMatrixValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape_in, unsigned int batch_size, DataType data_type, unsigned int m0, unsigned int k0, unsigned int v0, bool interleave, bool transpose) { GEMMLHSMatrixInfo lhs_info; diff --git a/tests/validation/fixtures/GEMMReshapeRHSMatrixFixture.h b/tests/validation/fixtures/GEMMReshapeRHSMatrixFixture.h index cdb3ec3944..0929faf04a 100644 --- a/tests/validation/fixtures/GEMMReshapeRHSMatrixFixture.h +++ b/tests/validation/fixtures/GEMMReshapeRHSMatrixFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,6 @@ template <typename TensorType, typename AccessorType, typename OperatorType, typ class GEMMReshapeRHSMatrixValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape_in, unsigned int batch_size, DataType data_type, unsigned int n0, unsigned int k0, unsigned int h0, bool interleave, bool transpose) { GEMMRHSMatrixInfo rhs_info; diff --git a/tests/validation/fixtures/GEMMTranspose1xWFixture.h b/tests/validation/fixtures/GEMMTranspose1xWFixture.h index 7caea1d3d8..3765515b57 100644 --- a/tests/validation/fixtures/GEMMTranspose1xWFixture.h +++ b/tests/validation/fixtures/GEMMTranspose1xWFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class GEMMTranspose1xWValidationFixture : public framework::Fixture { public: - template <typename...> void setup(size_t x, size_t y, DataType data_type) { _data_type = data_type; @@ -89,7 +88,7 @@ protected: // Create and configure function FunctionType f; - f.configure(&a, &b); + f.configure(a.info(), b.info()); ARM_COMPUTE_ASSERT(a.info()->is_resizable()); ARM_COMPUTE_ASSERT(b.info()->is_resizable()); @@ -106,7 +105,8 @@ protected: fill(AccessorType(b), 1); // Compute GEMM function - f.run(); + ITensorPack tensors{ { ACL_SRC, &a }, { ACL_DST, &b } }; + f.run(tensors); return b; } diff --git a/tests/validation/fixtures/GatherFixture.h b/tests/validation/fixtures/GatherFixture.h index 452a201f82..857b0387b7 100644 --- a/tests/validation/fixtures/GatherFixture.h +++ b/tests/validation/fixtures/GatherFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class GatherFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape indices_shape, int axis, DataType data_type) { _target = compute_target(input_shape, data_type, axis, indices_shape); @@ -67,9 +66,11 @@ protected: std::mt19937 gen(library->seed()); uint32_t *indices_ptr = static_cast<uint32_t *>(indices.data()); - std::uniform_int_distribution<uint32_t> dist_index(0, input_shape[actual_axis] - 1); - //Let's consider 1D indices - for(unsigned int ind = 0; ind < indices_shape[0]; ind++) + // 10% of the time the index is out-of-range. + uint32_t max_index = input_shape[actual_axis] + input_shape[actual_axis] / 9 + 1; + std::uniform_int_distribution<uint32_t> dist_index(0, max_index - 1); + + for(unsigned int ind = 0; ind < indices_shape.total_size(); ind++) { indices_ptr[ind] = dist_index(gen); } diff --git a/tests/validation/fixtures/Im2ColFixture.h b/tests/validation/fixtures/Im2ColFixture.h index b1fbd76eb2..5c7978f4ab 100644 --- a/tests/validation/fixtures/Im2ColFixture.h +++ b/tests/validation/fixtures/Im2ColFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,10 +45,9 @@ namespace validation using namespace arm_compute::misc::shape_calculator; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool batch_size_on_z> -class Im2ColValidationFixture : public framework::Fixture +class Im2ColOpValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, DataType data_type, const Size2D &kernel_dims, const PadStrideInfo &conv_info, const QuantizationInfo &quant_info, const DataLayout &data_layout, unsigned int num_groups) { @@ -88,7 +87,7 @@ protected: // Create and configure function FunctionType im2col_func; - im2col_func.configure(&src, &dst, _kernel_dims, _conv_info, _has_bias, Size2D(1U, 1U), _num_groups); + im2col_func.configure(src.info(), dst.info(), _kernel_dims, _conv_info, _has_bias, Size2D(1U, 1U), _num_groups); ARM_COMPUTE_ASSERT(src.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); @@ -103,8 +102,13 @@ protected: // Fill tensors fill(AccessorType(src)); + arm_compute::ITensorPack pack = + { + { arm_compute::TensorType::ACL_SRC, &src }, + { arm_compute::TensorType::ACL_DST, &dst } + }; // Compute function - im2col_func.run(); + im2col_func.run(pack); return dst; } diff --git a/tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h b/tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h new file mode 100644 index 0000000000..7374093f51 --- /dev/null +++ b/tests/validation/fixtures/IndirectConv2dAddressPrecalculationFixture.h @@ -0,0 +1,122 @@ +/* + * Copyright (c) 2022-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_FIXTURE +#define ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "tests/Globals.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/reference/IndirectConv2dAddressPrecalculation.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::misc::shape_calculator; + +template <typename TensorType, typename AccessorType, typename OperatorType> +class IndirectConv2dAddressPrecalculationValidationFixture : public framework::Fixture +{ +public: + void setup(unsigned int src_w, + unsigned int src_h, + unsigned int src_b, + unsigned int wei_w, + unsigned int wei_h, + unsigned int pad, + unsigned int stride, + unsigned int m0) + { + DirectConvComputeKernelInfo desc; + desc.m0 = m0; + desc.n0 = 1; // Not used by the kernel + desc.k0 = 1; // Not used by the kernel + desc.export_weights_to_cl_image = false; // Not used by the kernel + + const PadStrideInfo conv_info(stride, stride, pad, pad); + + const TensorShape shape_conv_src(23, // The input channels are not used by the kernel + src_w, + src_h, + src_b); + + const TensorShape shape_conv_wei(23, // The input channels are not used by the kernel + wei_w, + wei_h, + 23 // The output channels are not used by the kernel + ); + + // The result of the kernel does not change with the datatype. Hence, we can fix it to Fp16 for validation purposes + const DataType data_type = DataType::F16; + + _target = compute_target(shape_conv_src, shape_conv_wei, data_type, conv_info, desc); + _reference = compute_reference(shape_conv_src, shape_conv_wei, data_type, conv_info, desc); + } + +protected: + TensorType compute_target(TensorShape shape_conv_src, TensorShape shape_conv_wei, DataType data_type, const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc) + { + TensorInfo src_conv_info(shape_conv_src, 1, data_type, DataLayout::NHWC); + TensorInfo wei_conv_info(shape_conv_wei, 1, data_type, DataLayout::NHWC); + TensorType dst; + + // The output tensor will be auto-initialized within the function + + // Create and configure function + OperatorType func; + func.configure(&src_conv_info, &wei_conv_info, dst.info(), conv_info, desc); + + add_padding_x({ &dst }); + + // Allocate tensors + dst.allocator()->allocate(); + + // Compute GEMM LHS matrix reshape function + ITensorPack tensors = { { ACL_DST, &dst } }; + func.run(tensors); + + return dst; + } + + SimpleTensor<int32_t> compute_reference(TensorShape shape_conv_src, TensorShape shape_conv_wei, DataType data_type, const PadStrideInfo &conv_info, const DirectConvComputeKernelInfo &desc) + { + ARM_COMPUTE_UNUSED(data_type); + TensorShape shape_out = compute_indirect_buffer_shape(shape_conv_src, DataLayout::NHWC, shape_conv_wei, conv_info, desc); + TensorShape output_conv_shape = compute_deep_convolution_shape(shape_conv_src, DataLayout::NHWC, shape_conv_wei, conv_info); + + return reference::indirect_conv2d_addr_precalculation(shape_conv_src, shape_conv_wei, output_conv_shape, shape_out, conv_info); + } + + TensorType _target{}; + SimpleTensor<int32_t> _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_FIXTURE */
\ No newline at end of file diff --git a/tests/validation/fixtures/InstanceNormalizationLayerFixture.h b/tests/validation/fixtures/InstanceNormalizationLayerFixture.h index 0f8c155277..c26dd99f02 100644 --- a/tests/validation/fixtures/InstanceNormalizationLayerFixture.h +++ b/tests/validation/fixtures/InstanceNormalizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class InstanceNormalizationLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, DataLayout data_layout, bool in_place) { _target = compute_target(shape, data_type, data_layout, in_place); diff --git a/tests/validation/fixtures/L2NormalizeLayerFixture.h b/tests/validation/fixtures/L2NormalizeLayerFixture.h index 2dea66c98b..b8f4b1eaf3 100644 --- a/tests/validation/fixtures/L2NormalizeLayerFixture.h +++ b/tests/validation/fixtures/L2NormalizeLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -48,7 +48,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class L2NormalizeLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, DataLayout data_layout, int axis, float epsilon) { _target = compute_target(shape, data_type, data_layout, axis, epsilon); diff --git a/tests/validation/fixtures/LSTMLayerFixture.h b/tests/validation/fixtures/LSTMLayerFixture.h index f4bae86d30..a32e9adfe5 100644 --- a/tests/validation/fixtures/LSTMLayerFixture.h +++ b/tests/validation/fixtures/LSTMLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class LSTMLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape input_weights_shape, TensorShape recurrent_weights_shape, TensorShape cell_bias_shape, TensorShape output_cell_shape, TensorShape output_shape, TensorShape scratch_shape, ActivationLayerInfo info, float cell_threshold, float projection_threshold, DataType data_type, bool projection_opt, bool peephole_opt, bool use_layer_norm) @@ -458,7 +457,6 @@ protected: } input_gate = reference::activation_layer(input_gate, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); } - // Compute cell_state SimpleTensor<T> fully_connected_cell_state = reference::fully_connected_layer(input, input_to_cell_w, cell_bias, output_cell_shape); transposed_weights = reference::transpose(recurrent_to_cell_w); @@ -474,12 +472,13 @@ protected: fill(cell_bias, 8); cell_state_out = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, cell_state_out, cell_bias, data_type, ConvertPolicy::SATURATE); } - cell_state_out = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LOGISTIC)); + cell_state_out = reference::activation_layer(cell_state_out, info); cell_state_out = reference::pixel_wise_multiplication<T, T, T>(cell_state_out, input_gate, 1, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_EVEN, data_type); cell_state_out = reference::arithmetic_operation(reference::ArithmeticOperation::ADD, cell_state_out, pixelwise_mul, data_type, ConvertPolicy::SATURATE); + if(cell_threshold != 0.f) { - cell_state_out = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -cell_threshold, cell_threshold)); + cell_state_out = reference::activation_layer(cell_state_out, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, cell_threshold, -cell_threshold)); } // Compute output @@ -515,7 +514,6 @@ protected: output_state_out = reference::activation_layer(fully_connected_projection, ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, -projection_threshold, projection_threshold)); } } - std::vector<SimpleTensor<T>> scratch_inputs; if(!cifg_opt) { diff --git a/tests/validation/fixtures/LogicalFixture.h b/tests/validation/fixtures/LogicalFixture.h index c7dd1f6510..60dc963ba7 100644 --- a/tests/validation/fixtures/LogicalFixture.h +++ b/tests/validation/fixtures/LogicalFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2021 Arm Limited. + * Copyright (c) 2020-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -79,7 +79,6 @@ class LogicalBinaryOperationValidationFixture : public LogicalOperationValidatio using Parent = LogicalOperationValidationFixtureBase<TensorType, AccessorType, FunctionType, T>; public: - template <typename...> void setup(TensorShape shape0, TensorShape shape1) { Parent::_target = compute_target(shape0, shape1); @@ -135,7 +134,6 @@ class LogicalNotValidationFixture : public LogicalOperationValidationFixtureBase using Parent = LogicalOperationValidationFixtureBase<TensorType, AccessorType, FunctionType, T>; public: - template <typename...> void setup(TensorShape shape, DataType data_type) { Parent::_target = compute_target(shape, data_type); diff --git a/tests/validation/fixtures/MatMulFixture.h b/tests/validation/fixtures/MatMulFixture.h new file mode 100644 index 0000000000..ffd12e56d0 --- /dev/null +++ b/tests/validation/fixtures/MatMulFixture.h @@ -0,0 +1,612 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_MATMULFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_MATMULFIXTURE_H + +#include "arm_compute/core/Types.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" + +#include "src/core/utils/quantization/AsymmHelpers.h" +#include "tests/framework/Asserts.h" // Required for ARM_COMPUTE_ASSERT +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/ActivationLayer.h" +#include "tests/validation/reference/GEMM.h" +#include "tests/validation/reference/GEMMLowp.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/reference/ReshapeLayer.h" +#include "tests/validation/Validation.h" + +#include <limits> +#include <random> +#include <type_traits> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class MatMulGenericValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + int num_extra_runs, + Settings settings, + QuantizationInfo a_qinfo = QuantizationInfo(), + QuantizationInfo b_qinfo = QuantizationInfo(), + QuantizationInfo o_qinfo = QuantizationInfo()) + { + // For brevity, the input shapes are assumed to be not-transposed for both a and b matrices. + if (transpose_a) + { + permute(shape_a, PermutationVector(1U, 0U)); + } + if (transpose_b) + { + permute(shape_b, PermutationVector(1U, 0U)); + } + + _target = compute_target(shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, + num_extra_runs, settings, a_qinfo, b_qinfo, o_qinfo); + _reference = compute_reference(shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, + a_qinfo, b_qinfo, o_qinfo); + } + +protected: + template <typename U> + void fill(U &&tensor, int i, float lo = -1.f, float hi = 1.f) + { + switch (tensor.data_type()) + { + case DataType::BFLOAT16: + { + arm_compute::utils::uniform_real_distribution_16bit<bfloat16> distribution{float(lo), float(hi)}; + library->fill(tensor, distribution, i); + break; + } + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{float(lo), float(hi)}; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(lo, hi); + library->fill(tensor, distribution, i); + break; + } + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + { + library->fill_tensor_uniform(tensor, i); + break; + } + default: + { + ARM_COMPUTE_ERROR("Unsupported data type."); + } + } + } + + virtual TensorType compute_target(const TensorShape &shape_a, + const TensorShape &shape_b, + const TensorShape &output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + int num_extra_runs, + const Settings &settings, + QuantizationInfo a_qinfo, + QuantizationInfo b_qinfo, + QuantizationInfo o_qinfo) + { + // 1. Create Classes and configure function + // ---------------------------------------------------- + // Create tensors + // Configure relevant classes and matmul function + TensorType a = create_tensor<TensorType>(shape_a, data_type, 1, a_qinfo); + TensorType b = create_tensor<TensorType>(shape_b, data_type, 1, b_qinfo); + TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, o_qinfo); + + FunctionType matmul; + + // Configure MatMulInfo class + MatMulInfo mm_info; + mm_info.adj_lhs(transpose_a).adj_rhs(transpose_b); + + // Ensure values are dynamic + a.info()->set_are_values_constant(false); + b.info()->set_are_values_constant(false); + + // Configure operator + matmul.configure(&a, &b, &dst, mm_info, settings, act_info); + + // Assertions + ARM_COMPUTE_ASSERT(a.info()->is_resizable()); + ARM_COMPUTE_ASSERT(b.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + // Allocate tensors + a.allocator()->allocate(); + b.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!a.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!b.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // For multiple runs. + for (int i = 0; i < num_extra_runs; i++) + { + // Stress dynamic tensors by running multiple times. + // -------------------------------------------------------- + // Fill tensors with new seed + // Run function + const int seed_offset = num_extra_runs * 100; + fill(AccessorType(a), seed_offset); + fill(AccessorType(b), seed_offset + 1); + + matmul.run(); + } + + // 2. Final Run for reference comparison + // -------------------------------------------------------- + // Re-fill tensors same seed as reference run + // Compute MatMul operation + fill(AccessorType(a), 2); + fill(AccessorType(b), 3); + + matmul.run(); + + return dst; + } + + template <typename TT> + typename std::enable_if < !std::is_integral<TT>::value, SimpleTensor<TT >>::type + compute_reference_gemm(const SimpleTensor<TT> &a, + const SimpleTensor<TT> &b, + const SimpleTensor<TT> &c, + float alpha, + float beta, + const QuantizationInfo &o_qinfo) + { + ARM_COMPUTE_UNUSED(o_qinfo); + + return reference::gemm(a, b, c, alpha, beta); + } + + template <typename TT> + typename std::enable_if<std::is_integral<TT>::value, SimpleTensor<TT>>::type + compute_reference_gemm(const SimpleTensor<TT> &a, + const SimpleTensor<TT> &b, + const SimpleTensor<TT> &c, + float alpha, + float beta, + const QuantizationInfo &o_qinfo) + { + ARM_COMPUTE_UNUSED(alpha, beta); + + const auto aq = a.quantization_info().uniform(); + const auto bq = b.quantization_info().uniform(); + const auto oq = o_qinfo.uniform(); + + const auto multiplier = aq.scale * bq.scale / oq.scale; + + int32_t output_multiplier = 0; + int32_t output_shift = 0; + quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift); + std::vector<int32_t> output_multipliers{output_multiplier}; + std::vector<int32_t> output_shifts{output_shift}; + + //The lhs and rhs offsets are negated here to keep the reference aligned with the function implementation where the lhs and rhs offsets are also negated. + const auto tmp = reference::gemmlowp_matrix_multiply_core<int32_t>(a, b, c.shape(), -aq.offset, -bq.offset); + + auto output = reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, TT>( + tmp, output_multipliers, output_shifts, oq.offset, std::numeric_limits<int32_t>::lowest(), + std::numeric_limits<int32_t>::max()); + output.quantization_info(o_qinfo); + + return output; + } + + SimpleTensor<T> compute_reference(const TensorShape &a_shape, + const TensorShape &b_shape, + const TensorShape &output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + QuantizationInfo a_qinfo, + QuantizationInfo b_qinfo, + QuantizationInfo o_qinfo) + { + // We collapse dimensions > 2 onto dimension 2, i.e. 4D+ tensors will look like 3D + // This is necessary unless we choose to extend gemm reference for 4D+ tensors + TensorShape output_shape_collapsed = output_shape.collapsed_from(Window::DimZ); + TensorShape a_shape_collapsed = a_shape.collapsed_from(Window::DimZ); + TensorShape b_shape_collapsed = b_shape.collapsed_from(Window::DimZ); + + // Create reference + SimpleTensor<T> a{a_shape_collapsed, data_type, 1, a_qinfo}; + SimpleTensor<T> b{b_shape_collapsed, data_type, 1, b_qinfo}; + SimpleTensor<T> c{output_shape_collapsed, data_type, 1}; + + // Fill reference + fill(a, 2); + fill(b, 3); + + /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if transpose_a is set to true, then A is assumed to be (B x K x M), + therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K) + in order to be able to call reference implementation that works with (B x M x K) input. + Similarly, if transpose_b is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */ + + // Define transposed shapes + TensorShape a_transposed_shape(a.shape()); + a_transposed_shape.set(0, a.shape().y()); + a_transposed_shape.set(1, a.shape().x()); + + TensorShape b_transposed_shape(b.shape()); + b_transposed_shape.set(0, b.shape().y()); + b_transposed_shape.set(1, b.shape().x()); + + // Define transposed tensors + SimpleTensor<T> a_transposed{a_transposed_shape, data_type}; + SimpleTensor<T> b_transposed{b_transposed_shape, data_type}; + + // pretranspose a if necessary + if (transpose_a) + { + a_transposed = reference::permute<T>(a, PermutationVector(1U, 0U)); + } + // pretranspose b if necessary + if (transpose_b) + { + b_transposed = reference::permute<T>(b, PermutationVector(1U, 0U)); + } + + // Setting beta to 0 will effectively disable C for the + // computation of the reference: alpha * A * B + 0 * C + // Use transposed tensors if boolean enabled else use original tensors + auto result = compute_reference_gemm<T>((transpose_a) ? a_transposed : a, (transpose_b) ? b_transposed : b, c, + 1.0f, 0.f, o_qinfo); + + result = reference::activation_layer<T>(result, act_info, o_qinfo); + + // We reshape the gemm output back if the tensor is high dimensional + if (output_shape_collapsed != output_shape) + { + result = reference::reshape_layer(result, output_shape); + } + + return result; + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +/// TODO: (ONCPUML-1451) The current state of this fixture is interim and a longer-term testing method will be implemented later. +/// @note: Currently we support only a 2x2 test due to the lack of reorder ref. implementation. +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class MatMulFixedFormatFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + TensorType compute_target(const TensorShape &shape_a, + const TensorShape &shape_b, + const TensorShape &output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + int num_extra_runs, + const Settings &settings, + QuantizationInfo a_qinfo, + QuantizationInfo b_qinfo, + QuantizationInfo o_qinfo) override + { + // 1. Create Classes and configure function + // ---------------------------------------------------- + // Create tensors + // Configure relevant classes and matmul function + TensorType a = create_tensor<TensorType>(shape_a, data_type, 1, a_qinfo); + TensorType b = create_tensor<TensorType>(shape_b, data_type, 1, b_qinfo); + TensorType dst = create_tensor<TensorType>(output_shape, data_type, 1, o_qinfo); + + const auto weight_tensor_info = TensorInfo(*b.info()); + const TensorInfo new_tensor_info = prepare_weights(weight_tensor_info); + TensorType weights_transformed = create_tensor<TensorType>(new_tensor_info); + + // Configure MatMulInfo class + MatMulInfo mm_info; + mm_info.adj_lhs(transpose_a).adj_rhs(transpose_b); + + // Ensure values are dynamic + a.info()->set_are_values_constant(false); + b.info()->set_are_values_constant(false); + weights_transformed.info()->set_are_values_constant(false); + + FunctionType matmul; + + // Configure operator + matmul.configure(&a, &weights_transformed, &dst, mm_info, settings, act_info); + + // Assertions + ARM_COMPUTE_ASSERT(a.info()->is_resizable()); + ARM_COMPUTE_ASSERT(b.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + ARM_COMPUTE_ASSERT(weights_transformed.info()->is_resizable()); + + // Allocate tensors + a.allocator()->allocate(); + b.allocator()->allocate(); + dst.allocator()->allocate(); + weights_transformed.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!a.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!b.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!weights_transformed.info()->is_resizable()); + + // For multiple runs. + for (int i = 0; i < num_extra_runs; i++) + { + // Stress dynamic tensors by running multiple times. + // -------------------------------------------------------- + // Fill tensors with new seed + // Run function + const int seed_offset = num_extra_runs * 100; + this->fill(AccessorType(a), seed_offset); + this->fill(AccessorType(b), seed_offset + 1); + + matmul.run(); + } + + // 2. Final Run for reference comparison + // -------------------------------------------------------- + // Re-fill tensors same seed as reference run + // Compute MatMul operation + this->fill(AccessorType(a), 2); + this->fill(AccessorType(b), 3); + + rearrange_data(AccessorType(b), AccessorType(weights_transformed)); + + matmul.run(); + + return dst; + } + + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + int num_extra_runs, + Settings settings, + QuantizationInfo a_qinfo, + QuantizationInfo b_qinfo, + QuantizationInfo o_qinfo) + { + if (CPUInfo::get().has_bf16()) + { + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, num_extra_runs, settings, + a_qinfo, b_qinfo, o_qinfo); + } + } + +private: + TensorInfo prepare_weights(const TensorInfo tensor_info) + { + const DataLayout data_layout = tensor_info.data_layout(); + ARM_COMPUTE_EXPECT(data_layout == DataLayout::NCHW, framework::LogLevel::ERRORS); + const DataType data_type = tensor_info.data_type(); + const TensorShape tensor_shape = tensor_info.tensor_shape(); + const int H = tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)]; + const int W = tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)]; + ARM_COMPUTE_EXPECT(H <= 2 && W <= 2, framework::LogLevel::ERRORS); + + arm_compute::Strides strides_in_bytes = tensor_info.strides_in_bytes(); + strides_in_bytes.set(1, 32); + strides_in_bytes.set(2, 32); + + const size_t offset_first_element_in_bytes = tensor_info.offset_first_element_in_bytes(); + const size_t total_size_in_bytes = 32; + + const TensorShape TS(H, W); + + TensorInfo new_tensor_info = tensor_info; + new_tensor_info.init(TS, tensor_info.num_channels(), data_type, strides_in_bytes, offset_first_element_in_bytes, + total_size_in_bytes); + + return new_tensor_info; + } + + void rearrange_data(const AccessorType src, AccessorType dst) + { + const TensorShape src_tensor_shape = src.shape(); + const DataLayout data_layout = src.data_layout(); + ARM_COMPUTE_EXPECT(data_layout == DataLayout::NCHW, framework::LogLevel::ERRORS); + const unsigned int O = + src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES)]; // N=O + const unsigned int H = + src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT)]; + const unsigned int W = + src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH)]; + const unsigned int I = + src_tensor_shape[get_data_layout_dimension_index(data_layout, DataLayoutDimension::CHANNEL)]; // C=I + ARM_COMPUTE_EXPECT(H <= 2 && W <= 2, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(I == 1 && O == 1, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(src.num_elements() <= dst.num_elements(), framework::LogLevel::ERRORS); + + const T *src_ptr = reinterpret_cast<const T *>(src.data()); + T *dst_ptr = reinterpret_cast<T *>(dst.data()); + + // rearrange indexes for 2x2 input and weight + int dst_idx[] = {0, 4, 1, 5}; + for (int i = 0; i < 4; i++) + { + dst_ptr[dst_idx[i]] = src_ptr[i]; + } + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class MatMulValidationFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type) + { + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, ActivationLayerInfo(), 0, Settings()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class MatMulValidationWithDynamicTensorsFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + int num_extra_runs) + { + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, num_extra_runs, Settings()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class QuantizedMatMulValidationFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info, + int num_extra_runs, + QuantizationInfo a_qinfo, + QuantizationInfo b_qinfo, + QuantizationInfo o_qinfo) + { + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, num_extra_runs, Settings(), + a_qinfo, b_qinfo, o_qinfo); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class MatMulValidationWithActivationFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo act_info) + { + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, 0, Settings()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class MatMulValidationWithActivationAlphaBetaFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo::ActivationFunction function, + float alpha_beta) + { + ActivationLayerInfo act_info(function, alpha_beta, alpha_beta); + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, 0, Settings()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename Settings, typename T> +class QuantizedMatMulValidationWithActivationFixture + : public MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T> +{ +public: + void setup(TensorShape shape_a, + TensorShape shape_b, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + DataType data_type, + ActivationLayerInfo::ActivationFunction function, + float alpha_beta, + int num_extra_runs, + QuantizationInfo a_qinfo, + QuantizationInfo b_qinfo, + QuantizationInfo o_qinfo) + { + ActivationLayerInfo act_info(function, alpha_beta, alpha_beta); + MatMulGenericValidationFixture<TensorType, AccessorType, FunctionType, Settings, T>::setup( + shape_a, shape_b, output_shape, transpose_a, transpose_b, data_type, act_info, num_extra_runs, Settings(), + a_qinfo, b_qinfo, o_qinfo); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_MATMULFIXTURE_H diff --git a/tests/validation/fixtures/MatMulKernelFixture.h b/tests/validation/fixtures/MatMulKernelFixture.h new file mode 100644 index 0000000000..26072dff65 --- /dev/null +++ b/tests/validation/fixtures/MatMulKernelFixture.h @@ -0,0 +1,390 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_MATMULKERNELFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_MATMULKERNELFIXTURE_H + +#include "arm_compute/core/KernelDescriptors.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/CL/Helper.h" +#include "tests/framework/Asserts.h" // Required for ARM_COMPUTE_ASSERT +#include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/Validation.h" +#include "tests/validation/reference/GEMM.h" +#include "tests/validation/reference/GEMMLowp.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/reference/ReshapeLayer.h" +#include <cmath> +#include <random> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +using namespace arm_compute::opencl::kernels; + +template <typename T, typename KernelType, bool use_mmul = false> +class MatMulKernelGenericValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape output_shape, bool pretranspose_a, bool pretranspose_b, int M0, int N0, int K0, bool export_rhs_to_cl_image, DataType data_type, + bool enable_bias) + { + // This hash is used by random generators. There may be hash collisions but + // this is intentional as it's a very easy way to make the the current + // random generation process almost different for many test configurations, + // which were using the same set of values before. + _hash = M0 + N0 + K0 + shape_a[0] + shape_a[1] + shape_b[0] + shape_b[1] + enable_bias + export_rhs_to_cl_image; + + // Flag to create a bias + _enable_bias = enable_bias; + + // For brevity, the input shapes are assumed to be not-transposed for both Lhs and Rhs matrices. + QuantizationInfo lhs_q_info; + QuantizationInfo rhs_q_info; + QuantizationInfo dst_q_info; + + if(is_data_type_quantized(data_type)) + { + const int32_t t_max = static_cast<int32_t>(std::numeric_limits<T>::max()); + const int32_t t_min = static_cast<int32_t>(std::numeric_limits<T>::min()); + + std::mt19937 generator(library->seed() + _hash); + std::uniform_real_distribution<float> distribution_float(-5.0f, 3.0f); + std::uniform_int_distribution<int32_t> distribution_t(t_min, t_max); + + const float scale_lhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + const float scale_rhs = pow(2, distribution_float(generator)); // [2^-5, 2^3] + + const int32_t offset_lhs = distribution_t(generator); + const int32_t offset_rhs = distribution_t(generator); + + lhs_q_info = QuantizationInfo(scale_lhs, offset_lhs); + rhs_q_info = QuantizationInfo(scale_rhs, offset_rhs); + + const int m = shape_a.y(); + const int n = shape_b.x(); + const int k = shape_a.x(); + + const float bias_fraction = enable_bias ? 0.5f : 0.f; + + QuantizationHint q_hint = suggest_matmul_dst_q_info_and_bias(lhs_q_info, rhs_q_info, m, n, k, data_type, bias_fraction); + dst_q_info = q_hint.q_info; + _min_bias = q_hint.bias_min; + _max_bias = q_hint.bias_max; + } + + if(pretranspose_a) + { + permute(shape_a, PermutationVector(1U, 0U)); + } + + if(pretranspose_b) + { + permute(shape_b, PermutationVector(1U, 0U)); + } + + // Skip configurations unsupported by the device. + _device_supports_export_to_cl_image = image2d_from_buffer_supported(CLKernelLibrary::get().get_device()); + if(!_device_supports_export_to_cl_image && export_rhs_to_cl_image) + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + return; // Note: Also need to skip the validate in corresponding FIXTURE_DATA_TEST_CASEs. + } + + _device_supports_mmul = arm_matrix_multiply_supported(CLKernelLibrary::get().get_device()); + if(!_device_supports_mmul && use_mmul) + { + ARM_COMPUTE_TEST_INFO("cl_arm_matrix_multiply not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + return; // Note: Also need to skip the validate in corresponding FIXTURE_DATA_TEST_CASEs. + } + + _target = compute_target(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, M0, N0, K0, export_rhs_to_cl_image, data_type, lhs_q_info, rhs_q_info, dst_q_info); + _reference = compute_reference(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, data_type, lhs_q_info, rhs_q_info, dst_q_info); + } + +protected: + template <typename U> + void fill(U &&tensor, int i, float lo = -1.f, float hi = 1.f) + { + switch(tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ float(lo), float(hi) }; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(lo, hi); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + + template <typename U> + void fill_bias_s32(U &&tensor, int i, int32_t min, int32_t max) + { + std::uniform_int_distribution<int32_t> distribution(min, max); + library->fill(tensor, distribution, i); + } + + template <typename U, typename D> + void fill_constant(U &&tensor, D value) + { + library->fill_tensor_value(tensor, value); + } + + CLTensor compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, bool pretranspose_a, bool pretranspose_b, const int M0, const int N0, const int K0, + bool export_rhs_to_cl_image, DataType data_type, const QuantizationInfo &lhs_q_info, const QuantizationInfo &rhs_q_info, const QuantizationInfo &dst_q_info) + { + CLSynthetizeOperator<KernelType> matMul{}; + MatMulKernelInfo matmul_info; + matmul_info.adj_lhs = pretranspose_a; + matmul_info.adj_rhs = pretranspose_b; + matmul_info.m0 = M0; + matmul_info.n0 = N0; + matmul_info.k0 = K0; + matmul_info.export_rhs_to_cl_image = export_rhs_to_cl_image; + + bool is_quantized = is_data_type_quantized(data_type); + + // Create tensors + CLTensor a = create_tensor<CLTensor>(shape_a, data_type, 1, lhs_q_info); + CLTensor b = create_tensor<CLTensor>(shape_b, data_type, 1, rhs_q_info); + CLTensor bias = create_tensor<CLTensor>(output_shape[0], (is_quantized) ? DataType::S32 : data_type, 1, dst_q_info); + CLTensor dst = create_tensor<CLTensor>(output_shape, data_type, 1, dst_q_info); + + matMul.configure(a.info(), b.info(), (_enable_bias) ? bias.info() : nullptr, dst.info(), matmul_info); + ARM_COMPUTE_ASSERT(a.info()->is_resizable()); + ARM_COMPUTE_ASSERT(b.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + // Allocate tensors + a.allocator()->allocate(); + b.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!a.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!b.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(CLAccessor(a), _hash + 1); + fill(CLAccessor(b), _hash + 2); + + // Compute matMul kernel + ITensorPack tensors_pack({ { ACL_SRC_0, &a }, + { ACL_SRC_1, &b }, + { ACL_DST, &dst } + }); + + if(_enable_bias) + { + // Allocate, fill and add bias to TensorPack obj + bias.allocator()->allocate(); + if(is_quantized) + { + fill_bias_s32(CLAccessor(bias), _hash + 3, _min_bias, _max_bias); + } + else + { + fill(CLAccessor(bias), _hash + 3); + } + tensors_pack.add_tensor(ACL_SRC_2, &bias); + } + + matMul.run(tensors_pack); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &output_shape, bool pretranspose_a, bool pretranspose_b, DataType data_type, + const QuantizationInfo &lhs_q_info, const QuantizationInfo &rhs_q_info, const QuantizationInfo &dst_q_info) + { + // We collapse dimensions > 3 onto dimension 3, i.e. 5D+ tensors will look like 4D + // This is necessary unless we choose to extend gemm reference for 5D+ tensors + TensorShape output_shape_collapsed = output_shape.collapsed_from(Window::DimZ); + TensorShape shape_a_collapsed = shape_a.collapsed_from(Window::DimZ); + TensorShape shape_b_collapsed = shape_b.collapsed_from(Window::DimZ); + + // Create reference + SimpleTensor<T> a{ shape_a_collapsed, data_type, 1, lhs_q_info }; + SimpleTensor<T> b{ shape_b_collapsed, data_type, 1, rhs_q_info }; + SimpleTensor<T> c{ output_shape_collapsed, data_type, 1, dst_q_info }; + + // Fill reference + fill(a, _hash + 1); + fill(b, _hash + 2); + + /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if pretranspose_A is set to true, then A is assumed to be (B x K x M), + therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K) + in order to be able to call reference implementation that works with (B x M x K) input. + Similarly, if pretranspose_B is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */ + + // Define transposed shapes + TensorShape a_transposed_shape(a.shape()); + a_transposed_shape.set(0, a.shape().y()); + a_transposed_shape.set(1, a.shape().x()); + + TensorShape b_transposed_shape(b.shape()); + b_transposed_shape.set(0, b.shape().y()); + b_transposed_shape.set(1, b.shape().x()); + + // Define transposed tensors + SimpleTensor<T> a_transposed{ a_transposed_shape, data_type }; + SimpleTensor<T> b_transposed{ b_transposed_shape, data_type }; + + // pretranspose a if necessary + if(pretranspose_a) + { + a_transposed = reference::permute<T>(a, PermutationVector(1U, 0U)); + } + + // pretranspose b if necessary + if(pretranspose_b) + { + b_transposed = reference::permute<T>(b, PermutationVector(1U, 0U)); + } + + // Use transposed tensors if boolean enabled else use original tensors + SimpleTensor<T> result = gemm_reference<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c); + + // We reshape the gemm output back if the tensor is high dimensional + if(output_shape_collapsed != output_shape) + { + result = reference::reshape_layer(result, output_shape); + } + + return result; + } + + template <typename U = T> + typename std::enable_if < std::is_same<U, float>::value || std::is_same<U, half>::value, SimpleTensor<U >>::type gemm_reference(SimpleTensor<U> &a, SimpleTensor<U> &b, SimpleTensor<U> &c) + { + // Fill bias, then copy first dimension into subsequent dimensions to mimic broadcast + // of bias tensor from shape [dst.dimension(0)] to [dst.tensor_shape()] in target kernel + if(_enable_bias) + { + fill(c, _hash + 3); + const int n = c.shape().x(); + const int other_dims = c.shape().collapsed_from(1)[1]; + for(int i = 1; i < other_dims; ++i) // For all data, copy first n elements into remaining batches + { + memcpy(c.data() + i * n, c.data(), n * sizeof(T)); + } + } + // Setting beta to 0 will effectively disable C for the + // computation of the reference: alpha * A * B + 0 * C + return reference::gemm<U>(a, b, c, 1.0f, (_enable_bias) ? 1.0f : 0.f); + } + + template <typename U = T> + typename std::enable_if < std::is_same<U, int8_t>::value || std::is_same<U, uint8_t>::value, SimpleTensor<U >>::type gemm_reference(SimpleTensor<U> &a, SimpleTensor<U> &b, SimpleTensor<U> &c) + { + const UniformQuantizationInfo aq = a.quantization_info().uniform(); + const UniformQuantizationInfo bq = b.quantization_info().uniform(); + const UniformQuantizationInfo cq = c.quantization_info().uniform(); + + const SimpleTensor<int32_t> result = reference::gemmlowp_matrix_multiply_core<int32_t, U, U>(a, b, c.shape(), -aq.offset, -bq.offset); + + std::vector<int32_t> gemmlowp_multipliers{ 1 }; + std::vector<int32_t> gemmlowp_shifts{ 1 }; + const int gemmlowp_offset = cq.offset; + const float scale = aq.scale * bq.scale / cq.scale; + + quantization::calculate_quantized_multiplier(scale, &gemmlowp_multipliers[0], &gemmlowp_shifts[0]); + constexpr int32_t gemmlowp_min_bound = std::numeric_limits<int32_t>::min(); + constexpr int32_t gemmlowp_max_bound = std::numeric_limits<int32_t>::max(); + + SimpleTensor<int> bias{ c.shape(), DataType::S32 }; + if(_enable_bias) + { + // Identical to float implementation, fill and copy values of bias first dimension + fill_bias_s32(bias, _hash + 3, _min_bias, _max_bias); + const int n = bias.shape().x(); + const int other_dims = bias.shape().collapsed_from(1)[1]; + const unsigned int dt_size = sizeof(int32_t); + for(int i = 1; i < other_dims; ++i) + { + memcpy(bias.data() + i * n, bias.data(), n * dt_size); + } + } + else + { + fill_constant(bias, static_cast<int32_t>(0)); // effectively disable bias + } + + const SimpleTensor<U> final_result = reference::gemmlowp_quantize_down_scale_by_fixedpoint<int32_t, U>(result, bias, + gemmlowp_multipliers, gemmlowp_shifts, gemmlowp_offset, gemmlowp_min_bound, gemmlowp_max_bound); + + return final_result; + } + + CLTensor _target{}; + SimpleTensor<T> _reference{}; + bool _enable_bias{ false }; + bool _device_supports_export_to_cl_image{ true }; + bool _device_supports_mmul{ true }; + int32_t _min_bias{ 0 }; + int32_t _max_bias{ 0 }; + int32_t _hash{ 0 }; +}; + +template <typename T, typename KernelType, bool use_mmul = false> +class MatMulKernelValidationFixture : public MatMulKernelGenericValidationFixture<T, KernelType, use_mmul> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape output_shape, bool pretranspose_a, bool pretranspose_b, int M0, int N0, int K0, bool export_rhs_to_cl_image, DataType data_type) + { + MatMulKernelGenericValidationFixture<T, KernelType, use_mmul>::setup(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, M0, N0, K0, export_rhs_to_cl_image, data_type, + false /* enable bias */); + } +}; + +template <typename T, typename KernelType, bool use_mmul = false> +class MatMulKernelWithBiasValidation : public MatMulKernelGenericValidationFixture<T, KernelType, use_mmul> +{ +public: + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape output_shape, bool pretranspose_a, bool pretranspose_b, int M0, int N0, int K0, bool export_rhs_to_cl_image, DataType data_type) + { + MatMulKernelGenericValidationFixture<T, KernelType, use_mmul>::setup(shape_a, shape_b, output_shape, pretranspose_a, pretranspose_b, M0, N0, K0, export_rhs_to_cl_image, data_type, + true /* enable bias */); + } +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_MATMULKERNELFIXTURE_H diff --git a/tests/validation/fixtures/MaxUnpoolingLayerFixture.h b/tests/validation/fixtures/MaxUnpoolingLayerFixture.h index b2575aa26b..808e3ffabd 100644 --- a/tests/validation/fixtures/MaxUnpoolingLayerFixture.h +++ b/tests/validation/fixtures/MaxUnpoolingLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2021 Arm Limited. + * Copyright (c) 2020-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename PoolingFunctionTy class MaxUnpoolingLayerValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, PoolingLayerInfo pool_info, DataType data_type, DataLayout data_layout) { std::mt19937 gen(library->seed()); @@ -150,7 +149,6 @@ template <typename TensorType, typename AccessorType, typename F1, typename F2, class MaxUnpoolingLayerValidationFixture : public MaxUnpoolingLayerValidationGenericFixture<TensorType, AccessorType, F1, F2, T> { public: - template <typename...> void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, DataType data_type, DataLayout data_layout) { MaxUnpoolingLayerValidationGenericFixture<TensorType, AccessorType, F1, F2, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, data_layout, pad_stride_info, true), diff --git a/tests/validation/fixtures/MeanStdDevNormalizationLayerFixture.h b/tests/validation/fixtures/MeanStdDevNormalizationLayerFixture.h index 9868cd1abf..bf5d20790c 100644 --- a/tests/validation/fixtures/MeanStdDevNormalizationLayerFixture.h +++ b/tests/validation/fixtures/MeanStdDevNormalizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,30 +44,35 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class MeanStdDevNormalizationLayerValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape shape, DataType dt, bool in_place, float epsilon = 1e-8f) + void setup(TensorShape shape, DataType dt, bool in_place, float epsilon = 1e-8) { - _data_type = dt; - _target = compute_target(shape, dt, in_place, epsilon); - _reference = compute_reference(shape, dt, epsilon); + QuantizationInfo qi = QuantizationInfo(0.5f, 10); + _data_type = dt; + _target = compute_target(shape, dt, in_place, epsilon, qi); + _reference = compute_reference(shape, dt, epsilon, qi); } protected: template <typename U> - void fill(U &&src_tensor) + void fill(U &&tensor) { - static_assert(std::is_floating_point<T>::value || std::is_same<T, half>::value, "Only floating point data types supported."); - using DistributionType = typename std::conditional<std::is_same<T, half>::value, arm_compute::utils::uniform_real_distribution_16bit<T>, std::uniform_real_distribution<T>>::type; - - DistributionType distribution{ T(-1.0f), T(1.0f) }; - library->fill(src_tensor, distribution, 0); + if(is_data_type_float(_data_type)) + { + std::uniform_real_distribution<> distribution{ -1.0f, 1.0f }; + library->fill(tensor, distribution, 0); + } + else + { + std::uniform_int_distribution<> distribution{ 0, 255 }; + library->fill(tensor, distribution, 0); + } } - TensorType compute_target(TensorShape shape, DataType dt, bool in_place, float epsilon) + TensorType compute_target(TensorShape shape, DataType dt, bool in_place, float epsilon, QuantizationInfo qi) { // Create tensors - TensorType src = create_tensor<TensorType>(shape, dt, 1); - TensorType dst; + TensorType src = create_tensor<TensorType>(shape, dt, 1, qi); + TensorType dst = create_tensor<TensorType>(shape, dt, 1, qi); TensorType *dst_ptr = in_place ? &src : &dst; @@ -104,10 +109,10 @@ protected: } } - SimpleTensor<T> compute_reference(const TensorShape &shape, DataType dt, float epsilon) + SimpleTensor<T> compute_reference(const TensorShape &shape, DataType dt, float epsilon, QuantizationInfo qi) { // Create reference - SimpleTensor<T> ref_src{ shape, dt, 1 }; + SimpleTensor<T> ref_src{ shape, dt, 1, qi }; // Fill reference fill(ref_src); @@ -119,6 +124,7 @@ protected: SimpleTensor<T> _reference{}; DataType _data_type{}; }; + } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/fixtures/NonMaxSuppressionFixture.h b/tests/validation/fixtures/NonMaxSuppressionFixture.h index abfe909fc4..043b4731aa 100644 --- a/tests/validation/fixtures/NonMaxSuppressionFixture.h +++ b/tests/validation/fixtures/NonMaxSuppressionFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType> class NMSValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, unsigned int max_output_size, float score_threshold, float nms_threshold) { ARM_COMPUTE_ERROR_ON(max_output_size == 0); diff --git a/tests/validation/fixtures/NormalizationLayerFixture.h b/tests/validation/fixtures/NormalizationLayerFixture.h index 8a29e64826..ddaa3533f5 100644 --- a/tests/validation/fixtures/NormalizationLayerFixture.h +++ b/tests/validation/fixtures/NormalizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class NormalizationValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, NormType norm_type, int norm_size, float beta, bool is_scaled, DataType data_type, DataLayout data_layout) { NormalizationLayerInfo info(norm_type, norm_size, 5, beta, 1.f, is_scaled); @@ -119,7 +118,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class NormalizationValidationFixture : public NormalizationValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, NormType norm_type, int norm_size, float beta, bool is_scaled, DataType data_type, DataLayout data_layout) { NormalizationValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, norm_type, norm_size, beta, is_scaled, data_type, data_layout); diff --git a/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h b/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h index 4586afbb5d..5f2c865950 100644 --- a/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h +++ b/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class NormalizePlanarYUVLayerValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape0, TensorShape shape1, DataType dt, DataLayout data_layout, QuantizationInfo quantization_info) { _data_type = dt; @@ -144,7 +143,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class NormalizePlanarYUVLayerValidationFixture : public NormalizePlanarYUVLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape0, TensorShape shape1, DataType dt, DataLayout data_layout) { NormalizePlanarYUVLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, dt, data_layout, QuantizationInfo()); @@ -155,7 +153,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class NormalizePlanarYUVLayerValidationQuantizedFixture : public NormalizePlanarYUVLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape0, TensorShape shape1, DataType dt, DataLayout data_layout, QuantizationInfo quantization_info) { NormalizePlanarYUVLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, dt, data_layout, quantization_info); diff --git a/tests/validation/fixtures/PadLayerFixture.h b/tests/validation/fixtures/PadLayerFixture.h index 1951813df8..93b43616ff 100644 --- a/tests/validation/fixtures/PadLayerFixture.h +++ b/tests/validation/fixtures/PadLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PaddingFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, const PaddingList &padding, const PaddingMode mode) { PaddingList clamped_padding = padding; diff --git a/tests/validation/fixtures/PermuteFixture.h b/tests/validation/fixtures/PermuteFixture.h index 118971394a..b1b3845a8d 100644 --- a/tests/validation/fixtures/PermuteFixture.h +++ b/tests/validation/fixtures/PermuteFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PermuteValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, PermutationVector perm, DataType data_type) { _target = compute_target(input_shape, data_type, perm); diff --git a/tests/validation/fixtures/PixelWiseMultiplicationFixture.h b/tests/validation/fixtures/PixelWiseMultiplicationFixture.h index 7c643bd726..4345d8a13f 100644 --- a/tests/validation/fixtures/PixelWiseMultiplicationFixture.h +++ b/tests/validation/fixtures/PixelWiseMultiplicationFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationGenericValidationFixture : public framework::Fixture { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, @@ -76,9 +75,29 @@ protected: QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, ActivationLayerInfo act_info) { // Create tensors - TensorType src1 = create_tensor<TensorType>(shape0, dt_in1, 1, qinfo0); - TensorType src2 = create_tensor<TensorType>(shape1, dt_in2, 1, qinfo1); - TensorType dst = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), dt_out, 1, qinfo_out); + const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1); + TensorType src1 = create_tensor<TensorType>(shape0, dt_in1, 1, qinfo0); + TensorType src2 = create_tensor<TensorType>(shape1, dt_in2, 1, qinfo1); + TensorType dst = create_tensor<TensorType>(out_shape, dt_out, 1, qinfo_out); + + // Check whether do in-place computation and whether inputs are broadcast compatible + TensorType *actual_dst = &dst; + if(_is_inplace) + { + bool src1_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape0, 0) && (qinfo0 == qinfo_out) && (dt_in1 == dt_out); + bool src2_is_inplace = !arm_compute::detail::have_different_dimensions(out_shape, shape1, 0) && (qinfo1 == qinfo_out) && (dt_in2 == dt_out); + bool do_in_place = out_shape.total_size() != 0 && (src1_is_inplace || src2_is_inplace); + ARM_COMPUTE_ASSERT(do_in_place); + + if(src1_is_inplace) + { + actual_dst = &src1; + } + else + { + actual_dst = &src2; + } + } auto allocate_tensor = [](TensorType & t) { @@ -89,11 +108,12 @@ protected: // Create and configure function FunctionType multiply; - multiply.configure(&src1, &src2, (_is_inplace ? &src1 : &dst), scale, convert_policy, rounding_policy, act_info); + multiply.configure(&src1, &src2, actual_dst, scale, convert_policy, rounding_policy, act_info); allocate_tensor(src1); allocate_tensor(src2); + // If don't do in-place computation, still need to allocate original dst if(!_is_inplace) { allocate_tensor(dst); @@ -106,12 +126,7 @@ protected: // Compute function multiply.run(); - if(_is_inplace) - { - return src1; - } - - return dst; + return std::move(*actual_dst); } SimpleTensor<T3> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, DataType dt_out, @@ -122,16 +137,12 @@ protected: SimpleTensor<T1> src1{ shape0, dt_in1, 1, qinfo0 }; SimpleTensor<T2> src2{ shape1, dt_in2, 1, qinfo1 }; - // current in-place implementation only supports same metadata of input and output tensors. - // By ignoring output quantization information here, we can make test cases implementation much simpler. - QuantizationInfo output_qinfo = _is_inplace ? qinfo0 : qinfo_out; - // Fill reference fill(src1, 0); fill(src2, 1); - auto result = reference::pixel_wise_multiplication<T1, T2, T3>(src1, src2, scale, convert_policy, rounding_policy, dt_out, output_qinfo); - return act_info.enabled() ? reference::activation_layer(result, act_info, output_qinfo) : result; + auto result = reference::pixel_wise_multiplication<T1, T2, T3>(src1, src2, scale, convert_policy, rounding_policy, dt_out, qinfo_out); + return act_info.enabled() ? reference::activation_layer(result, act_info, qinfo_out) : result; } TensorType _target{}; @@ -143,7 +154,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationValidationFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2, T3> { public: - template <typename...> void setup(const TensorShape &shape, DataType dt_in1, DataType dt_in2, DataType dt_out, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, bool is_inplace) { PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2, T3>::setup(shape, shape, dt_in1, dt_in2, dt_out, scale, convert_policy, rounding_policy, @@ -155,7 +165,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationBroadcastValidationFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2, T3> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, DataType dt_out, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, bool is_inplace) { @@ -168,7 +177,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationValidationFloatFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> void setup(const TensorShape &shape, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, ActivationLayerInfo act_info, bool is_inplace) { PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, shape, dt_in1, dt_in2, dt_in2, scale, convert_policy, rounding_policy, @@ -180,7 +188,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationValidationIntegerFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> void setup(const TensorShape &shape, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, ActivationLayerInfo act_info, bool is_inplace) { PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, shape, dt_in1, dt_in2, dt_in2, scale, convert_policy, rounding_policy, @@ -192,7 +199,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationBroadcastValidationFloatFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, ActivationLayerInfo act_info, bool is_inplace) { @@ -205,7 +211,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationValidationQuantizedFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2, T3> { public: - template <typename...> void setup(const TensorShape &shape, DataType dt_in1, DataType dt_in2, DataType dt_out, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { @@ -218,7 +223,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PixelWiseMultiplicationBroadcastValidationQuantizedFixture : public PixelWiseMultiplicationGenericValidationFixture<TensorType, AccessorType, FunctionType, T1, T2, T3> { public: - template <typename...> void setup(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, DataType dt_out, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, QuantizationInfo qinfo0, QuantizationInfo qinfo1, QuantizationInfo qinfo_out, bool is_inplace) { diff --git a/tests/validation/fixtures/Pooling3dLayerFixture.h b/tests/validation/fixtures/Pooling3dLayerFixture.h new file mode 100644 index 0000000000..1bdf615fb1 --- /dev/null +++ b/tests/validation/fixtures/Pooling3dLayerFixture.h @@ -0,0 +1,164 @@ +/* + * Copyright (c) 2022-2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_TEST_POOLING_3D_LAYER_FIXTURE +#define ARM_COMPUTE_TEST_POOLING_3D_LAYER_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/Tensor.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/Pooling3dLayer.h" +#include <random> +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class Pooling3dLayerValidationGenericFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape, Pooling3dLayerInfo pool_info, DataType data_type, QuantizationInfo input_qinfo = QuantizationInfo(), QuantizationInfo output_qinfo = QuantizationInfo()) + { + _target = compute_target(shape, pool_info, data_type, input_qinfo, output_qinfo); + _reference = compute_reference(shape, pool_info, data_type, input_qinfo, output_qinfo); + } + +protected: + template <typename U> + void fill(U &&tensor) + { + if(tensor.data_type() == DataType::F32) + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, 0); + } + else if(tensor.data_type() == DataType::F16) + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f }; + library->fill(tensor, distribution, 0); + } + else // data type is quantized_asymmetric + { + library->fill_tensor_uniform(tensor, 0); + } + } + + TensorType compute_target(TensorShape shape, Pooling3dLayerInfo info, + DataType data_type, QuantizationInfo input_qinfo, QuantizationInfo output_qinfo) + { + // Create tensors + TensorType src = create_tensor<TensorType>(shape, data_type, 1, input_qinfo, DataLayout::NDHWC); + const TensorShape dst_shape = misc::shape_calculator::compute_pool3d_shape((src.info()->tensor_shape()), info); + TensorType dst = create_tensor<TensorType>(dst_shape, data_type, 1, output_qinfo, DataLayout::NDHWC); + + // Create and configure function + FunctionType pool_layer; + pool_layer.validate(src.info(), dst.info(), info); + pool_layer.configure(&src, &dst, info); + + ARM_COMPUTE_ASSERT(src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + // Allocate tensors + src.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(src)); + + // Compute function + pool_layer.run(); + return dst; + } + + SimpleTensor<T> compute_reference(TensorShape shape, Pooling3dLayerInfo info, DataType data_type, QuantizationInfo input_qinfo, QuantizationInfo output_qinfo) + { + // Create reference + SimpleTensor<T> src(shape, data_type, 1, input_qinfo, DataLayout::NDHWC); + // Fill reference + fill(src); + return reference::pooling_3d_layer<T>(src, info, output_qinfo); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class Pooling3dLayerValidationFixture : public Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, PoolingType pool_type, Size3D pool_size, Size3D stride, Padding3D padding, bool exclude_padding, DataType data_type) + { + Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, Pooling3dLayerInfo(pool_type, pool_size, stride, padding, exclude_padding), + data_type); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class Pooling3dLayerValidationQuantizedFixture : public Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, PoolingType pool_type, Size3D pool_size, Size3D stride, Padding3D padding, bool exclude_padding, DataType data_type, + QuantizationInfo input_qinfo = QuantizationInfo(), QuantizationInfo output_qinfo = QuantizationInfo()) + { + Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, Pooling3dLayerInfo(pool_type, pool_size, stride, padding, exclude_padding), + data_type, input_qinfo, output_qinfo); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class Pooling3dLayerGlobalValidationFixture : public Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, PoolingType pool_type, DataType data_type) + { + Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, Pooling3dLayerInfo(pool_type), data_type); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class SpecialPooling3dLayerValidationFixture : public Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape src_shape, Pooling3dLayerInfo pool_info, DataType data_type) + { + Pooling3dLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, pool_info, data_type); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_POOLING_3D_LAYER_FIXTURE */ diff --git a/tests/validation/fixtures/PoolingLayerFixture.h b/tests/validation/fixtures/PoolingLayerFixture.h index ec4e9f80dd..59c920868b 100644 --- a/tests/validation/fixtures/PoolingLayerFixture.h +++ b/tests/validation/fixtures/PoolingLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PoolingLayerValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, PoolingLayerInfo pool_info, DataType data_type, DataLayout data_layout, bool indices = false, QuantizationInfo input_qinfo = QuantizationInfo(), QuantizationInfo output_qinfo = QuantizationInfo(), bool mixed_layout = false) { @@ -161,10 +160,10 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PoolingLayerIndicesValidationFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> - void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, DataLayout data_layout) + void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, DataLayout data_layout, bool use_kernel_indices) { - PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, data_layout, pad_stride_info, exclude_padding), + PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, data_layout, pad_stride_info, exclude_padding, false, + true, use_kernel_indices), data_type, data_layout, true); } }; @@ -173,7 +172,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PoolingLayerValidationFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, DataLayout data_layout) { PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, data_layout, pad_stride_info, exclude_padding), @@ -185,7 +183,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PoolingLayerValidationMixedPrecisionFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, DataLayout data_layout, bool fp_mixed_precision = false) { PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, data_layout, pad_stride_info, exclude_padding, fp_mixed_precision), @@ -197,7 +194,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PoolingLayerValidationQuantizedFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, DataLayout data_layout = DataLayout::NCHW, QuantizationInfo input_qinfo = QuantizationInfo(), QuantizationInfo output_qinfo = QuantizationInfo()) { @@ -210,10 +206,9 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SpecialPoolingLayerValidationFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape src_shape, PoolingLayerInfo pool_info, DataType data_type) { - PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, pool_info, data_type, DataLayout::NCHW); + PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, pool_info, data_type, pool_info.data_layout); } }; @@ -221,7 +216,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class GlobalPoolingLayerValidationFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, PoolingType pool_type, DataType data_type, DataLayout data_layout = DataLayout::NCHW) { PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, data_layout), data_type, data_layout); diff --git a/tests/validation/fixtures/PriorBoxLayerFixture.h b/tests/validation/fixtures/PriorBoxLayerFixture.h index 40a9f5d976..0a76cfd155 100644 --- a/tests/validation/fixtures/PriorBoxLayerFixture.h +++ b/tests/validation/fixtures/PriorBoxLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class PriorBoxLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, PriorBoxLayerInfo info, DataType data_type, DataLayout data_layout) { TensorInfo input_info(input_shape, 1, data_type); diff --git a/tests/validation/fixtures/QLSTMLayerNormalizationFixture.h b/tests/validation/fixtures/QLSTMLayerNormalizationFixture.h index b496e4abad..e864b4affe 100644 --- a/tests/validation/fixtures/QLSTMLayerNormalizationFixture.h +++ b/tests/validation/fixtures/QLSTMLayerNormalizationFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2020-2021 Arm Limited. + * Copyright (c) 2020-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class QLSTMLayerNormalizationValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weight_shape, TensorShape bias_shape, DataType data_type, QuantizationInfo weight_qinfo) { ARM_COMPUTE_ERROR_ON(data_type != DataType::QSYMM16); diff --git a/tests/validation/fixtures/QuantizationLayerFixture.h b/tests/validation/fixtures/QuantizationLayerFixture.h index 8e8b920a9a..1b21967bda 100644 --- a/tests/validation/fixtures/QuantizationLayerFixture.h +++ b/tests/validation/fixtures/QuantizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class QuantizationValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type_in, DataType data_type_out, QuantizationInfo qinfo, QuantizationInfo qinfo_in) { _target = compute_target(shape, data_type_in, data_type_out, qinfo, qinfo_in); @@ -108,7 +107,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class QuantizationValidationFixture : public QuantizationValidationGenericFixture<TensorType, AccessorType, FunctionType, Tin, Tout> { public: - template <typename...> void setup(TensorShape shape, DataType data_type_in, DataType data_type_out, QuantizationInfo qinfo) { QuantizationValidationGenericFixture<TensorType, AccessorType, FunctionType, Tin, Tout>::setup(shape, data_type_in, data_type_out, qinfo, QuantizationInfo()); diff --git a/tests/validation/fixtures/RNNLayerFixture.h b/tests/validation/fixtures/RNNLayerFixture.h index 983cd43398..e9a05e7838 100644 --- a/tests/validation/fixtures/RNNLayerFixture.h +++ b/tests/validation/fixtures/RNNLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -42,7 +42,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class RNNLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape recurrent_weights_shape, TensorShape bias_shape, TensorShape output_shape, ActivationLayerInfo info, DataType data_type) { diff --git a/tests/validation/fixtures/ROIAlignLayerFixture.h b/tests/validation/fixtures/ROIAlignLayerFixture.h index 4c349bcd47..ad76dcbbd9 100644 --- a/tests/validation/fixtures/ROIAlignLayerFixture.h +++ b/tests/validation/fixtures/ROIAlignLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ROIAlignLayerGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout, QuantizationInfo qinfo, QuantizationInfo output_qinfo) { _rois_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::QASYMM16 : data_type; @@ -189,7 +188,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ROIAlignLayerFixture : public ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T, TRois> { public: - template <typename...> void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout) { ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T, TRois>::setup(input_shape, pool_info, rois_shape, data_type, data_layout, @@ -201,7 +199,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ROIAlignLayerQuantizedFixture : public ROIAlignLayerGenericFixture<TensorType, AccessorType, FunctionType, T, TRois> { public: - template <typename...> void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout, QuantizationInfo qinfo, QuantizationInfo output_qinfo) { diff --git a/tests/validation/fixtures/ROIPoolingLayerFixture.h b/tests/validation/fixtures/ROIPoolingLayerFixture.h index 763c5a7d76..4b46a6176d 100644 --- a/tests/validation/fixtures/ROIPoolingLayerFixture.h +++ b/tests/validation/fixtures/ROIPoolingLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2021 Arm Limited. + * Copyright (c) 2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ROIPoolingLayerGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout, QuantizationInfo qinfo, QuantizationInfo output_qinfo) { _target = compute_target(input_shape, data_type, data_layout, pool_info, rois_shape, qinfo, output_qinfo); @@ -174,7 +173,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ROIPoolingLayerQuantizedFixture : public ROIPoolingLayerGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout, QuantizationInfo qinfo, QuantizationInfo output_qinfo) { @@ -187,7 +185,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ROIPoolingLayerFixture : public ROIPoolingLayerGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, const ROIPoolingLayerInfo pool_info, TensorShape rois_shape, DataType data_type, DataLayout data_layout) { ROIPoolingLayerGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, pool_info, rois_shape, data_type, data_layout, diff --git a/tests/validation/fixtures/RangeFixture.h b/tests/validation/fixtures/RangeFixture.h index 9265e299b6..166613a318 100644 --- a/tests/validation/fixtures/RangeFixture.h +++ b/tests/validation/fixtures/RangeFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -55,7 +55,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class RangeFixture : public framework::Fixture { public: - template <typename...> void setup(const DataType data_type0, float start, float step, const QuantizationInfo qinfo0 = QuantizationInfo()) { _target = compute_target(data_type0, qinfo0, start, step); diff --git a/tests/validation/fixtures/ReduceMeanFixture.h b/tests/validation/fixtures/ReduceMeanFixture.h index 304630e9f5..e61941435c 100644 --- a/tests/validation/fixtures/ReduceMeanFixture.h +++ b/tests/validation/fixtures/ReduceMeanFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReduceMeanValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, Coordinates axis, bool keep_dims, QuantizationInfo quantization_info_input, QuantizationInfo quantization_info_output) { _target = compute_target(shape, data_type, axis, keep_dims, quantization_info_input, quantization_info_output); @@ -124,7 +123,13 @@ protected: { TensorShape output_shape = i == 0 ? src_shape : out.shape(); output_shape.set(axis[i], 1); - out = reference::reduction_operation<T, T>(i == 0 ? src : out, output_shape, axis[i], ReductionOperation::MEAN_SUM, quantization_info_output); + bool is_opencl = false; + +#ifdef ARM_COMPUTE_OPENCL_ENABLED + is_opencl = std::is_same<CLTensor, TensorType>::value; // Round down to zero on opencl to match kernel +#endif /* ARM_COMPUTE_OPENCL_ENABLED */ + out = reference::reduction_operation<T, T>(i == 0 ? src : out, output_shape, axis[i], ReductionOperation::MEAN_SUM, data_type, quantization_info_output, + is_opencl ? RoundingPolicy::TO_ZERO : RoundingPolicy::TO_NEAREST_UP); } if(!keep_dims) @@ -133,7 +138,7 @@ protected: std::sort(axis.begin(), axis.begin() + axis.num_dimensions()); for(unsigned int i = 0; i < axis.num_dimensions(); ++i) { - output_shape.remove_dimension(axis[i] - i); + output_shape.remove_dimension(axis[i] - i, false); } out = reference::reshape_layer(out, output_shape); @@ -149,7 +154,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReduceMeanQuantizedFixture : public ReduceMeanValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, Coordinates axis, bool keep_dims, QuantizationInfo quantization_info_input, QuantizationInfo quantization_info_output) { ReduceMeanValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, keep_dims, quantization_info_input, quantization_info_output); @@ -160,7 +164,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReduceMeanFixture : public ReduceMeanValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, Coordinates axis, bool keep_dims) { ReduceMeanValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, keep_dims, QuantizationInfo(), QuantizationInfo()); diff --git a/tests/validation/fixtures/ReductionOperationFixture.h b/tests/validation/fixtures/ReductionOperationFixture.h index fc422ad35b..b44f299486 100644 --- a/tests/validation/fixtures/ReductionOperationFixture.h +++ b/tests/validation/fixtures/ReductionOperationFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReductionOperationValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info, bool keep_dims = false) { const bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MAX) || (op == ReductionOperation::ARG_IDX_MIN); @@ -76,14 +75,14 @@ protected: if(tensor.data_type() == DataType::QASYMM8) { std::pair<int, int> bounds = get_quantized_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<uint8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<uint32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, 0); } else if(tensor.data_type() == DataType::QASYMM8_SIGNED) { std::pair<int, int> bounds = get_quantized_qasymm8_signed_bounds(tensor.quantization_info(), -1.0f, 1.0f); - std::uniform_int_distribution<int8_t> distribution(bounds.first, bounds.second); + std::uniform_int_distribution<int32_t> distribution(bounds.first, bounds.second); library->fill(tensor, distribution, 0); } @@ -135,7 +134,7 @@ protected: // Fill reference fill(src); - return reference::reduction_operation<T, T>(src, dst_shape, axis, op, quantization_info); + return reference::reduction_operation<T, T>(src, dst_shape, axis, op, data_type, quantization_info); } TensorType _target{}; @@ -149,7 +148,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReductionOperationQuantizedFixture : public ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info = QuantizationInfo(), bool keep_dims = false) { ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, op, quantization_info, keep_dims); @@ -160,7 +158,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReductionOperationFixture : public ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, unsigned int axis, ReductionOperation op, bool keep_dims = false) { ReductionOperationValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, axis, op, QuantizationInfo(), keep_dims); diff --git a/tests/validation/fixtures/RemapFixture.h b/tests/validation/fixtures/RemapFixture.h deleted file mode 100644 index 2cb8e67f62..0000000000 --- a/tests/validation/fixtures/RemapFixture.h +++ /dev/null @@ -1,168 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#ifndef ARM_COMPUTE_TEST_REMAP_FIXTURE -#define ARM_COMPUTE_TEST_REMAP_FIXTURE - -#include "arm_compute/core/TensorShape.h" -#include "arm_compute/core/Types.h" -#include "tests/AssetsLibrary.h" -#include "tests/Globals.h" -#include "tests/IAccessor.h" -#include "tests/framework/Asserts.h" -#include "tests/framework/Fixture.h" -#include "tests/validation/reference/Remap.h" - -#include <random> - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class RemapValidationGenericFixture : public framework::Fixture -{ -public: - template <typename...> - void setup(TensorShape shape, InterpolationPolicy policy, DataType data_type, BorderMode border_mode, DataLayout data_layout = DataLayout::NCHW) - { - std::mt19937 gen(library->seed()); - std::uniform_int_distribution<uint8_t> distribution(0, 255); - const T constant_border_value = static_cast<T>(distribution(gen)); - - _data_layout = data_layout; - _target = compute_target(shape, policy, data_type, border_mode, constant_border_value); - _reference = compute_reference(shape, policy, data_type, border_mode, constant_border_value); - } - -protected: - template <typename U> - void fill(U &&tensor, int i, float min, float max) - { - std::uniform_int_distribution<> distribution((int)min, (int)max); - library->fill(tensor, distribution, i); - } - - TensorType compute_target(TensorShape shape, InterpolationPolicy policy, DataType data_type, BorderMode border_mode, T constant_border_value) - { - if(_data_layout == DataLayout::NHWC) - { - permute(shape, PermutationVector(2U, 0U, 1U)); - } - - // Create tensors - TensorType src = create_tensor<TensorType>(shape, data_type, 1, QuantizationInfo(), _data_layout); - TensorType map_x = create_tensor<TensorType>(shape, DataType::F32, 1, QuantizationInfo(), _data_layout); - TensorType map_y = create_tensor<TensorType>(shape, DataType::F32, 1, QuantizationInfo(), _data_layout); - TensorType dst = create_tensor<TensorType>(shape, data_type, 1, QuantizationInfo(), _data_layout); - - // Create and configure function - FunctionType remap; - remap.configure(&src, &map_x, &map_y, &dst, policy, border_mode, constant_border_value); - - ARM_COMPUTE_ASSERT(src.info()->is_resizable()); - ARM_COMPUTE_ASSERT(map_x.info()->is_resizable()); - ARM_COMPUTE_ASSERT(map_y.info()->is_resizable()); - ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); - - // Allocate tensors - src.allocator()->allocate(); - map_x.allocator()->allocate(); - map_y.allocator()->allocate(); - dst.allocator()->allocate(); - - ARM_COMPUTE_ASSERT(!src.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!map_x.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!map_y.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); - - // Fill tensors - int max_val = std::max({ shape.x(), shape.y(), shape.z() }); - - fill(AccessorType(src), 0, 0, 255); - fill(AccessorType(map_x), 1, -5, max_val); - fill(AccessorType(map_y), 2, -5, max_val); - - // Compute function - remap.run(); - - return dst; - } - - SimpleTensor<T> compute_reference(const TensorShape shape, InterpolationPolicy policy, DataType data_type, BorderMode border_mode, T constant_border_value) - { - ARM_COMPUTE_ERROR_ON(data_type != DataType::U8); - - // Create reference - SimpleTensor<T> src{ shape, data_type }; - SimpleTensor<float> map_x{ shape, DataType::F32 }; - SimpleTensor<float> map_y{ shape, DataType::F32 }; - - // Create the valid mask Tensor - _valid_mask = SimpleTensor<T> { shape, data_type }; - - // Fill reference - int max_val = std::max({ shape.x(), shape.y(), shape.z() }); - - fill(src, 0, 0, 255); - fill(map_x, 1, -5, max_val); - fill(map_y, 2, -5, max_val); - - // Compute reference - return reference::remap<T>(src, map_x, map_y, _valid_mask, policy, border_mode, constant_border_value); - } - - TensorType _target{}; - SimpleTensor<T> _reference{}; - SimpleTensor<T> _valid_mask{}; - DataLayout _data_layout{}; -}; - -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class RemapValidationFixture : public RemapValidationGenericFixture<TensorType, AccessorType, FunctionType, T> -{ -public: - template <typename...> - void setup(TensorShape shape, InterpolationPolicy policy, DataType data_type, BorderMode border_mode) - { - RemapValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, policy, data_type, border_mode); - } -}; - -template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class RemapValidationMixedLayoutFixture : public RemapValidationGenericFixture<TensorType, AccessorType, FunctionType, T> -{ -public: - template <typename...> - void setup(TensorShape shape, InterpolationPolicy policy, DataType data_type, BorderMode border_mode, DataLayout data_layout) - { - RemapValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, policy, data_type, border_mode, data_layout); - } -}; - -} // namespace validation -} // namespace test -} // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_REMAP_FIXTURE */ diff --git a/tests/validation/fixtures/ReorderFixture.h b/tests/validation/fixtures/ReorderFixture.h new file mode 100644 index 0000000000..8e28484c48 --- /dev/null +++ b/tests/validation/fixtures/ReorderFixture.h @@ -0,0 +1,123 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE_H + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/Reorder.h" +#include "src/core/NEON/kernels/arm_gemm/utils.hpp" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/** [ReorderLayer fixture] **/ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ReorderValidationFixture : public framework::Fixture +{ +public: + void check_hardware_supports(WeightFormat output_wf){ + if(!Scheduler::get().cpu_info().has_sve() && output_wf!=WeightFormat::OHWIo4){ + _hardware_supports = false; + } + if (Scheduler::get().cpu_info().has_sve() && arm_gemm::utils::get_vector_length<float>() != 8 && output_wf==WeightFormat::OHWIo8) + { + _hardware_supports = false; + } + } + + void setup(TensorShape input_shape, TensorShape output_shape, WeightFormat input_wf, WeightFormat output_wf, DataType data_type) + { + check_hardware_supports(output_wf); + if (_hardware_supports){ + _target = compute_target(input_shape, output_shape, input_wf, output_wf, data_type); + _reference = compute_reference(input_shape, output_shape, output_wf, data_type); + } + } + + protected: + template <typename U> + void fill(U &&tensor) + { + library->fill_tensor_uniform(tensor, 0); + } + + TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, WeightFormat input_wf, WeightFormat output_wf, DataType data_type) + { + // Create tensors + TensorType src = create_tensor<TensorType>(input_shape, data_type); + TensorType dst = create_tensor<TensorType>(output_shape, data_type); + + // Create and configure function + FunctionType reorder; + + reorder.configure(&src, &dst, input_wf, output_wf); + + ARM_COMPUTE_ASSERT(src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + // Allocate tensors + src.allocator()->allocate(); + dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + + // Fill tensors + fill(AccessorType(src)); + + // Compute function + reorder.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, WeightFormat output_wf, DataType data_type) + { + // Create reference + SimpleTensor<T> src{ input_shape, data_type }; + + // Fill reference + fill(src); + + return reference::reorder_layer<T>(src, output_shape, output_wf); + } + + bool _hardware_supports = true; + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; +/** [ReorderLayer fixture] **/ +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE_H diff --git a/tests/validation/fixtures/ReorgLayerFixture.h b/tests/validation/fixtures/ReorgLayerFixture.h index 76a486e3d7..f87017190e 100644 --- a/tests/validation/fixtures/ReorgLayerFixture.h +++ b/tests/validation/fixtures/ReorgLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReorgLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, int32_t stride, DataType data_type, DataLayout data_layout) { _target = compute_target(input_shape, stride, data_type, data_layout); diff --git a/tests/validation/fixtures/ReshapeLayerFixture.h b/tests/validation/fixtures/ReshapeLayerFixture.h index 8a98379ef2..5be431f8cf 100644 --- a/tests/validation/fixtures/ReshapeLayerFixture.h +++ b/tests/validation/fixtures/ReshapeLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_RESHAPE_LAYER_FIXTURE -#define ARM_COMPUTE_TEST_RESHAPE_LAYER_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_RESHAPELAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_RESHAPELAYERFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -31,6 +31,7 @@ #include "tests/IAccessor.h" #include "tests/framework/Asserts.h" #include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" #include "tests/validation/reference/ReshapeLayer.h" namespace arm_compute @@ -41,13 +42,12 @@ namespace validation { /** [ReshapeLayer fixture] **/ template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class ReshapeLayerValidationFixture : public framework::Fixture +class ReshapeLayerGenericValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape input_shape, TensorShape output_shape, DataType data_type) + void setup(TensorShape input_shape, TensorShape output_shape, DataType data_type, bool add_x_padding = false) { - _target = compute_target(input_shape, output_shape, data_type); + _target = compute_target(input_shape, output_shape, data_type, add_x_padding); _reference = compute_reference(input_shape, output_shape, data_type); } @@ -58,7 +58,7 @@ protected: library->fill_tensor_uniform(tensor, i); } - TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type) + TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type, bool add_x_padding = false) { // Check if indeed the input shape can be reshape to the output one ARM_COMPUTE_ASSERT(input_shape.total_size() == output_shape.total_size()); @@ -75,6 +75,12 @@ protected: ARM_COMPUTE_ASSERT(src.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + if(add_x_padding) + { + // Add random padding in x dimension + add_padding_x({ &src, &dst }); + } + // Allocate tensors src.allocator()->allocate(); dst.allocator()->allocate(); @@ -105,8 +111,27 @@ protected: TensorType _target{}; SimpleTensor<T> _reference{}; }; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ReshapeLayerValidationFixture : public ReshapeLayerGenericValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, TensorShape output_shape, DataType data_type) + { + ReshapeLayerGenericValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, output_shape, data_type); + } +}; +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ReshapeLayerPaddedValidationFixture : public ReshapeLayerGenericValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, TensorShape output_shape, DataType data_type) + { + ReshapeLayerGenericValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, output_shape, data_type, true /* add_x_padding */); + } +}; /** [ReshapeLayer fixture] **/ } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_RESHAPE_LAYER_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_RESHAPELAYERFIXTURE_H diff --git a/tests/validation/fixtures/ReverseFixture.h b/tests/validation/fixtures/ReverseFixture.h index 9d047a0067..856bff7b12 100644 --- a/tests/validation/fixtures/ReverseFixture.h +++ b/tests/validation/fixtures/ReverseFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_REVERSE_FIXTURE -#define ARM_COMPUTE_TEST_REVERSE_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_REVERSEFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_REVERSEFIXTURE_H #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorShape.h" @@ -45,11 +45,11 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ReverseValidationFixture : public framework::Fixture { public: - template <typename...> - void setup(TensorShape shape, TensorShape axis_shape, DataType data_type) + void setup(TensorShape shape, TensorShape axis_shape, DataType data_type, bool use_negative_axis = false, bool use_inverted_axis = false) { - _target = compute_target(shape, axis_shape, data_type); - _reference = compute_reference(shape, axis_shape, data_type); + _num_dims = shape.num_dimensions(); + _target = compute_target(shape, axis_shape, data_type, use_negative_axis, use_inverted_axis); + _reference = compute_reference(shape, axis_shape, data_type, use_negative_axis, use_inverted_axis); } protected: @@ -58,16 +58,25 @@ protected: { library->fill_tensor_uniform(tensor, 0); } - std::vector<int> generate_random_axis() + std::vector<int32_t> generate_random_axis(bool use_negative = false) { - std::vector<int> axis_v = { 0, 1, 2, 3 }; - std::mt19937 g(0); + std::vector<int32_t> axis_v; + if(use_negative) + { + axis_v = { -1, -2, -3, -4 }; + } + else + { + axis_v = { 0, 1, 2, 3 }; + } + axis_v = std::vector<int32_t>(axis_v.begin(), axis_v.begin() + _num_dims); + std::mt19937 g(library->seed()); std::shuffle(axis_v.begin(), axis_v.end(), g); return axis_v; } - TensorType compute_target(const TensorShape &shape, const TensorShape &axis_shape, DataType data_type) + TensorType compute_target(const TensorShape &shape, const TensorShape &axis_shape, DataType data_type, bool use_negative_axis, bool use_inverted_axis = false) { // Create tensors TensorType src = create_tensor<TensorType>(shape, data_type, 1); @@ -76,7 +85,7 @@ protected: // Create and configure function FunctionType reverse_func; - reverse_func.configure(&src, &dst, &axis); + reverse_func.configure(&src, &dst, &axis, use_inverted_axis); ARM_COMPUTE_ASSERT(src.info()->is_resizable()); ARM_COMPUTE_ASSERT(axis.info()->is_resizable()); @@ -95,8 +104,8 @@ protected: fill(AccessorType(src)); { auto axis_data = AccessorType(axis); - auto axis_v = generate_random_axis(); - std::copy(axis_v.begin(), axis_v.begin() + axis_shape.x(), static_cast<int32_t *>(axis_data.data())); + auto axis_v = generate_random_axis(use_negative_axis); + std::copy(axis_v.begin(), axis_v.begin() + axis_shape.total_size(), static_cast<int32_t *>(axis_data.data())); } // Compute function @@ -105,24 +114,25 @@ protected: return dst; } - SimpleTensor<T> compute_reference(const TensorShape &shape, const TensorShape &axis_shape, DataType data_type) + SimpleTensor<T> compute_reference(const TensorShape &shape, const TensorShape &axis_shape, DataType data_type, bool use_negative_axis, bool use_inverted_axis = false) { // Create reference - SimpleTensor<T> src{ shape, data_type }; - SimpleTensor<uint32_t> axis{ axis_shape, DataType::U32 }; + SimpleTensor<T> src{ shape, data_type }; + SimpleTensor<int32_t> axis{ axis_shape, DataType::S32 }; // Fill reference fill(src); - auto axis_v = generate_random_axis(); - std::copy(axis_v.begin(), axis_v.begin() + axis_shape.x(), axis.data()); + auto axis_v = generate_random_axis(use_negative_axis); + std::copy(axis_v.begin(), axis_v.begin() + axis_shape.total_size(), axis.data()); - return reference::reverse<T>(src, axis); + return reference::reverse<T>(src, axis, use_inverted_axis); } TensorType _target{}; SimpleTensor<T> _reference{}; + unsigned int _num_dims{}; }; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_REVERSE_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_REVERSEFIXTURE_H diff --git a/tests/validation/fixtures/ScaleFixture.h b/tests/validation/fixtures/ScaleFixture.h index b719a22fdf..86d89d71f7 100644 --- a/tests/validation/fixtures/ScaleFixture.h +++ b/tests/validation/fixtures/ScaleFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,15 +21,10 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_SCALE_FIXTURE -#define ARM_COMPUTE_TEST_SCALE_FIXTURE - -#include "arm_compute/core/TensorShape.h" -#include "arm_compute/core/Types.h" -#include "tests/AssetsLibrary.h" -#include "tests/Globals.h" -#include "tests/IAccessor.h" -#include "tests/framework/Asserts.h" +#ifndef ACL_TESTS_VALIDATION_FIXTURES_SCALEFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_SCALEFIXTURE_H + +#include "tests/framework/Asserts.h" // Required for ARM_COMPUTE_ASSERT #include "tests/framework/Fixture.h" #include "tests/validation/reference/Permute.h" #include "tests/validation/reference/Scale.h" @@ -44,23 +39,23 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ScaleValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout, InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy, - bool align_corners, bool mixed_layout) + bool align_corners, bool mixed_layout, QuantizationInfo output_quantization_info) { - _shape = shape; - _policy = policy; - _border_mode = border_mode; - _sampling_policy = sampling_policy; - _data_type = data_type; - _quantization_info = quantization_info; - _align_corners = align_corners; - _mixed_layout = mixed_layout; + _shape = shape; + _policy = policy; + _border_mode = border_mode; + _sampling_policy = sampling_policy; + _data_type = data_type; + _input_quantization_info = quantization_info; + _output_quantization_info = output_quantization_info; + _align_corners = align_corners; + _mixed_layout = mixed_layout; generate_scale(shape); - std::mt19937 generator(library->seed()); - std::uniform_int_distribution<uint8_t> distribution_u8(0, 255); + std::mt19937 generator(library->seed()); + std::uniform_int_distribution<uint32_t> distribution_u8(0, 255); _constant_border_value = static_cast<T>(distribution_u8(generator)); _target = compute_target(shape, data_layout); @@ -144,7 +139,7 @@ protected: } // Create tensors - TensorType src = create_tensor<TensorType>(shape, _data_type, 1, _quantization_info, data_layout); + TensorType src = create_tensor<TensorType>(shape, _data_type, 1, _input_quantization_info, data_layout); const int idx_width = get_data_layout_dimension_index(data_layout, DataLayoutDimension::WIDTH); const int idx_height = get_data_layout_dimension_index(data_layout, DataLayoutDimension::HEIGHT); @@ -152,7 +147,7 @@ protected: TensorShape shape_scaled(shape); shape_scaled.set(idx_width, shape[idx_width] * _scale_x, /* apply_dim_correction = */ false); shape_scaled.set(idx_height, shape[idx_height] * _scale_y, /* apply_dim_correction = */ false); - TensorType dst = create_tensor<TensorType>(shape_scaled, _data_type, 1, _quantization_info, data_layout); + TensorType dst = create_tensor<TensorType>(shape_scaled, _data_type, 1, _output_quantization_info, data_layout); // Create and configure function FunctionType scale; @@ -188,12 +183,12 @@ protected: SimpleTensor<T> compute_reference(const TensorShape &shape) { // Create reference - SimpleTensor<T> src{ shape, _data_type, 1, _quantization_info }; + SimpleTensor<T> src{ shape, _data_type, 1, _input_quantization_info }; // Fill reference fill(src); - return reference::scale<T>(src, _scale_x, _scale_y, _policy, _border_mode, _constant_border_value, _sampling_policy, /* ceil_policy_scale */ false, _align_corners); + return reference::scale<T>(src, _scale_x, _scale_y, _policy, _border_mode, _constant_border_value, _sampling_policy, /* ceil_policy_scale */ false, _align_corners, _output_quantization_info); } TensorType _target{}; @@ -204,7 +199,8 @@ protected: T _constant_border_value{}; SamplingPolicy _sampling_policy{}; DataType _data_type{}; - QuantizationInfo _quantization_info{}; + QuantizationInfo _input_quantization_info{}; + QuantizationInfo _output_quantization_info{}; bool _align_corners{ false }; bool _mixed_layout{ false }; float _scale_x{ 1.f }; @@ -215,7 +211,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ScaleValidationQuantizedFixture : public ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, DataLayout data_layout, InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy, bool align_corners) { @@ -227,14 +222,34 @@ public: border_mode, sampling_policy, align_corners, - mixed_layout); + mixed_layout, + quantization_info); + } +}; +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> +class ScaleValidationDifferentOutputQuantizedFixture : public ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, DataType data_type, QuantizationInfo input_quantization_info, QuantizationInfo output_quantization_info, DataLayout data_layout, InterpolationPolicy policy, + BorderMode border_mode, SamplingPolicy sampling_policy, + bool align_corners) + { + ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, + data_type, + input_quantization_info, + data_layout, + policy, + border_mode, + sampling_policy, + align_corners, + mixed_layout, + output_quantization_info); } }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> class ScaleValidationFixture : public ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, DataLayout data_layout, InterpolationPolicy policy, BorderMode border_mode, SamplingPolicy sampling_policy, bool align_corners) { ScaleValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, @@ -245,10 +260,11 @@ public: border_mode, sampling_policy, align_corners, - mixed_layout); + mixed_layout, + QuantizationInfo()); } }; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_SCALE_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_SCALEFIXTURE_H diff --git a/tests/validation/fixtures/ScatterLayerFixture.h b/tests/validation/fixtures/ScatterLayerFixture.h new file mode 100644 index 0000000000..af161ef98b --- /dev/null +++ b/tests/validation/fixtures/ScatterLayerFixture.h @@ -0,0 +1,254 @@ +/* + * Copyright (c) 2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_SCATTERLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_SCATTERLAYERFIXTURE_H + +#include "arm_compute/core/Utils.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "tests/Globals.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/Validation.h" +#include "tests/validation/reference/ScatterLayer.h" +#include "tests/SimpleTensor.h" + +#include <random> +#include <cstdint> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ScatterGenericValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape src_shape, TensorShape updates_shape, TensorShape indices_shape, + TensorShape out_shape, DataType data_type, ScatterInfo scatter_info, bool inplace, bool padding, + QuantizationInfo src_qinfo = QuantizationInfo(), QuantizationInfo o_qinfo = QuantizationInfo()) + { + // this is for improving randomness across tests + _hash = src_shape[0] + src_shape[1] + src_shape[2] + src_shape[3] + src_shape[4] + src_shape[5] + + updates_shape[0] + updates_shape[1] + updates_shape[2] + updates_shape[3] + + updates_shape[4] + updates_shape[5] + + indices_shape[0] + indices_shape[1] + indices_shape[2] + indices_shape[3]; + + _target = compute_target(src_shape, updates_shape, indices_shape, out_shape, data_type, scatter_info, inplace, padding, src_qinfo, o_qinfo); + _reference = compute_reference(src_shape, updates_shape, indices_shape, out_shape, data_type,scatter_info, src_qinfo , o_qinfo); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::F32: + case DataType::F16: + { + std::uniform_real_distribution<float> distribution(-10.f, 10.f); + library->fill(tensor, distribution, i); + break; + } + case DataType::S32: + case DataType::S16: + case DataType::S8: + { + std::uniform_int_distribution<int32_t> distribution(-100, 100); + library->fill(tensor, distribution, i); + break; + } + case DataType::U32: + case DataType::U16: + case DataType::U8: + { + std::uniform_int_distribution<uint32_t> distribution(0, 200); + library->fill(tensor, distribution, i); + break; + } + default: + { + ARM_COMPUTE_ERROR("Unsupported data type."); + } + } + } + + // This is used to fill indices tensor with S32 datatype. + // Used to prevent ONLY having values that are out of bounds. + template <typename U> + void fill_indices(U &&tensor, int i, const TensorShape &shape) + { + // Calculate max indices the shape should contain. Add an arbitrary value to allow testing for some out of bounds values (In this case min dimension) + const int32_t max = std::min({shape[0] , shape[1], shape[2]}) + 1; + library->fill_tensor_uniform(tensor, i, static_cast<int32_t>(0), static_cast<int32_t>(max)); + } + + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, + const TensorShape &out_shape, DataType data_type, const ScatterInfo info, bool inplace, bool padding, + QuantizationInfo a_qinfo, QuantizationInfo o_qinfo) + { + // 1. Create relevant tensors using ScatterInfo data structure. + // ---------------------------------------------------- + // In order - src, updates, indices, output. + TensorType src = create_tensor<TensorType>(shape_a, data_type, 1, a_qinfo); + TensorType updates = create_tensor<TensorType>(shape_b, data_type, 1, a_qinfo); + TensorType indices = create_tensor<TensorType>(shape_c, DataType::S32, 1, QuantizationInfo()); + TensorType dst = create_tensor<TensorType>(out_shape, data_type, 1, o_qinfo); + + FunctionType scatter; + + // Configure operator + // When scatter_info.zero_initialization is true, pass nullptr for src + // because dst does not need to be initialized with src values. + if(info.zero_initialization) + { + scatter.configure(nullptr, &updates, &indices, &dst, info); + } + else + { + if(inplace) + { + scatter.configure(&src, &updates, &indices, &src, info); + } + else + { + scatter.configure(&src, &updates, &indices, &dst, info); + } + } + + // Assertions + ARM_COMPUTE_ASSERT(src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(updates.info()->is_resizable()); + ARM_COMPUTE_ASSERT(indices.info()->is_resizable()); + ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); + + if(padding) + { + add_padding_x({ &src, &updates, &indices}); + + if(!inplace) + { + add_padding_x({ &dst }); + } + } + + // Allocate tensors + src.allocator()->allocate(); + updates.allocator()->allocate(); + indices.allocator()->allocate(); + + if(!inplace) + { + dst.allocator()->allocate(); + } + + ARM_COMPUTE_ASSERT(!src.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!updates.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!indices.info()->is_resizable()); + + if(!inplace) + { + ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); + } + + // Fill update (a) and indices (b) tensors. + fill(AccessorType(src), 0 + _hash); + fill(AccessorType(updates), 1+ _hash); + fill_indices(AccessorType(indices), 2 + _hash, out_shape); + + scatter.run(); + + if(inplace) + { + return src; + } + else + { + return dst; + } + } + + SimpleTensor<T> compute_reference(const TensorShape &a_shape, const TensorShape &b_shape, const TensorShape &c_shape, + const TensorShape &out_shape, DataType data_type, ScatterInfo info, QuantizationInfo a_qinfo, QuantizationInfo o_qinfo) + { + // Output Quantization not currently in use - fixture should be extended to support this. + ARM_COMPUTE_UNUSED(o_qinfo); + TensorShape src_shape = a_shape; + TensorShape updates_shape = b_shape; + TensorShape indices_shape = c_shape; + const int num_ind_dims = c_shape.num_dimensions(); + + // 1. Collapse batch index into a single dim if necessary for update tensor and indices tensor. + if(num_ind_dims >= 3) + { + indices_shape = indices_shape.collapsed_from(1); + updates_shape = updates_shape.collapsed_from(updates_shape.num_dimensions() - (num_ind_dims -1)); // Collapses batch dims + } + + // 2. Collapse data dims into a single dim. + // Collapse all src dims into 2 dims. First one holding data, the other being the index we iterate over. + src_shape.collapse(updates_shape.num_dimensions() - 1); // Collapse all data dims into single dim. + src_shape = src_shape.collapsed_from(1); // Collapse all index dims into a single dim + updates_shape.collapse(updates_shape.num_dimensions() - 1); // Collapse data dims (all except last dim which is batch dim) + + // Create reference tensors + SimpleTensor<T> src{ src_shape, data_type, 1, a_qinfo }; + SimpleTensor<T> updates{updates_shape, data_type, 1, QuantizationInfo() }; + SimpleTensor<int32_t> indices{ indices_shape, DataType::S32, 1, QuantizationInfo() }; + + // Fill reference + fill(src, 0 + _hash); + fill(updates, 1 + _hash); + fill_indices(indices, 2 + _hash, out_shape); + + // Calculate individual reference using collapsed shapes + return reference::scatter_layer<T>(src, updates, indices, out_shape, info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; + int32_t _hash{}; +}; + +// This fixture will use the same shape for updates as indices. +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ScatterValidationFixture : public ScatterGenericValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape src_shape, TensorShape update_shape, TensorShape indices_shape, + TensorShape out_shape, DataType data_type, ScatterFunction func, bool zero_init, bool inplace, bool padding) + { + ScatterGenericValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, update_shape, + indices_shape, out_shape, data_type, ScatterInfo(func, zero_init), inplace, padding, + QuantizationInfo(), QuantizationInfo()); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_SCATTERLAYERFIXTURE_H diff --git a/tests/validation/fixtures/ScharrFixture.h b/tests/validation/fixtures/ScharrFixture.h index e67d0178b0..b54a9d29e6 100644 --- a/tests/validation/fixtures/ScharrFixture.h +++ b/tests/validation/fixtures/ScharrFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -66,7 +66,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class ScharrValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, BorderMode border_mode, Format format, GradientDimension gradient_dimension) { // Generate a random constant value diff --git a/tests/validation/fixtures/SelectFixture.h b/tests/validation/fixtures/SelectFixture.h index 7b035574f7..8cb6f062f9 100644 --- a/tests/validation/fixtures/SelectFixture.h +++ b/tests/validation/fixtures/SelectFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -63,7 +63,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SelectValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, bool has_same_same_rank, DataType data_type) { TensorShape condition_shape = detail::select_condition_shape(shape, has_same_same_rank); diff --git a/tests/validation/fixtures/SliceOperationsFixtures.h b/tests/validation/fixtures/SliceOperationsFixtures.h index 267cdd5123..b1f91ea2e0 100644 --- a/tests/validation/fixtures/SliceOperationsFixtures.h +++ b/tests/validation/fixtures/SliceOperationsFixtures.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,7 +45,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SliceFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, Coordinates starts, Coordinates ends, DataType data_type) { _target = compute_target(shape, starts, ends, data_type); @@ -108,7 +107,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class StridedSliceFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, Coordinates starts, Coordinates ends, BiStrides strides, int32_t begin_mask, int32_t end_mask, int32_t shrink_mask, diff --git a/tests/validation/fixtures/SoftmaxLayerFixture.h b/tests/validation/fixtures/SoftmaxLayerFixture.h index 352e215dee..f4bf8df9c0 100644 --- a/tests/validation/fixtures/SoftmaxLayerFixture.h +++ b/tests/validation/fixtures/SoftmaxLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SoftmaxValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, float beta, size_t axis) { _quantization_info = quantization_info; @@ -131,7 +130,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SoftmaxValidationFixture : public SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T, IS_LOG> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, float beta, size_t axis) { SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T, IS_LOG>::setup(shape, @@ -146,7 +144,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SoftmaxValidationQuantizedFixture : public SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T, IS_LOG> { public: - template <typename...> void setup(TensorShape shape, DataType data_type, QuantizationInfo quantization_info, float beta, size_t axis) { SoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T, IS_LOG>::setup(shape, diff --git a/tests/validation/fixtures/SpaceToBatchFixture.h b/tests/validation/fixtures/SpaceToBatchFixture.h index cfa7775c21..964e511301 100644 --- a/tests/validation/fixtures/SpaceToBatchFixture.h +++ b/tests/validation/fixtures/SpaceToBatchFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -39,7 +39,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SpaceToBatchLayerValidationGenericFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape block_shape_shape, TensorShape paddings_shape, TensorShape output_shape, DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info) { @@ -140,7 +139,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SpaceToBatchLayerValidationFixture : public SpaceToBatchLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape block_shape_shape, TensorShape paddings_shape, TensorShape output_shape, DataType data_type, DataLayout data_layout) { @@ -152,7 +150,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SpaceToBatchLayerValidationQuantizedFixture : public SpaceToBatchLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: - template <typename...> void setup(TensorShape input_shape, TensorShape block_shape_shape, TensorShape paddings_shape, TensorShape output_shape, DataType data_type, DataLayout data_layout, QuantizationInfo quantization_info) { diff --git a/tests/validation/fixtures/SpaceToDepthFixture.h b/tests/validation/fixtures/SpaceToDepthFixture.h index 7448ec546b..2d2e9fad7d 100644 --- a/tests/validation/fixtures/SpaceToDepthFixture.h +++ b/tests/validation/fixtures/SpaceToDepthFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #ifndef ARM_COMPUTE_TEST_SPACE_TO_DEPTH_LAYER_FIXTURE #define ARM_COMPUTE_TEST_SPACE_TO_DEPTH_LAYER_FIXTURE +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "tests/Globals.h" #include "tests/framework/Asserts.h" #include "tests/framework/Fixture.h" @@ -39,7 +40,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class SpaceToDepthLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape output_shape, const int block_shape, DataType data_type, DataLayout data_layout) { _target = compute_target(input_shape, output_shape, block_shape, data_type, data_layout); @@ -69,6 +69,12 @@ protected: TensorType input = create_tensor<TensorType>(input_shape, data_type, 1, QuantizationInfo(), data_layout); TensorType output = create_tensor<TensorType>(output_shape, data_type, 1, QuantizationInfo(), data_layout); + auto calc_out_shape = misc::shape_calculator::compute_space_to_depth_shape(input.info(), block_shape); + ARM_COMPUTE_ASSERT(output_shape[0] == calc_out_shape[0]); + ARM_COMPUTE_ASSERT(output_shape[1] == calc_out_shape[1]); + ARM_COMPUTE_ASSERT(output_shape[2] == calc_out_shape[2]); + ARM_COMPUTE_ASSERT(output_shape[3] == calc_out_shape[3]); + // Create and configure function FunctionType space_to_depth; space_to_depth.configure(&input, &output, block_shape); diff --git a/tests/validation/fixtures/SplitFixture.h b/tests/validation/fixtures/SplitFixture.h index 3006f21467..203925329c 100644 --- a/tests/validation/fixtures/SplitFixture.h +++ b/tests/validation/fixtures/SplitFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename ITensorType, typename AccessorType, type class SplitFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, unsigned int axis, unsigned int splits, DataType data_type) { _target = compute_target(shape, axis, splits, data_type); @@ -149,7 +148,6 @@ template <typename TensorType, typename ITensorType, typename AccessorType, type class SplitShapesFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, unsigned int axis, std::vector<TensorShape> split_shapes, DataType data_type) { _target = compute_target(shape, axis, split_shapes, data_type); diff --git a/tests/validation/fixtures/StackLayerFixture.h b/tests/validation/fixtures/StackLayerFixture.h index f729e059b0..7dd8fe47dc 100644 --- a/tests/validation/fixtures/StackLayerFixture.h +++ b/tests/validation/fixtures/StackLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_STACK_LAYER_FIXTURE -#define ARM_COMPUTE_TEST_STACK_LAYER_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_STACKLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_STACKLAYERFIXTURE_H #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorShape.h" @@ -52,10 +52,9 @@ template <typename TensorType, typename AbstractTensorType, typename AccessorTyp class StackLayerValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape_src, int axis, DataType data_type, int num_tensors) { - _target = compute_target(shape_src, axis, data_type, num_tensors); + _target = compute_target(shape_src, axis, data_type, num_tensors, false /* add_x_padding */); _reference = compute_reference(shape_src, axis, data_type, num_tensors); } @@ -66,7 +65,7 @@ protected: library->fill_tensor_uniform(tensor, i); } - TensorType compute_target(TensorShape shape_src, int axis, DataType data_type, int num_tensors) + TensorType compute_target(TensorShape shape_src, int axis, DataType data_type, int num_tensors, bool add_x_padding) { std::vector<TensorType> tensors(num_tensors); std::vector<AbstractTensorType *> src(num_tensors); @@ -91,6 +90,11 @@ protected: // Allocate and fill the input tensors for(int i = 0; i < num_tensors; ++i) { + if(add_x_padding) + { + add_padding_x({&tensors[i]}, DataLayout::NHWC); + } + ARM_COMPUTE_ASSERT(tensors[i].info()->is_resizable()); tensors[i].allocator()->allocate(); ARM_COMPUTE_ASSERT(!tensors[i].info()->is_resizable()); @@ -99,6 +103,11 @@ protected: fill(AccessorType(tensors[i]), i); } + if(add_x_padding) + { + add_padding_x({&dst}, DataLayout::NHWC); + } + // Allocate output tensor dst.allocator()->allocate(); @@ -132,7 +141,21 @@ protected: TensorType _target{}; SimpleTensor<T> _reference{}; }; + +template <typename TensorType, typename AbstractTensorType, typename AccessorType, typename FunctionType, typename T> +class StackLayerWithPaddingValidationFixture : + public StackLayerValidationFixture<TensorType, AbstractTensorType, AccessorType, FunctionType, T> +{ +public: + using Parent = StackLayerValidationFixture<TensorType, AbstractTensorType, AccessorType, FunctionType, T>; + + void setup(TensorShape shape_src, int axis, DataType data_type, int num_tensors) + { + Parent::_target = Parent::compute_target(shape_src, axis, data_type, num_tensors, true /* add_x_padding */); + Parent::_reference = Parent::compute_reference(shape_src, axis, data_type, num_tensors); + } +}; } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_STACK_LAYER_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_STACKLAYERFIXTURE_H diff --git a/tests/validation/fixtures/TileFixture.h b/tests/validation/fixtures/TileFixture.h index 35773d9a83..979eee5ab1 100644 --- a/tests/validation/fixtures/TileFixture.h +++ b/tests/validation/fixtures/TileFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class TileValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type, const Multiples &multiples) { _target = compute_target(shape, data_type, multiples); diff --git a/tests/validation/fixtures/TransposeFixture.h b/tests/validation/fixtures/TransposeFixture.h index 6f032d4889..212c76cc9a 100644 --- a/tests/validation/fixtures/TransposeFixture.h +++ b/tests/validation/fixtures/TransposeFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_TRANSPOSE_FIXTURE -#define ARM_COMPUTE_TEST_TRANSPOSE_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_TRANSPOSEFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_TRANSPOSEFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -32,7 +32,7 @@ #include "tests/IAccessor.h" #include "tests/framework/Asserts.h" #include "tests/framework/Fixture.h" -#include "tests/validation/reference/Transpose.h" +#include "tests/validation/reference/Permute.h" namespace arm_compute { @@ -44,7 +44,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class TransposeValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape shape, DataType data_type) { _target = compute_target(shape, data_type); @@ -98,7 +97,7 @@ protected: // Fill reference fill(src); - return reference::transpose<T>(src); + return reference::permute<T>(src, PermutationVector(1U, 0U)); } TensorType _target{}; @@ -107,4 +106,4 @@ protected: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_TRANSPOSE_FIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_TRANSPOSEFIXTURE_H diff --git a/tests/validation/fixtures/UNIT/DynamicTensorFixture.h b/tests/validation/fixtures/UNIT/DynamicTensorFixture.h index bdf43050e6..3e96dcbf2d 100644 --- a/tests/validation/fixtures/UNIT/DynamicTensorFixture.h +++ b/tests/validation/fixtures/UNIT/DynamicTensorFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2021 Arm Limited. + * Copyright (c) 2019-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -127,7 +127,6 @@ class DynamicTensorType3SingleFunction : public framework::Fixture using T = float; public: - template <typename...> void setup(TensorShape input_level0, TensorShape input_level1) { input_l0 = input_level0; @@ -251,7 +250,6 @@ class DynamicTensorType3ComplexFunction : public framework::Fixture using T = float; public: - template <typename...> void setup(std::vector<TensorShape> input_shapes, TensorShape weights_shape, TensorShape bias_shape, std::vector<TensorShape> output_shapes, PadStrideInfo info) { num_iterations = input_shapes.size(); @@ -390,7 +388,6 @@ class DynamicTensorType2PipelineFunction : public framework::Fixture using T = float; public: - template <typename...> void setup(std::vector<TensorShape> input_shapes) { _data_type = DataType::F32; diff --git a/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h b/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h index af9f776ebc..f5e6071340 100644 --- a/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h +++ b/tests/validation/fixtures/UNIT/WeightsRetentionFixture.h @@ -74,10 +74,10 @@ protected: TensorType compute_target() { // Create tensors - TensorType w1 = create_tensor<TensorType>(TensorShape(180000U, 150U), DataType::F32, 1); - TensorType b1 = create_tensor<TensorType>(TensorShape(150U), DataType::F32, 1); - TensorType src = create_tensor<TensorType>(TensorShape(1U, 150U, 1200U, _max_batches), DataType::F32, 1); - TensorType dst = create_tensor<TensorType>(TensorShape(150U, _max_batches), DataType::F32, 1); + TensorType w1 = create_tensor<TensorType>(TensorShape(6000U, 15U), DataType::F32, 1); + TensorType b1 = create_tensor<TensorType>(TensorShape(15U), DataType::F32, 1); + TensorType src = create_tensor<TensorType>(TensorShape(1U, 15U, 400U, _max_batches), DataType::F32, 1); + TensorType dst = create_tensor<TensorType>(TensorShape(15U, _max_batches), DataType::F32, 1); // Create and configure function FullyConnectedFunction fc_layer_1; @@ -105,9 +105,9 @@ protected: int diff = _max_batches - _cur_batches; auto new_src_padding = PaddingSize(src_padding.top, src_padding.right, src_padding.bottom + diff, src_padding.left); auto new_dst_padding = PaddingSize(dst_padding.top, dst_padding.right, dst_padding.bottom + diff, dst_padding.left); - src.allocator()->info().set_tensor_shape(TensorShape(1U, 150U, 1200U, _cur_batches)).set_is_resizable(true).extend_padding(new_src_padding); + src.allocator()->info().set_tensor_shape(TensorShape(1U, 15U, 400U, _cur_batches)).set_is_resizable(true).extend_padding(new_src_padding); src.allocator()->info().set_is_resizable(false); - dst.allocator()->info().set_tensor_shape(TensorShape(150U, _cur_batches)).set_is_resizable(true).extend_padding(new_dst_padding); + dst.allocator()->info().set_tensor_shape(TensorShape(15U, _cur_batches)).set_is_resizable(true).extend_padding(new_dst_padding); dst.allocator()->info().set_is_resizable(false); // Configure FC info @@ -129,16 +129,16 @@ protected: SimpleTensor<T> compute_reference() { // Create reference - SimpleTensor<T> w1{ TensorShape(180000U, 150U), DataType::F32 }; - SimpleTensor<T> b1{ TensorShape(150U), DataType::F32 }; - SimpleTensor<T> src{ TensorShape(1U, 150U, 1200U, _cur_batches), DataType::F32 }; + SimpleTensor<T> w1{ TensorShape(6000U, 15U), DataType::F32 }; + SimpleTensor<T> b1{ TensorShape(15U), DataType::F32 }; + SimpleTensor<T> src{ TensorShape(1U, 15U, 400U, _cur_batches), DataType::F32 }; // Fill reference fill(src, 5); fill(w1, 1); fill(b1, 2); - return reference::fully_connected_layer(src, w1, b1, TensorShape(150U, _cur_batches)); + return reference::fully_connected_layer(src, w1, b1, TensorShape(15U, _cur_batches)); } protected: diff --git a/tests/validation/fixtures/UnstackFixture.h b/tests/validation/fixtures/UnstackFixture.h index 4f8b280735..30b7dd5539 100644 --- a/tests/validation/fixtures/UnstackFixture.h +++ b/tests/validation/fixtures/UnstackFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -47,7 +47,6 @@ template <typename TensorType, typename ITensorType, typename AccessorType, type class UnstackValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, int axis, int num, DataType data_type) { _target = compute_target(input_shape, axis, num, data_type); diff --git a/tests/validation/fixtures/WeightsReshapeFixture.h b/tests/validation/fixtures/WeightsReshapeFixture.h index 0b3e76d677..68bd8b689d 100644 --- a/tests/validation/fixtures/WeightsReshapeFixture.h +++ b/tests/validation/fixtures/WeightsReshapeFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -45,10 +45,9 @@ namespace validation using namespace arm_compute::misc::shape_calculator; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class WeightsReshapeValidationFixture : public framework::Fixture +class WeightsReshapeOpValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, DataType data_type, bool has_bias, unsigned int num_groups) { const TensorShape output_shape = compute_weights_reshaped_shape(TensorInfo(input_shape, 1, data_type), has_bias, num_groups); @@ -73,7 +72,7 @@ protected: // Create and configure function FunctionType weights_reshape_func; - weights_reshape_func.configure(&src, (has_bias ? &bias : nullptr), &dst, num_groups); + weights_reshape_func.configure(src.info(), (has_bias ? bias.info() : nullptr), dst.info(), num_groups); ARM_COMPUTE_ASSERT(src.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); @@ -99,8 +98,18 @@ protected: fill(AccessorType(bias), 1); } + arm_compute::ITensorPack pack = + { + { arm_compute::TensorType::ACL_SRC, &src }, + { arm_compute::TensorType::ACL_DST, &dst } + }; + + if(has_bias) + { + pack.add_const_tensor(arm_compute::TensorType::ACL_BIAS, &bias); + } // Compute function - weights_reshape_func.run(); + weights_reshape_func.run(pack); return dst; } diff --git a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h index e99eb910e8..20b678b36c 100644 --- a/tests/validation/fixtures/WinogradConvolutionLayerFixture.h +++ b/tests/validation/fixtures/WinogradConvolutionLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2021 Arm Limited. + * Copyright (c) 2018-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_WINOGRAD_LAYER_FIXTURE -#define ARM_COMPUTE_TEST_WINOGRAD_LAYER_FIXTURE +#ifndef ACL_TESTS_VALIDATION_FIXTURES_WINOGRADCONVOLUTIONLAYERFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_WINOGRADCONVOLUTIONLAYERFIXTURE_H #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" @@ -55,7 +55,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class WinogradConvolutionLayerFastMathValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, Size2D dilation, DataType data_type, ActivationLayerInfo act_info, const DataLayout &data_layout) @@ -230,7 +229,7 @@ protected: SimpleTensor<T1> filter_transform_out = reference::winograd_filter_transform<T1>(weights_t1, filter_transform_shape, winograd_info); SimpleTensor<T1> batched_gemm = reference::gemm<T1>(input_transform_out, filter_transform_out, dummy_c, 1.0f, 0.0f); SimpleTensor<T1> conv_out = reference::winograd_output_transform<T1>(batched_gemm, bias_t1, output_transform_shape, winograd_info); - SimpleTensor<T> conv_out_t(std::move(copy_tensor<T, T1>(conv_out))); + SimpleTensor<T> conv_out_t(copy_tensor<T, T1>(conv_out)); return (act_info.enabled()) ? reference::activation_layer<T>(conv_out_t, act_info) : conv_out_t; } @@ -243,7 +242,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class WinogradInputTransformValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, WinogradInfo winograd_info, DataLayout data_layout, DataType data_type) { TensorShape output_shape = compute_winograd_input_transform_shape(TensorInfo(input_shape, 1, data_type), winograd_info); @@ -355,7 +353,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class WinogradFilterTransformValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, Size2D output_tile, DataLayout data_layout, DataType data_type) { WinogradInfo winograd_info(output_tile, Size2D(input_shape[0], input_shape[1]), Size2D() /* Not needed */, PadStrideInfo() /* Not needed */, DataLayout::NCHW /* Not needed */); @@ -470,7 +467,6 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ class WinogradOutputTransformValidationFixture : public framework::Fixture { public: - template <typename...> void setup(TensorShape input_shape, WinogradInfo winograd_info, DataType data_type, ActivationLayerInfo act_info = ActivationLayerInfo()) { _target = compute_target(input_shape, winograd_info, data_type, act_info); @@ -588,4 +584,4 @@ protected: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_WINOGRAD_LAYER_FIXTURE */
\ No newline at end of file +#endif // ACL_TESTS_VALIDATION_FIXTURES_WINOGRADCONVOLUTIONLAYERFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h new file mode 100644 index 0000000000..ca4de11a15 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DepthwiseConv2dFixture.h @@ -0,0 +1,246 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DEPTHWISECONV2DFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DEPTHWISECONV2DFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/DepthwiseConv2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuDepthwiseConv2d.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/reference/DepthwiseConvolutionLayer.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuDepthwiseConv2dValidationGenericFixture : public framework::Fixture +{ +public: + using TBias = typename std::conditional<std::is_same<typename std::decay<T>::type, uint8_t>::value || + std::is_same<typename std::decay<T>::type, int8_t>::value, + int32_t, + T>::type; // If T: uint8_t or int8_t then TBias: int32_t, otherwise TBias: T + + void setup(TensorShape input_shape, + Size2D kernel_size, + const PadStrideInfo &pad_stride, + const Size2D &dilation, + const unsigned int depth_multiplier, + const DataType data_type, + const DataLayout data_layout) + { + ARM_COMPUTE_ERROR_ON(data_layout != + DataLayout::NHWC); // Dynamic fusion depthwise conv2d only supports NHWC layout + + DepthwiseConv2dAttributes dwc_conv2d_attr; + const Padding2D padding_2d(pad_stride.pad_left(), pad_stride.pad_right(), pad_stride.pad_top(), + pad_stride.pad_bottom()); + dwc_conv2d_attr.pad(padding_2d) + .stride(Size2D(pad_stride.stride().first, pad_stride.stride().second)) + .dilation(dilation) + .depth_multiplier(depth_multiplier) + .dimension_rounding_type(pad_stride.round()); + + // Calculate Output and Weight Shapes + TensorShape weights_shape = TensorShape(kernel_size.width, kernel_size.height); + + const TensorInfo in_info(input_shape, 1, data_type); + const TensorInfo we_info(weights_shape, 1, data_type); + + const ConvolutionInfo info{pad_stride, depth_multiplier, ActivationLayerInfo(), dilation}; + const TensorShape output_shape = + misc::shape_calculator::compute_depthwise_convolution_shape(in_info, we_info, info); + + weights_shape.set(2, output_shape.z()); + const TensorShape bias_shape = TensorShape(weights_shape[2]); + + _data_type = data_type; + _data_layout = data_layout; + _target = compute_target(input_shape, weights_shape, bias_shape, dwc_conv2d_attr); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, dwc_conv2d_attr); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch (tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-1.0f, 1.0f}; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + + // Given input is in nchw format + TensorType compute_target(TensorShape input_shape, + TensorShape weights_shape, + const TensorShape &bias_shape, + const DepthwiseConv2dAttributes dwc_conv2d_attr) + { + ARM_COMPUTE_ERROR_ON(_data_layout != DataLayout::NHWC); + + // Our test shapes are assumed in NCHW data layout, thus the permutation + permute(input_shape, PermutationVector(2U, 0U, 1U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + ITensorInfo *input_info = context.create_tensor_info(TensorInfo(input_shape, 1, _data_type, _data_layout)); + ITensorInfo *weight_info = context.create_tensor_info(TensorInfo(weights_shape, 1, _data_type, _data_layout)); + ITensorInfo *bias_info = context.create_tensor_info(TensorInfo(bias_shape, 1, _data_type, _data_layout)); + ITensorInfo *dst_info = context.create_tensor_info(); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, input_info, weight_info, bias_info, dwc_conv2d_attr); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_input{}; + TensorType t_weight{}; + TensorType t_bias{}; + TensorType t_dst{}; + + // Initialize user tensors + t_input.allocator()->init(*input_info); + t_weight.allocator()->init(*weight_info); + t_bias.allocator()->init(*bias_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_input.allocator()->allocate(); + t_weight.allocator()->allocate(); + t_bias.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_input), 0); + fill(AccessorType(t_weight), 1); + fill(AccessorType(t_bias), 2); + + // Run runtime + runtime.run({&t_input, &t_weight, &t_bias, &t_dst}); + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &input_shape, + const TensorShape &weights_shape, + const TensorShape &bias_shape, + const TensorShape &output_shape, + DepthwiseConv2dAttributes dwc_conv2d_attr) + { + // Create reference + SimpleTensor<T> src{input_shape, _data_type, 1}; + SimpleTensor<T> weight{weights_shape, _data_type, 1}; + SimpleTensor<TBias> bias{bias_shape, _data_type, 1}; + + fill(src, 0); + fill(weight, 1); + fill(bias, 2); + + auto src_nchw = src; + auto weights_nchw = weight; + auto bias_nchw = bias; + auto output_shape_nchw = output_shape; + + PadStrideInfo legacy_pad_stride(dwc_conv2d_attr.stride().x(), dwc_conv2d_attr.stride().y(), + dwc_conv2d_attr.pad().left, dwc_conv2d_attr.pad().right, + dwc_conv2d_attr.pad().top, dwc_conv2d_attr.pad().bottom, + DimensionRoundingType{}); + auto dst_nchw = + reference::depthwise_convolution(src_nchw, weights_nchw, bias_nchw, output_shape_nchw, legacy_pad_stride, + dwc_conv2d_attr.depth_multiplier(), dwc_conv2d_attr.dilation()); + return dst_nchw; + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; + DataType _data_type{}; + DataLayout _data_layout{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuDepthwiseConv2dValidationFixture + : public DynamicFusionGpuDepthwiseConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, + Size2D kernel_size, + const PadStrideInfo &info, + const Size2D &dilation, + const unsigned int depth_multiplier, + DataType data_type, + DataLayout data_layout) + { + DynamicFusionGpuDepthwiseConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, kernel_size, info, dilation, depth_multiplier, data_type, data_layout); + } +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DEPTHWISECONV2DFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h new file mode 100644 index 0000000000..1f4e223b93 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/DirectConv2dFixture.h @@ -0,0 +1,411 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/Conv2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuConv2d.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/reference/ConvolutionLayer.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +template <typename U> +void fill(U &&tensor, int i) +{ + switch (tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-1.0f, 1.0f}; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } +} + +} // namespace + +/** General Conv2d fixture + * Adapted from tests/validation/fixtures/ConvolutionLayerFixture.h + * TODO: Parameterize to be fully backend agnostic: COMPMID-5760; remove Gpu from name + */ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuConv2dValidationGenericFixture : public framework::Fixture +{ +public: + using TBias = typename std::conditional<std::is_same<typename std::decay<T>::type, uint8_t>::value || + std::is_same<typename std::decay<T>::type, int8_t>::value, + int32_t, + T>::type; // If T: uint8_t or int8_t then TBias: int32_t, otherwise TBias: T + + void setup(TensorShape input_shape, + TensorShape weights_shape, + TensorShape bias_shape, + TensorShape output_shape, + const PadStrideInfo &info, + const Size2D &dilation, + DataType data_type, + DataLayout data_layout, + QuantizationInfo quantization_info, + QuantizationInfo weight_quantization_info) + { + ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion conv2d only supports NHWC layout + const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, dilation); + _data_type = data_type; + _data_layout = data_layout; + _is_quantized = is_data_type_quantized_asymmetric(data_type); + _quantization_info = quantization_info; + _weight_quantization_info = weight_quantization_info; + _bias_data_type = _is_quantized ? DataType::S32 : data_type; + _target = compute_target(input_shape, weights_shape, bias_shape, conv2d_attr); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr); + } + +protected: + // Given input is in nchw format + TensorType compute_target(TensorShape input_shape, + TensorShape weights_shape, + const TensorShape &bias_shape, + Conv2dAttributes conv2d_attr) + { + ARM_COMPUTE_ERROR_ON(_data_layout != DataLayout::NHWC); + permute(input_shape, PermutationVector(2U, 0U, 1U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + CLScheduler::get().default_reinit(); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + ITensorInfo *input_info = context.create_tensor_info(TensorInfo(input_shape, 1, _data_type, _data_layout)); + ITensorInfo *weight_info = context.create_tensor_info(TensorInfo(weights_shape, 1, _data_type, _data_layout)); + ITensorInfo *bias_info = context.create_tensor_info(TensorInfo(bias_shape, 1, _data_type, _data_layout)); + ITensorInfo *dst_info = context.create_tensor_info(); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, input_info, weight_info, bias_info, conv2d_attr); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + // Construct user tensors + TensorType t_input{}; + TensorType t_weight{}; + TensorType t_bias{}; + TensorType t_dst{}; + + // Initialize user tensors + t_input.allocator()->init(*input_info); + t_weight.allocator()->init(*weight_info); + t_bias.allocator()->init(*bias_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_input.allocator()->allocate(); + t_weight.allocator()->allocate(); + t_bias.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_input), 0); + fill(AccessorType(t_weight), 1); + fill(AccessorType(t_bias), 2); + + // Run runtime + runtime.run({&t_input, &t_weight, &t_bias, &t_dst}); + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &input_shape, + const TensorShape &weights_shape, + const TensorShape &bias_shape, + const TensorShape &output_shape, + Conv2dAttributes conv2d_attr) + { + // Create reference + SimpleTensor<T> src{input_shape, _data_type, 1, _quantization_info}; + SimpleTensor<T> weight{weights_shape, _data_type, 1, _weight_quantization_info}; + SimpleTensor<TBias> bias{bias_shape, _data_type, 1, _quantization_info}; + + fill(src, 0); + fill(weight, 1); + fill(bias, 2); + + auto src_nchw = src; + auto weights_nchw = weight; + auto bias_nchw = bias; + auto output_shape_nchw = output_shape; + + PadStrideInfo legacy_pad_stride(conv2d_attr.stride().x(), conv2d_attr.stride().y(), conv2d_attr.pad().left, + conv2d_attr.pad().right, conv2d_attr.pad().top, conv2d_attr.pad().bottom, + DimensionRoundingType{}); + auto dst_nchw = reference::convolution_layer(src_nchw, weights_nchw, bias_nchw, output_shape_nchw, + legacy_pad_stride, conv2d_attr.dilation()); + return dst_nchw; + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; + DataType _data_type{}; + DataType _bias_data_type{}; + DataLayout _data_layout{}; + QuantizationInfo _quantization_info{}; + QuantizationInfo _weight_quantization_info{}; + bool _is_quantized = false; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuConv2dValidationFixture + : public DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, + TensorShape weights_shape, + TensorShape output_shape, + TensorShape bias_shape, + const PadStrideInfo &info, + const Size2D &dialation, + DataType data_type, + DataLayout data_layout, + QuantizationInfo quantization_info) + { + DynamicFusionGpuConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, weights_shape, output_shape, bias_shape, info, dialation, data_type, data_layout, + quantization_info, quantization_info); + } +}; + +/** Specific Conv2d method: Direct Conv2d fixture + * Adapted from tests/validation/fixtures/DirectConvolutionLayerFixture.h + * TODO: Parameterize to be fully backend agnostic: COMPMID-5760 + */ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionDirectConv2dValidationGenericFixture : public framework::Fixture +{ +public: + using TBias = + typename std::conditional<std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value, int32_t, T>::type; + + void setup(TensorShape input_shape, + int stride_x, + int stride_y, + int pad_x, + int pad_y, + unsigned int kernel_size, + unsigned int num_kernels, + DataType data_type, + QuantizationInfo quantization_info, + DataLayout data_layout) + { + ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion conv2d only supports NHWC layout + + TensorShape weights_shape(kernel_size, kernel_size, input_shape.z(), num_kernels); + const TensorShape bias_shape(num_kernels); + const PadStrideInfo info(stride_x, stride_y, pad_x, pad_y, DimensionRoundingType::FLOOR); + const DataType bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; + + const Conv2dAttributes conv2d_attr = convert_pad_stride_info_to_conv_attr(info, {1U, 1U} /* dilation */); + + TensorInfo input_info = TensorInfo(input_shape, 1, data_type); + TensorInfo weights_info = TensorInfo(weights_shape, 1, data_type); + + const TensorShape output_shape = + misc::shape_calculator::compute_deep_convolution_shape(input_info, weights_info, info); + + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, conv2d_attr, data_type, + bias_data_type, quantization_info, data_layout); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, + bias_data_type, quantization_info); + } + +protected: + TensorType compute_target(TensorShape input_shape, + TensorShape weights_shape, + const TensorShape &bias_shape, + TensorShape output_shape, + const Conv2dAttributes &conv2d_attr, + DataType data_type, + DataType bias_data_type, + QuantizationInfo quantization_info, + const DataLayout &data_layout) + { + ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); + ARM_COMPUTE_UNUSED(quantization_info); + // Dataset shapes are in NCHW layout + permute(input_shape, PermutationVector(2U, 0U, 1U)); + permute(weights_shape, PermutationVector(2U, 0U, 1U)); + permute(output_shape, PermutationVector(2U, 0U, 1U)); + + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + auto input_info = context.create_tensor_info(TensorInfo(input_shape, 1, data_type, data_layout)); + auto weight_info = context.create_tensor_info(TensorInfo(weights_shape, 1, data_type, data_layout)); + auto bias_info = context.create_tensor_info(TensorInfo(bias_shape, 1, bias_data_type, data_layout)); + auto dst_info = context.create_tensor_info(); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, input_info, weight_info, bias_info, conv2d_attr); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + // Construct user tensors + TensorType t_input{}; + TensorType t_weight{}; + TensorType t_bias{}; + TensorType t_dst{}; + + // Initialize user tensors + t_input.allocator()->init(*input_info); + t_weight.allocator()->init(*weight_info); + t_bias.allocator()->init(*bias_info); + t_dst.allocator()->init(*dst_info); + + ARM_COMPUTE_ASSERT(t_input.info()->is_resizable()); + ARM_COMPUTE_ASSERT(t_weight.info()->is_resizable()); + ARM_COMPUTE_ASSERT(t_bias.info()->is_resizable()); + ARM_COMPUTE_ASSERT(t_dst.info()->is_resizable()); + + // Allocate and fill user tensors + t_input.allocator()->allocate(); + t_weight.allocator()->allocate(); + t_bias.allocator()->allocate(); + t_dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!t_input.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!t_weight.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!t_bias.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!t_dst.info()->is_resizable()); + + fill(AccessorType(t_input), 0); + fill(AccessorType(t_weight), 1); + fill(AccessorType(t_bias), 2); + + // Run runtime + runtime.run({&t_input, &t_weight, &t_bias, &t_dst}); + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &input_shape, + const TensorShape &weights_shape, + const TensorShape &bias_shape, + const TensorShape &output_shape, + const PadStrideInfo &info, + DataType data_type, + DataType bias_data_type, + QuantizationInfo quantization_info) + { + // Create reference + SimpleTensor<T> src{input_shape, data_type, 1, quantization_info}; + SimpleTensor<T> weights{weights_shape, data_type, 1, quantization_info}; + SimpleTensor<TBias> bias{bias_shape, bias_data_type, 1, quantization_info}; + + // Fill reference + fill(src, 0); + fill(weights, 1); + fill(bias, 2); + + SimpleTensor<T> dst = reference::convolution_layer<T>(src, weights, bias, output_shape, info); + return dst; + } + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionDirectConv2dValidationFixture + : public DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, + int stride_x, + int stride_y, + int pad_x, + int pad_y, + unsigned int kernel_size, + unsigned int num_kernels, + DataType data_type, + DataLayout data_layout) + { + DynamicFusionDirectConv2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, QuantizationInfo(), + data_layout); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_DIRECTCONV2DFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h new file mode 100644 index 0000000000..69bd0efbdc --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h @@ -0,0 +1,273 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_ELEMENTWISEBINARYFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_ELEMENTWISEBINARYFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/reference/ElementwiseOperations.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuElementwiseBinaryValidationGenericFixture : public framework::Fixture +{ +public: + void setup(ArithmeticOperation ref_op, + const TensorShape &shape0, + const TensorShape &shape1, + const TensorShape &shape2, + DataType data_type, + bool is_inplace, + bool fuse_two_ops = false) + { + _ref_op = ref_op; + _is_inplace = is_inplace; + _data_type = data_type; + _fuse = fuse_two_ops; + ARM_COMPUTE_ERROR_ON_MSG(_fuse && shape2.total_size() == 0, "No shape2 provided for fusion of two ops."); + ARM_COMPUTE_ERROR_ON_MSG(_fuse && _is_inplace, "In place for fusing case not supported yet."); + _target = compute_target(shape0, shape1, shape2); + _reference = compute_reference(shape0, shape1, shape2); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + if (is_data_type_float(tensor.data_type())) + { + switch (_ref_op) + { + case ArithmeticOperation::DIV: + library->fill_tensor_uniform_ranged(tensor, i, {std::pair<float, float>(-0.001f, 0.001f)}); + break; + case ArithmeticOperation::POWER: + library->fill_tensor_uniform(tensor, i, 0.0f, 5.0f); + break; + default: + library->fill_tensor_uniform(tensor, i); + } + } + else if (tensor.data_type() == DataType::S32) + { + switch (_ref_op) + { + case ArithmeticOperation::DIV: + library->fill_tensor_uniform_ranged(tensor, i, {std::pair<int32_t, int32_t>(-1U, 1U)}); + break; + default: + library->fill_tensor_uniform(tensor, i); + } + } + else + { + library->fill_tensor_uniform(tensor, i); + } + } + + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) + { + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Fuse first element wise binary Op + ITensorInfo *lhs_info = context.create_tensor_info(TensorInfo(shape0, 1, _data_type)); + ITensorInfo *rhs_info = context.create_tensor_info(TensorInfo(shape1, 1, _data_type)); + ITensorInfo *dst_info = context.create_tensor_info(); + + ITensorInfo *rhs_info_fuse = nullptr; + + ITensorInfo *ans_info = FunctionType::create_op(sketch, lhs_info, rhs_info); + + if (_fuse) + { + rhs_info_fuse = context.create_tensor_info(TensorInfo(shape2, 1, _data_type)); + ITensorInfo *ans2_info = FunctionType::create_op(sketch, ans_info, rhs_info_fuse); + GpuOutput::create_op(sketch, ans2_info, dst_info); + } + else + { + GpuOutput::create_op(sketch, ans_info, dst_info); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_lhs{}; + TensorType t_rhs{}; + TensorType t_rhs_fuse{}; + TensorType t_dst{}; + + // Initialize user tensors + t_lhs.allocator()->init(*lhs_info); + t_rhs.allocator()->init(*rhs_info); + t_dst.allocator()->init(*dst_info); + if (_fuse) + { + t_rhs_fuse.allocator()->init(*rhs_info_fuse); + } + + // Allocate and fill user tensors + // Instead of using ACL allocator, the user can choose to import memory into the tensors + t_lhs.allocator()->allocate(); + t_rhs.allocator()->allocate(); + t_dst.allocator()->allocate(); + if (_fuse) + { + t_rhs_fuse.allocator()->allocate(); + } + + fill(AccessorType(t_lhs), 0); + fill(AccessorType(t_rhs), 1); + if (_fuse) + { + fill(AccessorType(t_rhs_fuse), 2); + } + + // Run runtime + if (_fuse) + { + runtime.run({&t_lhs, &t_rhs, &t_rhs_fuse, &t_dst}); + } + else + { + runtime.run({&t_lhs, &t_rhs, &t_dst}); + } + + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) + { + const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1); + const TensorShape out_shape_fuse = TensorShape::broadcast_shape(out_shape, shape1); + + // Create reference + SimpleTensor<T> ref_lhs{shape0, _data_type, 1, QuantizationInfo()}; + SimpleTensor<T> ref_rhs{shape1, _data_type, 1, QuantizationInfo()}; + SimpleTensor<T> ref_rhs_fuse{shape2, _data_type, 1, QuantizationInfo()}; + SimpleTensor<T> ref_dst{out_shape, _data_type, 1, QuantizationInfo()}; + SimpleTensor<T> ref_dst_fuse{out_shape_fuse, _data_type, 1, QuantizationInfo()}; + + // Fill reference + fill(ref_lhs, 0); + fill(ref_rhs, 1); + + reference::arithmetic_operation<T>(_ref_op, ref_lhs, ref_rhs, ref_dst, ConvertPolicy::WRAP); + if (_fuse) + { + fill(ref_rhs_fuse, 2); + reference::arithmetic_operation<T>(_ref_op, ref_dst, ref_rhs_fuse, ref_dst_fuse, ConvertPolicy::WRAP); + } + SimpleTensor<T> *ret = _fuse ? &ref_dst_fuse : &ref_dst; + return *ret; + } + + ArithmeticOperation _ref_op{ArithmeticOperation::ADD}; + TensorType _target{}; + SimpleTensor<T> _reference{}; + DataType _data_type{}; + DataLayout _data_layout{}; + bool _is_inplace{false}; + bool _fuse{false}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuElementwiseBinaryOneOpValidationFixture + : public DynamicFusionGpuElementwiseBinaryValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(ArithmeticOperation ref_op, const TensorShape &shape0, DataType data_type, bool is_inplace) + { + DynamicFusionGpuElementwiseBinaryValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + ref_op, shape0, shape0, TensorShape(), data_type, is_inplace); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuElementwiseBinaryBroadcastOneOpValidationFixture + : public DynamicFusionGpuElementwiseBinaryValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(ArithmeticOperation ref_op, + const TensorShape &shape0, + const TensorShape &shape1, + DataType data_type, + bool is_inplace) + { + DynamicFusionGpuElementwiseBinaryValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + ref_op, shape0, shape1, TensorShape(), data_type, is_inplace); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuElementwiseBinaryTwoOpsValidationFixture + : public DynamicFusionGpuElementwiseBinaryValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(ArithmeticOperation ref_op, + const TensorShape &shape0, + const TensorShape &shape1, + const TensorShape &shape2, + DataType data_type, + bool is_inplace, + bool fuse_two_ops) + { + DynamicFusionGpuElementwiseBinaryValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + ref_op, shape0, shape1, shape2, data_type, is_inplace, fuse_two_ops); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_ELEMENTWISEBINARYFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/MatMulKernelFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/MatMulKernelFixture.h new file mode 100644 index 0000000000..4c1cc94d3d --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/MatMulKernelFixture.h @@ -0,0 +1,297 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_MATMULKERNELFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_MATMULKERNELFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/MatMulAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuMatMul.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/reference/GEMM.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/reference/ReshapeLayer.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +template <typename U> +void fill(U &&tensor, int i) +{ + switch (tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-1.0f, 1.0f}; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } +} + +} // namespace +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuMatMulValidationGenericFixture : public framework::Fixture +{ +public: + void setup(TensorShape lhs_shape, + TensorShape rhs_shape, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + int M0, + int N0, + int K0, + bool export_rhs_to_cl_image, + DataType data_type) + { + //For brevity, the input shapes are assumed to be not-transposed for both a and b matrices. + if (transpose_a) + { + permute(lhs_shape, PermutationVector(1U, 0U)); + } + if (transpose_b) + { + permute(rhs_shape, PermutationVector(1U, 0U)); + } + + // Skip configurations unsupported by the device. + _device_supports_export_to_cl_image = image2d_from_buffer_supported(CLKernelLibrary::get().get_device()); + if (!_device_supports_export_to_cl_image && export_rhs_to_cl_image) + { + ARM_COMPUTE_TEST_INFO("cl_khr_image2d_from_buffer not supported. TEST skipped"); + framework::ARM_COMPUTE_PRINT_INFO(); + return; // Note: Also need to skip the validate in corresponding FIXTURE_DATA_TEST_CASEs. + } + + _target = compute_target(lhs_shape, rhs_shape, transpose_a, transpose_b, M0, N0, K0, export_rhs_to_cl_image, + data_type); + _reference = compute_reference(lhs_shape, rhs_shape, output_shape, transpose_a, transpose_b, data_type); + } + +protected: + TensorType compute_target(TensorShape &shape_a, + TensorShape &shape_b, + bool transpose_a, + bool transpose_b, + int M0, + int N0, + int K0, + bool export_rhs_to_cl_image, + DataType data_type) + { + ARM_COMPUTE_UNUSED(export_rhs_to_cl_image); + CLScheduler::get().default_reinit(); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + ITensorInfo *lhs_info = context.create_tensor_info(TensorInfo(shape_a, 1, data_type)); + ITensorInfo *rhs_info = context.create_tensor_info(TensorInfo(shape_b, 1, data_type)); + ITensorInfo *dst_info = context.create_tensor_info(); + + MatMulAttributes matmul_attr{}; + matmul_attr.adj_lhs(transpose_a); + matmul_attr.adj_rhs(transpose_b); + + GpuMatMulSettings matmul_settings{}; + matmul_settings.m0(M0); + matmul_settings.n0(N0); + matmul_settings.k0(K0); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, lhs_info, rhs_info, matmul_attr, matmul_settings); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_lhs{}; + TensorType t_rhs{}; + TensorType t_dst{}; + + // Initialize user tensors + t_lhs.allocator()->init(*lhs_info); + t_rhs.allocator()->init(*rhs_info); + t_dst.allocator()->init(*dst_info); + + ARM_COMPUTE_ASSERT(t_lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(t_rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(t_dst.info()->is_resizable()); + + // Allocate and fill user tensors + t_lhs.allocator()->allocate(); + t_rhs.allocator()->allocate(); + t_dst.allocator()->allocate(); + + ARM_COMPUTE_ASSERT(!t_lhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!t_rhs.info()->is_resizable()); + ARM_COMPUTE_ASSERT(!t_dst.info()->is_resizable()); + + fill(AccessorType(t_lhs), 0); + fill(AccessorType(t_rhs), 1); + + // Run runtime + runtime.run({&t_lhs, &t_rhs, &t_dst}); + + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape_a, + const TensorShape &shape_b, + const TensorShape &output_shape, + bool pretranspose_a, + bool pretranspose_b, + DataType data_type) + { + // We collapse dimensions > 3 onto dimension 3, i.e. 5D+ tensors will look like 3D + // This is necessary unless we choose to extend gemm reference for 5D+ tensors + TensorShape output_shape_collapsed = output_shape.collapsed_from(Window::DimZ); + TensorShape shape_a_collapsed = shape_a.collapsed_from(Window::DimZ); + TensorShape shape_b_collapsed = shape_b.collapsed_from(Window::DimZ); + + // Create reference + SimpleTensor<T> a{shape_a_collapsed, data_type, 1}; + SimpleTensor<T> b{shape_b_collapsed, data_type, 1}; + SimpleTensor<T> c{output_shape_collapsed, data_type, 1}; + + // Fill reference + fill(a, 0); + fill(b, 1); + + /* Note: Assuming the usual batch matmul dimensions A = (B x M x K), B = (B x K x N), if pretranspose_A is set to true, then A is assumed to be (B x K x M), + therefore, A must be pre-transposed before passing it to the fixture. And, we transpose A again in the fixture to make it (B x M x K) + in order to be able to call reference implementation that works with (B x M x K) input. + Similarly, if pretranspose_B is set to true, then B is assumed to be (B x N x K), B must be pre-transposed before passing it to the fixture. */ + + // Define transposed shapes + TensorShape a_transposed_shape(a.shape()); + a_transposed_shape.set(0, a.shape().y()); + a_transposed_shape.set(1, a.shape().x()); + + TensorShape b_transposed_shape(b.shape()); + b_transposed_shape.set(0, b.shape().y()); + b_transposed_shape.set(1, b.shape().x()); + + // Define transposed tensors + SimpleTensor<T> a_transposed{a_transposed_shape, data_type}; + SimpleTensor<T> b_transposed{b_transposed_shape, data_type}; + + //pretranspose a if necessary + if (pretranspose_a) + { + a_transposed = reference::permute<T>(a, PermutationVector(1U, 0U)); + } + + // pretranspose b if necessary + if (pretranspose_b) + { + b_transposed = reference::permute<T>(b, PermutationVector(1U, 0U)); + } + + // Use transposed tensors if boolean enabled else use original tensors + SimpleTensor<T> result = + reference::gemm<T>((pretranspose_a) ? a_transposed : a, (pretranspose_b) ? b_transposed : b, c, 1.0f, 0.f); + + // We reshape the gemm output back if the tensor is high dimensional + if (output_shape_collapsed != output_shape) + { + // std::cout << "called reshape: \n"; + result = reference::reshape_layer(result, output_shape); + } + + return result; + } + + CLTensor _target{}; + SimpleTensor<T> _reference{}; + bool _device_supports_export_to_cl_image{false}; + bool _device_supports_mmul{false}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuMatMulValidationFixture + : public DynamicFusionGpuMatMulValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape lhs_shape, + TensorShape rhs_shape, + TensorShape output_shape, + bool transpose_a, + bool transpose_b, + int M0, + int N0, + int K0, + bool export_rhs_to_cl_image, + DataType data_type) + { + ARM_COMPUTE_UNUSED(export_rhs_to_cl_image); + DynamicFusionGpuMatMulValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + lhs_shape, rhs_shape, output_shape, transpose_a, transpose_b, M0, N0, K0, + false /* export_rhs_to_cl_image bias */, data_type); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_MATMULKERNELFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h new file mode 100644 index 0000000000..b0c7143d91 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h @@ -0,0 +1,188 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_POOL2DFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_POOL2DFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/Pool2dAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuPool2d.h" + +#include "src/dynamic_fusion/utils/Utils.h" +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/PoolingLayer.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuPool2dValidationGenericFixture : public framework::Fixture +{ +public: + void setup(TensorShape input_shape, const Pool2dAttributes &pool_attr, DataType data_type) + { + _target = compute_target(input_shape, pool_attr, data_type); + _reference = compute_reference( + input_shape, convert_pool_attr_to_pool_info(pool_attr, true /* mixed_precision */), data_type); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + switch (tensor.data_type()) + { + case DataType::F16: + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-1.0f, 1.0f}; + library->fill(tensor, distribution, i); + break; + } + case DataType::F32: + { + std::uniform_real_distribution<float> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + + // Given input is in nchw format + TensorType compute_target(TensorShape input_shape, const Pool2dAttributes &pool_attr, const DataType data_type) + { + CLScheduler::get().default_reinit(); + + // Change shape due to NHWC data layout, test shapes are NCHW + permute(input_shape, PermutationVector(2U, 0U, 1U)); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + auto input_info = context.create_tensor_info(TensorInfo(input_shape, 1, data_type, DataLayout::NHWC)); + auto dst_info = context.create_tensor_info(); + + // Create Pool2dSettings + GpuPool2dSettings pool_settings = GpuPool2dSettings(); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, input_info, pool_attr, pool_settings); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + // Construct user tensors + TensorType t_input{}; + TensorType t_dst{}; + + // Initialize user tensors + t_input.allocator()->init(*input_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_input.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_input), 0); + + // Run runtime + runtime.run({&t_input, &t_dst}); + return t_dst; + } + + SimpleTensor<T> compute_reference(TensorShape shape, PoolingLayerInfo pool_info, DataType data_type) + { + // Create reference + SimpleTensor<T> src(shape, data_type, 1, QuantizationInfo()); + // Fill reference + fill(src, 0); + return reference::pooling_layer<T>(src, pool_info, QuantizationInfo(), nullptr, DataLayout::NCHW); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuPool2dValidationFixture + : public DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, + PoolingType pool_type, + Size2D pool_size, + Padding2D pad, + Size2D stride, + bool exclude_padding, + DataType data_type) + { + DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, + Pool2dAttributes().pool_type(pool_type).pool_size(pool_size).pad(pad).stride(stride).exclude_padding( + exclude_padding), + data_type); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuPool2dSpecialValidationFixture + : public DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape input_shape, Pool2dAttributes pool_attr, DataType data_type) + { + DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + input_shape, pool_attr, data_type); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_POOL2DFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h new file mode 100644 index 0000000000..c9ffbccbc7 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ActivationFixture.h @@ -0,0 +1,207 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_ACTIVATIONFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_ACTIVATIONFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/ActivationLayer.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, typename... TArgs> +class DynamicFusionActivationValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape, bool fuse, DataType data_type, ActivationLayerInfo act_info, TArgs... args) + { + _fuse = fuse; + _data_type = data_type; + _function = act_info.activation(); + _target = compute_target(shape, args...); + _reference = compute_reference(shape, act_info); + } + +protected: + std::vector<T> get_boundary_values(T min, T max) + { + // This function will return a vector filled with the following values that can + // represent two partitions derived from equivalent partitioning. + // * Lower partition: min, min + delta, lower quarter (nominal), center - delta + // * Upper partition: center, center + delta, upper quarter (nominal), max - delta, max + const auto delta = is_data_type_float(_data_type) ? T(0.1f) : T(1); + const auto center_value = (min + max) / 2; + const auto lower_quarter = (min + center_value) / 2; + const auto upper_quarter = (center_value + max) / 2; + + std::vector<T> boundary_values{}; + + // To ensure all the inserted values are within the given range after subtracing/adding delta + auto insert_values = [&boundary_values, &min, &max](const std::initializer_list<T> &new_values) + { + for (auto &v : new_values) + { + if (v >= min && v <= max) + { + boundary_values.emplace_back(v); + } + } + }; + + insert_values({min, static_cast<T>(min + delta), static_cast<T>(lower_quarter), + static_cast<T>(center_value - delta)}); // lower partition + insert_values({static_cast<T>(center_value), static_cast<T>(center_value + delta), + static_cast<T>(upper_quarter), static_cast<T>(max - delta), max}); // upper partition + + return boundary_values; + } + + template <typename U> + void fill(U &&tensor) + { + float min_bound = 0; + float max_bound = 0; + std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<T>(_function, _data_type); + library->fill_static_values(tensor, get_boundary_values(static_cast<T>(min_bound), static_cast<T>(max_bound))); + } + + TensorType compute_target(const TensorShape &shape, TArgs... args) + { + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + ITensorInfo *src_info = context.create_tensor_info(TensorInfo(shape, 1, _data_type)); + ITensorInfo *dst_info = context.create_tensor_info(TensorInfo(shape, 1, _data_type)); + + ITensorInfo *ans_0_info = FunctionType::create_op(sketch, src_info, args...); + if (_fuse) + { + ITensorInfo *ans_1_info = FunctionType::create_op(sketch, ans_0_info, args...); + GpuOutput::create_op(sketch, ans_1_info, dst_info); + } + else + { + GpuOutput::create_op(sketch, ans_0_info, dst_info); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst{}; + + // Initialize user tensors + t_src.allocator()->init(*src_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_src)); + + // Run runtime + runtime.run({&t_src, &t_dst}); + + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape, ActivationLayerInfo act_info) + { + // Create reference + SimpleTensor<T> src{shape, _data_type, 1}; + + // Fill reference + fill(src); + + auto tmp = reference::activation_layer<T>(src, act_info); + + if (_fuse) + { + auto dst = reference::activation_layer<T>(tmp, act_info); + return dst; + } + else + { + return tmp; + } + } + +protected: + ActivationLayerInfo::ActivationFunction _function{}; + bool _fuse{false}; + DataType _data_type{}; + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionSigmoidValidationFixture + : public DynamicFusionActivationValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, bool fuse, DataType data_type) + { + ActivationLayerInfo act_info{ActivationLayerInfo::ActivationFunction::LOGISTIC}; + DynamicFusionActivationValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, fuse, + data_type, act_info); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionTanhValidationFixture + : public DynamicFusionActivationValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, bool fuse, DataType data_type) + { + ActivationLayerInfo act_info{ActivationLayerInfo::ActivationFunction::TANH, 1.0f, 1.0f}; + DynamicFusionActivationValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, fuse, + data_type, act_info); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_ACTIVATIONFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h new file mode 100644 index 0000000000..08fffb305b --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/CastFixture.h @@ -0,0 +1,186 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CASTFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CASTFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/CastAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/DepthConvertLayer.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2> +class DynamicFusionCastValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape, DataType dt_in, DataType dt_out, ConvertPolicy policy) + { + _target = compute_target(shape, dt_in, dt_out, policy); + _reference = compute_reference(shape, dt_in, dt_out, policy); + } + +protected: + template <typename U> + void fill(U &&tensor, int i, DataType dt_in, DataType dt_out) + { + // Restricting range to avoid inf values + if (dt_out == DataType::F16) + { + constexpr int signed_min = -32000; + constexpr int signed_max = 32000; + constexpr int unsigned_min = 0; + constexpr int unsigned_max = 65000; + + switch (dt_in) + { + case DataType::U8: + case DataType::QASYMM8: + case DataType::QASYMM8_SIGNED: + case DataType::S8: + case DataType::F32: + { + library->fill_tensor_uniform(tensor, i); + break; + } + case DataType::U16: + { + library->fill_tensor_uniform(tensor, i, static_cast<uint16_t>(unsigned_min), + static_cast<uint16_t>(unsigned_max)); + break; + } + case DataType::S16: + { + library->fill_tensor_uniform(tensor, i, static_cast<int16_t>(signed_min), + static_cast<int16_t>(signed_max)); + break; + } + case DataType::U32: + { + library->fill_tensor_uniform(tensor, i, static_cast<uint32_t>(unsigned_min), + static_cast<uint32_t>(unsigned_max)); + break; + } + case DataType::S32: + { + library->fill_tensor_uniform(tensor, i, static_cast<int32_t>(signed_min), + static_cast<int32_t>(signed_max)); + break; + } + default: + ARM_COMPUTE_ERROR("NOT SUPPORTED!"); + } + } + else + { + library->fill_tensor_uniform(tensor, i); + } + } + + // Given input is in nchw format + TensorType + compute_target(const TensorShape &shape, const DataType dt_in, const DataType dt_out, const ConvertPolicy policy) + { + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + // Here, we use DataLayout::NCHW just for the test. However, the optimal data layout to + // be used with dynamic fusion is NHWC + ITensorInfo *src_info = + context.create_tensor_info(TensorInfo(shape, 1, dt_in, DataLayout::NCHW)); // layout is not important + ITensorInfo *dst_info = context.create_tensor_info(); + + CastAttributes attributes; + attributes.convert_policy(policy).data_type(dt_out); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, src_info, attributes); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst{}; + + // Initialize user tensors + t_src.allocator()->init(*src_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_src), 0, dt_in, dt_out); + + // Run runtime + runtime.run({&t_src, &t_dst}); + return t_dst; + } + + SimpleTensor<T2> + compute_reference(const TensorShape &shape, const DataType dt_in, const DataType dt_out, const ConvertPolicy policy) + { + // Create reference + SimpleTensor<T1> src{shape, dt_in, 1}; + + // Fill reference + fill(src, 0, dt_in, dt_out); + + return reference::depth_convert<T1, T2>(src, dt_out, policy, 0); + } + + TensorType _target{}; + SimpleTensor<T2> _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CASTFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h new file mode 100644 index 0000000000..e8f6f83e42 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ClampFixture.h @@ -0,0 +1,171 @@ +/* + * Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/framework/Fixture.h" +#include "tests/validation/reference/ActivationLayer.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionClampValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape, ClampAttributes attributes, bool fuse, DataType data_type) + { + // CLAMP is implemented as LU_BOUNDED_RELU with the alpha and beta variables swapped. + ActivationLayerInfo act_info{ ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, attributes.max_val(), attributes.min_val() }; + + _fuse = fuse; + _attributes = attributes; + _data_type = data_type; + _target = compute_target(shape, attributes); + _reference = compute_reference(shape, act_info); + } + +protected: + std::vector<T> get_boundary_values(T min, T max) + { + // This function will return a vector filled with the following values that can + // represent two partitions derived from equivalent partitioning. + // * Lower partition: min, min + delta, lower quarter (nominal), center - delta + // * Upper partition: center, center + delta, upper quarter (nominal), max - delta, max + const auto delta = is_data_type_float(_data_type) ? T(0.1f) : T(1); + const auto center_value = (min + max) / 2; + const auto lower_quarter = (min + center_value) / 2; + const auto upper_quarter = (center_value + max) / 2; + + std::vector<T> boundary_values{}; + + // To ensure all the inserted values are within the given range after subtracing/adding delta + auto insert_values = [&boundary_values, &min, &max](const std::initializer_list<T> &new_values) + { + for(auto &v : new_values) + { + if(v >= min && v <= max) + { + boundary_values.emplace_back(v); + } + } + }; + + insert_values({ min, static_cast<T>(min + delta), static_cast<T>(lower_quarter), static_cast<T>(center_value - delta) }); // lower partition + insert_values({ static_cast<T>(center_value), static_cast<T>(center_value + delta), static_cast<T>(upper_quarter), static_cast<T>(max - delta), max }); // upper partition + + return boundary_values; + } + + template <typename U> + void fill(U &&tensor) + { + float min_bound = 0; + float max_bound = 0; + std::tie(min_bound, max_bound) = get_activation_layer_test_bounds<T>(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, _data_type); + library->fill_static_values(tensor, get_boundary_values(static_cast<T>(min_bound), static_cast<T>(max_bound))); + } + + TensorType compute_target(const TensorShape &shape, ClampAttributes attributes) + { + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &context }; + + // Create sketch tensors + ITensorInfo* src_info = context.create_tensor_info(TensorInfo(shape, 1, _data_type)); + ITensorInfo* dst_info = context.create_tensor_info(TensorInfo(shape, 1, _data_type)); + + ITensorInfo *ans_0_info = FunctionType::create_op(sketch, src_info, attributes); + if(_fuse) + { + ITensorInfo *ans_1_info = FunctionType::create_op(sketch, ans_0_info, attributes); + GpuOutput::create_op(sketch, ans_1_info, dst_info); + } + else + { + GpuOutput::create_op(sketch, ans_0_info, dst_info); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst{}; + + // Initialize user tensors + t_src.allocator()->init(*src_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_src)); + + // Run runtime + runtime.run({ &t_src, &t_dst }); + + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape, ActivationLayerInfo act_info) + { + // Create reference + SimpleTensor<T> src{ shape, _data_type, 1, _quantization_info }; + + // Fill reference + fill(src); + + auto dst = reference::activation_layer<T>(src, act_info, _quantization_info); + return dst; + } + +protected: + QuantizationInfo _quantization_info{}; + ClampAttributes _attributes{}; + bool _fuse{ false }; + DataType _data_type{}; + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_CLAMPFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h new file mode 100644 index 0000000000..f02aa5e36a --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h @@ -0,0 +1,239 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_MULFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_MULFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/Globals.h" +#include "tests/validation/reference/PixelWiseMultiplication.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/* We use a separate test fixture for Multiplication op instead of reusing ElementwiseBinaryFixture to avoid exposing + * the internal enum ElementwiseOp to the public utils/TypePrinters.h as required by the data test case macros + * to print the test data. + */ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionMulValidationFixture : public framework::Fixture +{ +public: + void setup(const TensorShape &shape0, + const TensorShape &shape1, + const TensorShape &shape2, + DataType data_type, + bool is_inplace, + bool fuse_two_ops = false) + { + _data_type = data_type; + _is_inplace = is_inplace; + _fuse = fuse_two_ops; + ARM_COMPUTE_ERROR_ON_MSG(_fuse && shape2.total_size() == 0, "No shape2 provided for fusion of two ops."); + ARM_COMPUTE_ERROR_ON_MSG(_fuse && _is_inplace, "In place for fusing case not supported yet."); + _target = compute_target(shape0, shape1, shape2); + _reference = compute_reference(shape0, shape1, shape2); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + library->fill_tensor_uniform(tensor, i); + } + + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) + { + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Fuse first multiplication op + ITensorInfo *lhs_info = context.create_tensor_info(TensorInfo(shape0, 1, _data_type)); + ITensorInfo *rhs_info = context.create_tensor_info(TensorInfo(shape1, 1, _data_type)); + ITensorInfo *dst_info = context.create_tensor_info(); + + ITensorInfo *rhs_info_fuse = nullptr; + + ITensorInfo *ans_info = FunctionType::create_op(sketch, lhs_info, rhs_info); + + if (_fuse) + { + rhs_info_fuse = context.create_tensor_info(TensorInfo(shape2, 1, _data_type)); + ITensorInfo *ans2_info = FunctionType::create_op(sketch, ans_info, rhs_info_fuse); + GpuOutput::create_op(sketch, ans2_info, dst_info); + } + else + { + GpuOutput::create_op(sketch, ans_info, dst_info); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_lhs{}; + TensorType t_rhs{}; + TensorType t_rhs_fuse{}; + TensorType t_dst{}; + + // Initialize user tensors + t_lhs.allocator()->init(*lhs_info); + t_rhs.allocator()->init(*rhs_info); + t_dst.allocator()->init(*dst_info); + if (_fuse) + { + t_rhs_fuse.allocator()->init(*rhs_info_fuse); + } + + // Allocate and fill user tensors + // Instead of using ACL allocator, the user can choose to import memory into the tensors + t_lhs.allocator()->allocate(); + t_rhs.allocator()->allocate(); + t_dst.allocator()->allocate(); + if (_fuse) + { + t_rhs_fuse.allocator()->allocate(); + } + + fill(AccessorType(t_lhs), 0); + fill(AccessorType(t_rhs), 1); + if (_fuse) + { + fill(AccessorType(t_rhs_fuse), 2); + } + + // Run runtime + if (_fuse) + { + runtime.run({&t_lhs, &t_rhs, &t_rhs_fuse, &t_dst}); + } + else + { + runtime.run({&t_lhs, &t_rhs, &t_dst}); + } + + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) + { + // Create reference + SimpleTensor<T> ref_lhs{shape0, _data_type, 1, QuantizationInfo()}; + SimpleTensor<T> ref_rhs{shape1, _data_type, 1, QuantizationInfo()}; + SimpleTensor<T> ref_rhs_fuse{shape2, _data_type, 1, QuantizationInfo()}; + + // Fill reference + fill(ref_lhs, 0); + fill(ref_rhs, 1); + SimpleTensor<T> ref_dst = reference::pixel_wise_multiplication<T, T, T>( + ref_lhs, ref_rhs, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_UP, _data_type, + QuantizationInfo()); + if (_fuse) + { + fill(ref_rhs_fuse, 2); + SimpleTensor<T> ref_dst_fuse = reference::pixel_wise_multiplication<T, T, T>( + ref_dst, ref_rhs_fuse, 1.f, ConvertPolicy::SATURATE, RoundingPolicy::TO_NEAREST_UP, _data_type, + QuantizationInfo()); + return ref_dst_fuse; + } + return ref_dst; + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; + DataType _data_type{}; + bool _is_inplace{false}; + bool _fuse{false}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionMulOneOpValidationFixture + : public DynamicFusionMulValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape0, DataType data_type, bool is_inplace) + { + DynamicFusionMulValidationFixture<TensorType, AccessorType, FunctionType, T>::setup( + shape0, shape0, TensorShape(), data_type, is_inplace); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionMulBroadcastValidationFixture + : public DynamicFusionMulValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, bool is_inplace) + { + DynamicFusionMulValidationFixture<TensorType, AccessorType, FunctionType, T>::setup( + shape0, shape1, TensorShape(), data_type, is_inplace); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionMulTwoOpsValidationFixture + : public DynamicFusionMulValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(const TensorShape &shape0, + const TensorShape &shape1, + const TensorShape &shape2, + DataType data_type, + bool is_inplace, + bool fuse_two_ops) + { + DynamicFusionMulValidationFixture<TensorType, AccessorType, FunctionType, T>::setup( + shape0, shape1, shape2, data_type, is_inplace, fuse_two_ops); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_MULFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h new file mode 100644 index 0000000000..bde3360940 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ReshapeFixture.h @@ -0,0 +1,137 @@ +/* + * Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESHAPEFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESHAPEFIXTURE_H + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/ReshapeAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadContext.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuReshape.h" + +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/Globals.h" +#include "tests/validation/reference/ReshapeLayer.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionGpuReshapeLayerValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape input_shape, TensorShape output_shape, DataType data_type) + { + _target = compute_target(input_shape, output_shape, data_type); + _reference = compute_reference(input_shape, output_shape, data_type); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + library->fill_tensor_uniform(tensor, i); + } + + TensorType compute_target(TensorShape &input_shape, TensorShape &output_shape, DataType data_type) + { + // Check if indeed the input shape can be reshape to the output one + ARM_COMPUTE_ASSERT(input_shape.total_size() == output_shape.total_size()); + + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + ITensorInfo *src_info = context.create_tensor_info(TensorInfo(input_shape, 1, data_type)); + ITensorInfo *dst_info = context.create_tensor_info(TensorInfo(output_shape, 1, data_type)); + ReshapeAttributes attributes; + attributes.shape(output_shape); + + ITensorInfo *ans_info = FunctionType::create_op(sketch, src_info, attributes); + GpuOutput::create_op(sketch, ans_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst{}; + // Initialize user tensors + t_src.allocator()->init(*src_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_src), 0); + + // Run runtime + runtime.run({&t_src, &t_dst}); + + return t_dst; + } + + SimpleTensor<T> + compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, DataType data_type) + { + // Create reference + SimpleTensor<T> src{input_shape, data_type}; + + // Fill reference + fill(src, 0); + + return reference::reshape_layer<T>(src, output_shape); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; +/** [ReshapeLayer fixture] **/ +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESHAPEFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h new file mode 100644 index 0000000000..711767b66f --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/ResizeFixture.h @@ -0,0 +1,272 @@ +/* +* Copyright (c) 2022-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESIZEFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESIZEFIXTURE_H + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/ResizeAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/CL/CLAccessor.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/SimpleTensor.h" +#include "tests/validation/reference/Permute.h" +#include "tests/validation/reference/Scale.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionResizeGenericValidationFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape, + DataType data_type, + QuantizationInfo quantization_info, + DataLayout data_layout, + InterpolationPolicy interpolation_policy, + SamplingPolicy sampling_policy, + bool align_corners, + QuantizationInfo output_quantization_info) + { + _shape = shape; + _interpolation_policy = interpolation_policy; + _sampling_policy = sampling_policy; + _data_type = data_type; + _input_quantization_info = quantization_info; + _output_quantization_info = output_quantization_info; + _align_corners = align_corners; + _data_layout = data_layout; + + ARM_COMPUTE_ERROR_ON(data_layout != DataLayout::NHWC); // Dynamic fusion resize supports only NHWC layout + + generate_scale(shape); + + std::mt19937 generator(library->seed()); + std::uniform_int_distribution<uint32_t> distribution_u8(0, 255); + + _target = compute_target(shape); + _reference = compute_reference(shape); + } + +protected: + void generate_scale(const TensorShape &shape) + { + static constexpr float _min_scale{0.25f}; + static constexpr float _max_scale{3.f}; + + constexpr float max_width{8192.0f}; + constexpr float max_height{6384.0f}; + constexpr float min_width{1.f}; + constexpr float min_height{1.f}; + + std::mt19937 generator(library->seed()); + std::uniform_real_distribution<float> distribution_float(_min_scale, _max_scale); + + auto generate = [&](size_t input_size, float min_output, float max_output) -> int + { + const float generated_scale = distribution_float(generator); + const int output_size = static_cast<int>( + utility::clamp(static_cast<float>(input_size) * generated_scale, min_output, max_output)); + return output_size; + }; + + // Input shape is always given in NCHW layout. NHWC is dealt by permute in compute_target() + const int idx_width = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::HEIGHT); + + _output_width = generate(shape[idx_width], min_width, max_width); + _output_height = generate(shape[idx_height], min_height, max_height); + } + + template <typename U> + void fill(U &&tensor) + { + if (tensor.data_type() == DataType::F32) + { + std::uniform_real_distribution<float> distribution(-5.0f, 5.0f); + library->fill(tensor, distribution, 0); + } + else if (tensor.data_type() == DataType::F16) + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-5.0f, 5.0f}; + library->fill(tensor, distribution, 0); + } + else if (is_data_type_quantized(tensor.data_type())) + { + std::uniform_int_distribution<> distribution(0, 100); + library->fill(tensor, distribution, 0); + } + else + { + library->fill_tensor_uniform(tensor, 0); + } + } + + TensorType compute_target(TensorShape shape) + { + // Our test shapes are assumed in NCHW data layout, thus the permutation + permute(shape, PermutationVector(2U, 0U, 1U)); + + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + // Create sketch tensors + ITensorInfo *src_info = context.create_tensor_info(TensorInfo(shape, 1, _data_type, _data_layout)); + src_info->set_quantization_info(_input_quantization_info); + ITensorInfo *dst_info = context.create_tensor_info(); + + ResizeAttributes attributes; + attributes.align_corners(_align_corners) + .sampling_policy(_sampling_policy) + .interpolation_policy(_interpolation_policy) + .output_width(_output_width) + .output_height(_output_height); + + ITensorInfo *scale_result_info = FunctionType::create_op(sketch, src_info, attributes); + GpuOutput::create_op(sketch, scale_result_info, dst_info); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_src{}; + TensorType t_dst{}; + + // Initialize user tensors + t_src.allocator()->init(*src_info); + t_dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + t_src.allocator()->allocate(); + t_dst.allocator()->allocate(); + + fill(AccessorType(t_src)); + + // Run runtime + runtime.run({&t_src, &t_dst}); + + return t_dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape) + { + // Create reference + SimpleTensor<T> src{shape, _data_type, 1, _input_quantization_info}; + + // Reference code is NCHW, so the input shapes are NCHW + const int idx_width = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::WIDTH); + const int idx_height = get_data_layout_dimension_index(DataLayout::NCHW, DataLayoutDimension::HEIGHT); + + const float scale_x = static_cast<float>(_output_width) / shape[idx_width]; + const float scale_y = static_cast<float>(_output_height) / shape[idx_height]; + + // Fill reference + fill(src); + + return reference::scale<T>(src, scale_x, scale_y, _interpolation_policy, BorderMode::REPLICATE, + static_cast<T>(0), _sampling_policy, /* ceil_policy_scale */ false, _align_corners, + _output_quantization_info); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; + TensorShape _shape{}; + InterpolationPolicy _interpolation_policy{}; + SamplingPolicy _sampling_policy{}; + DataType _data_type{}; + DataLayout _data_layout{}; + QuantizationInfo _input_quantization_info{}; + QuantizationInfo _output_quantization_info{}; + bool _align_corners{false}; + int _output_width{0}; + int _output_height{0}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionResizeValidationFixture + : public DynamicFusionResizeGenericValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, + DataType data_type, + DataLayout data_layout, + InterpolationPolicy policy, + SamplingPolicy sampling_policy, + bool align_corners) + { + DynamicFusionResizeGenericValidationFixture<TensorType, AccessorType, FunctionType, T>::setup( + shape, data_type, QuantizationInfo(), data_layout, policy, sampling_policy, align_corners, + QuantizationInfo()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T, bool mixed_layout = false> +class DynamicFusionResizeQuantizedValidationFixture + : public DynamicFusionResizeGenericValidationFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, + DataType data_type, + QuantizationInfo quantization_info, + DataLayout data_layout, + InterpolationPolicy policy, + SamplingPolicy sampling_policy, + bool align_corners) + { + DynamicFusionResizeGenericValidationFixture<TensorType, AccessorType, FunctionType, T>::setup( + shape, data_type, quantization_info, data_layout, policy, sampling_policy, align_corners, + quantization_info); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_RESIZEFIXTURE_H diff --git a/tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h new file mode 100644 index 0000000000..175d4ff889 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/SoftmaxFixture.h @@ -0,0 +1,158 @@ +/* +* Copyright (c) 2023-2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_SOFTMAXFIXTURE_H +#define ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_SOFTMAXFIXTURE_H + +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/attributes/SoftmaxAttributes.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" + +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/SimpleTensor.h" +#include "tests/validation/reference/SoftmaxLayer.h" +#include "tests/validation/Validation.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionSoftmaxValidationGenericFixture : public framework::Fixture +{ +public: + void setup(TensorShape shape, DataType data_type, float beta, size_t axis, bool is_log) + { + _reference = compute_reference(shape, data_type, beta, axis, is_log); + _target = compute_target(shape, data_type, beta, axis, is_log); + } + +protected: + template <typename U> + void fill(U &&tensor) + { + if (tensor.data_type() == DataType::F32) + { + std::uniform_real_distribution<float> distribution(-10.0f, 10.0f); + library->fill(tensor, distribution, 0); + } + else if (tensor.data_type() == DataType::F16) + { + arm_compute::utils::uniform_real_distribution_16bit<half> distribution{-10.0f, 10.0f}; + library->fill(tensor, distribution, 0); + } + else if (!is_data_type_quantized(tensor.data_type())) + { + std::uniform_int_distribution<> distribution(0, 100); + library->fill(tensor, distribution, 0); + } + else + { + library->fill_tensor_uniform(tensor, 0); + } + } + + TensorType compute_target(const TensorShape &shape, DataType data_type, float beta, int32_t axis, bool is_log) + { + // Create a new workload sketch + CLCompileContext cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + GpuWorkloadContext context = GpuWorkloadContext{&cl_compile_ctx}; + GpuWorkloadSketch sketch{&context}; + + SoftmaxAttributes softmax_attr{}; + softmax_attr.axis(axis).beta(beta).is_log_softmax(is_log); + ITensorInfo *src_info = context.create_tensor_info(shape, 1, data_type); + ITensorInfo *dst_info = context.create_tensor_info(shape, 1, data_type); + FunctionType::create_op(sketch, src_info, dst_info, softmax_attr); + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + // Instead of using ACL allocated memory, the user can choose to import memory into the tensors + for (auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + // Construct user tensors + TensorType src{}; + TensorType dst{}; + + // Initialize user tensors + src.allocator()->init(*src_info); + dst.allocator()->init(*dst_info); + + // Allocate and fill user tensors + src.allocator()->allocate(); + dst.allocator()->allocate(); + fill(AccessorType(src)); + + // Run runtime + runtime.run({&src, &dst}); + + return dst; + } + + SimpleTensor<T> + compute_reference(const TensorShape &shape, DataType data_type, float beta, int32_t axis, bool is_log) + { + // Create reference + SimpleTensor<T> src{shape, data_type, 1}; + + // Fill reference + fill(src); + + return reference::softmax_layer<T>(src, beta, axis, is_log); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DynamicFusionSoftmaxValidationFixture + : public DynamicFusionSoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + void setup(TensorShape shape, DataType data_type, float beta, size_t axis, bool is_log) + { + DynamicFusionSoftmaxValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup( + shape, data_type, beta, axis, is_log); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute + +#endif // ACL_TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_SOFTMAXFIXTURE_H diff --git a/tests/validation/reference/ActivationLayer.cpp b/tests/validation/reference/ActivationLayer.cpp index 664b969125..2172362bdd 100644 --- a/tests/validation/reference/ActivationLayer.cpp +++ b/tests/validation/reference/ActivationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020,2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #include "ActivationLayer.h" #include "arm_compute/core/Types.h" + #include "tests/validation/Helpers.h" namespace arm_compute @@ -40,7 +41,7 @@ SimpleTensor<T> activation_layer(const SimpleTensor<T> &src, ActivationLayerInfo ARM_COMPUTE_UNUSED(oq_info); // Create reference - SimpleTensor<T> dst{ src.shape(), src.data_type(), 1 }; + SimpleTensor<T> dst{src.shape(), src.data_type(), 1}; // Compute reference const T a(info.a()); @@ -48,7 +49,7 @@ SimpleTensor<T> activation_layer(const SimpleTensor<T> &src, ActivationLayerInfo #if defined(_OPENMP) #pragma omp parallel for #endif /* _OPENMP */ - for(int i = 0; i < src.num_elements(); ++i) + for (int i = 0; i < src.num_elements(); ++i) { dst[i] = activate_float<T>(src[i], a, b, info.activation()); } @@ -57,7 +58,8 @@ SimpleTensor<T> activation_layer(const SimpleTensor<T> &src, ActivationLayerInfo } template <> -SimpleTensor<uint8_t> activation_layer<uint8_t>(const SimpleTensor<uint8_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info) +SimpleTensor<uint8_t> +activation_layer<uint8_t>(const SimpleTensor<uint8_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info) { const QuantizationInfo dst_qinfo = oq_info.empty() ? src.quantization_info() : oq_info; @@ -68,7 +70,8 @@ SimpleTensor<uint8_t> activation_layer<uint8_t>(const SimpleTensor<uint8_t> &src } template <> -SimpleTensor<int8_t> activation_layer<int8_t>(const SimpleTensor<int8_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info) +SimpleTensor<int8_t> +activation_layer<int8_t>(const SimpleTensor<int8_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info) { const QuantizationInfo dst_qinfo = oq_info.empty() ? src.quantization_info() : oq_info; @@ -79,7 +82,8 @@ SimpleTensor<int8_t> activation_layer<int8_t>(const SimpleTensor<int8_t> &src, A } template <> -SimpleTensor<int16_t> activation_layer<int16_t>(const SimpleTensor<int16_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info) +SimpleTensor<int16_t> +activation_layer<int16_t>(const SimpleTensor<int16_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info) { const QuantizationInfo dst_qinfo = oq_info.empty() ? src.quantization_info() : oq_info; @@ -88,9 +92,14 @@ SimpleTensor<int16_t> activation_layer<int16_t>(const SimpleTensor<int16_t> &src SimpleTensor<int16_t> dst = convert_to_symmetric<int16_t>(dst_tmp, dst_qinfo); return dst; } -template SimpleTensor<int32_t> activation_layer(const SimpleTensor<int32_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); -template SimpleTensor<float> activation_layer(const SimpleTensor<float> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); -template SimpleTensor<half> activation_layer(const SimpleTensor<half> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); +template SimpleTensor<int32_t> +activation_layer(const SimpleTensor<int32_t> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); +template SimpleTensor<float> +activation_layer(const SimpleTensor<float> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); +template SimpleTensor<half> +activation_layer(const SimpleTensor<half> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); +template SimpleTensor<bfloat16> +activation_layer(const SimpleTensor<bfloat16> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ActivationLayer.h b/tests/validation/reference/ActivationLayer.h index 8aad1af63e..7f896bd696 100644 --- a/tests/validation/reference/ActivationLayer.h +++ b/tests/validation/reference/ActivationLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020,2022,2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_ACTIVATION_LAYER_H -#define ARM_COMPUTE_TEST_ACTIVATION_LAYER_H +#ifndef ACL_TESTS_VALIDATION_REFERENCE_ACTIVATIONLAYER_H +#define ACL_TESTS_VALIDATION_REFERENCE_ACTIVATIONLAYER_H #include "tests/SimpleTensor.h" #include "tests/validation/Helpers.h" @@ -40,7 +40,7 @@ inline T activate_float(T x, T a, T b, ActivationLayerInfo::ActivationFunction a { T ret; - switch(activation) + switch (activation) { case ActivationLayerInfo::ActivationFunction::ABS: ret = std::abs(x); @@ -61,13 +61,13 @@ inline T activate_float(T x, T a, T b, ActivationLayerInfo::ActivationFunction a ret = std::min<T>(a, std::max<T>(b, x)); break; case ActivationLayerInfo::ActivationFunction::LEAKY_RELU: - ret = (x > 0) ? x : a * x; + ret = x > static_cast<T>(0) ? x : static_cast<T>(a * x); break; case ActivationLayerInfo::ActivationFunction::SOFT_RELU: - ret = std::log(static_cast<T>(1) + std::exp(x)); + ret = std::log(static_cast<T>(1) + std::exp(static_cast<double>(x))); break; case ActivationLayerInfo::ActivationFunction::ELU: - ret = (x > 0) ? x : a * (std::exp(x) - static_cast<T>(1)); + ret = x > static_cast<T>(0) ? x : static_cast<T>(a * (std::exp(x) - static_cast<T>(1))); break; case ActivationLayerInfo::ActivationFunction::SQRT: ret = std::sqrt(x); @@ -82,7 +82,14 @@ inline T activate_float(T x, T a, T b, ActivationLayerInfo::ActivationFunction a ret = x; break; case ActivationLayerInfo::ActivationFunction::HARD_SWISH: - ret = x * ((std::min(std::max(static_cast<T>(x + 3), static_cast<T>(0.0f)), static_cast<T>(6.0f))) * 0.166666667f); + ret = x * ((std::min(std::max(static_cast<T>(x + 3), static_cast<T>(0.0f)), static_cast<T>(6.0f))) * + 0.166666667f); + break; + case ActivationLayerInfo::ActivationFunction::SWISH: + ret = static_cast<T>(x) / (static_cast<T>(1) + std::exp(-a * x)); + break; + case ActivationLayerInfo::ActivationFunction::GELU: + ret = x * 0.5f * (1 + erf(x / std::sqrt(2.0f))); break; default: ARM_COMPUTE_ERROR("Unsupported activation function"); @@ -93,9 +100,11 @@ inline T activate_float(T x, T a, T b, ActivationLayerInfo::ActivationFunction a } template <typename T> -SimpleTensor<T> activation_layer(const SimpleTensor<T> &src, ActivationLayerInfo info, const QuantizationInfo &oq_info = QuantizationInfo()); +SimpleTensor<T> activation_layer(const SimpleTensor<T> &src, + ActivationLayerInfo info, + const QuantizationInfo &oq_info = QuantizationInfo()); } // namespace reference } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_ACTIVATION_LAYER_H */ +#endif // ACL_TESTS_VALIDATION_REFERENCE_ACTIVATIONLAYER_H diff --git a/tests/validation/reference/BatchToSpaceLayer.cpp b/tests/validation/reference/BatchToSpaceLayer.cpp index 404ee73cac..63d121f59b 100644 --- a/tests/validation/reference/BatchToSpaceLayer.cpp +++ b/tests/validation/reference/BatchToSpaceLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018 Arm Limited. + * Copyright (c) 2018, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,8 +23,10 @@ */ #include "BatchToSpaceLayer.h" +#include "arm_compute/core/Validate.h" #include "tests/validation/Helpers.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" namespace arm_compute { namespace test @@ -35,32 +37,37 @@ namespace reference { // Batch to Space template <typename T> -SimpleTensor<T> batch_to_space(const SimpleTensor<T> &src, const SimpleTensor<int32_t> &block_shape, const TensorShape &dst_shape) +SimpleTensor<T> batch_to_space(const SimpleTensor<T> &src, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, const TensorShape &dst_shape) { - ARM_COMPUTE_ERROR_ON(block_shape[0] <= 0); - ARM_COMPUTE_ERROR_ON(block_shape[1] <= 0); - SimpleTensor<T> result(dst_shape, src.data_type()); + ARM_COMPUTE_ERROR_ON(block_shape[0] < 1); + ARM_COMPUTE_ERROR_ON(block_shape[1] < 1); + const auto expected_dst_shape = misc::shape_calculator::compute_batch_to_space_shape(DataLayout::NCHW, src.shape(), block_shape[0], block_shape[1], crop_info); + ARM_COMPUTE_ERROR_ON(arm_compute::detail::have_different_dimensions(expected_dst_shape, dst_shape, 0)); + ARM_COMPUTE_UNUSED(expected_dst_shape); - int in_pos = 0; - const auto width_in = static_cast<int>(src.shape()[0]); - const auto height_in = static_cast<int>(src.shape()[1]); - const auto z_in = static_cast<int>(src.shape()[2]); - const auto batch_in = static_cast<int>(src.shape()[3]); + SimpleTensor<T> result(dst_shape, src.data_type()); + int out_pos = 0; + const auto width_out = static_cast<int>(dst_shape[0]); + const auto height_out = static_cast<int>(dst_shape[1]); + const auto z_out = static_cast<int>(dst_shape[2]); + const auto batch_out = static_cast<int>(dst_shape[3]); - for(int batch = 0; batch < batch_in; ++batch) + for(int batch = 0; batch < batch_out; ++batch) { - for(int z = 0; z < z_in; ++z) + for(int z = 0; z < z_out; ++z) { - for(int y = 0; y < height_in; ++y) + for(int y = 0; y < height_out; ++y) { - for(int x = 0; x < width_in; ++x) + for(int x = 0; x < width_out; ++x) { - const int r = src.shape()[3] / (block_shape[0] * block_shape[1]); - const int out_x = (block_shape[0] * x + (batch / r) % block_shape[0]); - const int out_y = (block_shape[1] * y + (batch / r) / block_shape[0]); - const int out_pos = out_x + dst_shape[0] * out_y + z * dst_shape[0] * dst_shape[1] + (batch % r) * dst_shape[0] * dst_shape[1] * dst_shape[2]; - result[out_pos] = src[in_pos]; - ++in_pos; + const int x_c = x + crop_info.left; + const int y_c = y + crop_info.top; + const int in_batch = batch + ((x_c % block_shape[0]) + (y_c % block_shape[1]) * (block_shape[0])) * dst_shape[3]; + const int in_x = x_c / block_shape[0]; + const int in_y = y_c / block_shape[1]; + const int in_pos = in_x + src.shape()[0] * in_y + z * src.shape()[0] * src.shape()[1] + in_batch * src.shape()[0] * src.shape()[1] * src.shape()[2]; + result[out_pos] = src[in_pos]; + ++out_pos; } } } @@ -68,8 +75,8 @@ SimpleTensor<T> batch_to_space(const SimpleTensor<T> &src, const SimpleTensor<in return result; } -template SimpleTensor<float> batch_to_space(const SimpleTensor<float> &src, const SimpleTensor<int32_t> &block_shape, const TensorShape &dst_shape); -template SimpleTensor<half> batch_to_space(const SimpleTensor<half> &src, const SimpleTensor<int32_t> &block_shape, const TensorShape &dst_shape); +template SimpleTensor<float> batch_to_space(const SimpleTensor<float> &src, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, const TensorShape &dst_shape); +template SimpleTensor<half> batch_to_space(const SimpleTensor<half> &src, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, const TensorShape &dst_shape); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/BatchToSpaceLayer.h b/tests/validation/reference/BatchToSpaceLayer.h index 52556cb53f..a37bfc3373 100644 --- a/tests/validation/reference/BatchToSpaceLayer.h +++ b/tests/validation/reference/BatchToSpaceLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 Arm Limited. + * Copyright (c) 2018-2019, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #ifndef ARM_COMPUTE_TEST_BATCH_TO_SPACE_LAYER_H #define ARM_COMPUTE_TEST_BATCH_TO_SPACE_LAYER_H +#include "arm_compute/core/Types.h" #include "tests/SimpleTensor.h" #include "tests/validation/Helpers.h" @@ -36,7 +37,7 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> batch_to_space(const SimpleTensor<T> &src, const SimpleTensor<int32_t> &block_shape, const TensorShape &dst_shape); +SimpleTensor<T> batch_to_space(const SimpleTensor<T> &src, const std::vector<int32_t> &block_shape, const CropInfo &crop_info, const TensorShape &dst_shape); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Conv3D.cpp b/tests/validation/reference/Conv3D.cpp new file mode 100644 index 0000000000..e4010a507a --- /dev/null +++ b/tests/validation/reference/Conv3D.cpp @@ -0,0 +1,260 @@ +/* + * Copyright (c) 2021, 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "Conv3D.h" + +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/quantization/AsymmHelpers.h" +#include "support/AclRequires.h" +#include "tests/validation/reference/UtilsQuantizedAsymm.h" + +// Source/Destination Tensor shape indices (N D H W C) +constexpr unsigned int batch_dim = 4u; +constexpr unsigned int depth_dim = 3u; +constexpr unsigned int height_dim = 2u; +constexpr unsigned int width_dim = 1u; +constexpr unsigned int channel_dim = 0u; + +// Weight tensor shape indices (D H W Cin Cout) +constexpr unsigned int weights_depth_dim = 4u; +constexpr unsigned int weights_height_dim = 3u; +constexpr unsigned int weights_width_dim = 2u; +constexpr unsigned int weights_CHin_dim = 1u; +constexpr unsigned int weights_CHout_dim = 0u; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +namespace +{ +inline bool is_valid_pixel(int i, int min, int max) +{ + return (i >= min && i < max); +} + +// Evaluate the weights against an element in a given tensor. +template < typename T, typename TB, typename std::enable_if < validation::is_floating_point<T>::value &&validation::is_floating_point<TB>::value, int >::type = 0 > +T calculate_conv3d(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const Size3D &dilation, int batch, + int z_start, int y_start, int x_start, int ch_out, UniformQuantizationInfo oq_info) +{ + ARM_COMPUTE_UNUSED(oq_info); + + const unsigned int weights_width = weights.shape()[weights_width_dim]; + const unsigned int weights_height = weights.shape()[weights_height_dim]; + const unsigned int weights_depth = weights.shape()[weights_depth_dim]; + + const unsigned int src_channels = src.shape()[channel_dim]; + const unsigned int src_width = src.shape()[width_dim]; + const unsigned int src_height = src.shape()[height_dim]; + const unsigned int src_depth = src.shape()[depth_dim]; + + T total(0); + for(unsigned int weight_d = 0; weight_d < weights_depth; ++weight_d) + { + const int idx_z = z_start + dilation.depth * weight_d; + for(unsigned int weight_y = 0; weight_y < weights_height; ++weight_y) + { + const int idx_y = y_start + dilation.height * weight_y; + for(unsigned int weight_x = 0; weight_x < weights_width; ++weight_x) + { + const int idx_x = x_start + dilation.width * weight_x; + + //Check if the point is within padding + const bool is_x_valid = is_valid_pixel(idx_x, 0, src_width); + const bool is_y_valid = is_valid_pixel(idx_y, 0, src_height); + const bool is_z_valid = is_valid_pixel(idx_z, 0, src_depth); + const bool is_invalid_pixel = !(is_x_valid && is_y_valid && is_z_valid); + if(is_invalid_pixel) + { + continue; + } + + for(unsigned int ch_in = 0; ch_in < src_channels; ++ch_in) + { + const T *in_ptr = src.data(); + const T *w_ptr = weights.data(); + + const int in_offset = coord2index(src.shape(), Coordinates{ ch_in, idx_x, idx_y, idx_z, batch }); + const int weight_offset = coord2index(weights.shape(), Coordinates{ ch_out, ch_in, weight_x, weight_y, weight_d }); + T input_value = in_ptr[in_offset]; + T weight_value = w_ptr[weight_offset]; + total += (input_value * weight_value); + } + } + } + } + + const TB *b_ptr = bias.data(); + TB bias_value = b_ptr[ch_out]; + + return total + bias_value; +} + +template < typename T, typename TB, ARM_COMPUTE_REQUIRES_TA(std::is_same<T, uint8_t>::value || std::is_same<T, int8_t>::value) > +T calculate_conv3d(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, const Size3D &dilation, int batch, + int z_start, int y_start, int x_start, int ch_out, UniformQuantizationInfo oq_info) +{ + const unsigned int weights_width = weights.shape()[weights_width_dim]; + const unsigned int weights_height = weights.shape()[weights_height_dim]; + const unsigned int weights_depth = weights.shape()[weights_depth_dim]; + + const unsigned int src_channels = src.shape()[channel_dim]; + const unsigned int src_width = src.shape()[width_dim]; + const unsigned int src_height = src.shape()[height_dim]; + const unsigned int src_depth = src.shape()[depth_dim]; + + const UniformQuantizationInfo iq_info = src.quantization_info().uniform(); + const UniformQuantizationInfo wq_info = weights.quantization_info().uniform(); + + const int input_offset = -iq_info.offset; + const float input_scale = iq_info.scale; + int weights_offset = -wq_info.offset; + float weights_scale = wq_info.scale; + const int output_offset = oq_info.offset; + const float output_scale = oq_info.scale; + + int output_multiplier = 0; + int output_shift = 0; + const float multiplier = input_scale * weights_scale / output_scale; + arm_compute::quantization::calculate_quantized_multiplier(multiplier, &output_multiplier, &output_shift); + + int32_t total(0); + for(unsigned int weight_d = 0; weight_d < weights_depth; ++weight_d) + { + const int idx_z = z_start + dilation.depth * weight_d; + for(unsigned int weight_y = 0; weight_y < weights_height; ++weight_y) + { + const int idx_y = y_start + dilation.height * weight_y; + for(unsigned int weight_x = 0; weight_x < weights_width; ++weight_x) + { + const int idx_x = x_start + dilation.width * weight_x; + + //Check if the point is within padding + const bool is_x_valid = is_valid_pixel(idx_x, 0, src_width); + const bool is_y_valid = is_valid_pixel(idx_y, 0, src_height); + const bool is_z_valid = is_valid_pixel(idx_z, 0, src_depth); + const bool is_invalid_pixel = !(is_x_valid && is_y_valid && is_z_valid); + if(is_invalid_pixel) + { + continue; + } + + for(unsigned int ch_in = 0; ch_in < src_channels; ++ch_in) + { + const T *in_ptr = src.data(); + const T *w_ptr = weights.data(); + + const int in_offset = coord2index(src.shape(), Coordinates{ ch_in, idx_x, idx_y, idx_z, batch }); + const int weight_offset = coord2index(weights.shape(), Coordinates{ ch_out, ch_in, weight_x, weight_y, weight_d }); + T input_value = in_ptr[in_offset]; + T weight_value = w_ptr[weight_offset]; + total += ((input_value + input_offset) * (weight_value + weights_offset)); + } + } + } + } + + const TB *b_ptr = bias.data(); + TB bias_value = b_ptr[ch_out]; + + total += bias_value; + + return validation::quantize_down_scale_by_fixedpoint(total, output_multiplier, output_shift, output_offset, + std::numeric_limits<T>::lowest(), std::numeric_limits<T>::max()); +} +} // namespace + +template <typename T, typename TB> +SimpleTensor<T> conv3d(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, const Conv3dInfo &conv3d_info) +{ + // Compute reference + const unsigned int batch_size = src.shape()[batch_dim]; + const unsigned int dst_width = dst.shape()[width_dim]; + const unsigned int dst_height = dst.shape()[height_dim]; + const unsigned int dst_depth = dst.shape()[depth_dim]; + const unsigned int src_channels = src.shape()[channel_dim]; + const unsigned int weights_out_ch = weights.shape()[weights_CHout_dim]; + const unsigned int dst_channels = dst.shape()[channel_dim]; + const size_t pad_left = conv3d_info.padding.left; + const size_t pad_top = conv3d_info.padding.top; + const size_t pad_front = conv3d_info.padding.front; + const size_t stride_x = conv3d_info.stride.x(); + const size_t stride_y = conv3d_info.stride.y(); + const size_t stride_z = conv3d_info.stride.z(); + + const TensorShape dst_shape = arm_compute::misc::shape_calculator::compute_conv3d_shape(src.shape(), weights.shape(), conv3d_info); + + ARM_COMPUTE_UNUSED(src_channels, weights_out_ch, dst_channels, dst_shape, weights_CHin_dim); + // Number of batches of source and destination tensors must match. + ARM_COMPUTE_ERROR_ON(src.shape()[batch_dim] != dst.shape()[batch_dim]); + // Input channels in the source and weights must match. + ARM_COMPUTE_ERROR_ON(src_channels != weights.shape()[weights_CHin_dim]); + // Weight channels in the destination and weights must match. + ARM_COMPUTE_ERROR_ON(weights_out_ch != dst_channels); + // Bias must match the number of destination channels. + ARM_COMPUTE_ERROR_ON(bias.shape()[0] != dst_channels); + // Compare given dst tensor shape with expected shape. + ARM_COMPUTE_ERROR_ON(dst.shape() != dst_shape); + + for(unsigned int batch = 0; batch < batch_size; ++batch) + { + for(unsigned int z_out = 0; z_out < dst_depth; ++z_out) + { + const int z_start = (z_out * stride_z) - pad_front; + for(unsigned int y_out = 0; y_out < dst_height; ++y_out) + { + const int y_start = (y_out * stride_y) - pad_top; + for(unsigned int x_out = 0; x_out < dst_width; ++x_out) + { + const int x_start = (x_out * stride_x) - pad_left; + for(unsigned int ch_out = 0; ch_out < dst_channels; ++ch_out) + { + T *out_ptr = dst.data(); + + const int out_offset = coord2index(dst.shape(), Coordinates{ ch_out, x_out, y_out, z_out, batch }); + out_ptr[out_offset] = calculate_conv3d<T, TB>(src, weights, bias, conv3d_info.dilation, batch, z_start, y_start, x_start, ch_out, dst.quantization_info().uniform()); + } + } + } + } + } + return dst; +} + +template SimpleTensor<float> conv3d(const SimpleTensor<float> &src, const SimpleTensor<float> &weights, const SimpleTensor<float> &bias, SimpleTensor<float> &dst, + const Conv3dInfo &conv3d_info); +template SimpleTensor<half> conv3d(const SimpleTensor<half> &src, const SimpleTensor<half> &weights, const SimpleTensor<half> &bias, SimpleTensor<half> &dst, + const Conv3dInfo &conv3d_info); +template SimpleTensor<uint8_t> conv3d(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &weights, const SimpleTensor<int32_t> &bias, SimpleTensor<uint8_t> &dst, + const Conv3dInfo &conv3d_info); +template SimpleTensor<int8_t> conv3d(const SimpleTensor<int8_t> &src, const SimpleTensor<int8_t> &weights, const SimpleTensor<int32_t> &bias, SimpleTensor<int8_t> &dst, + const Conv3dInfo &conv3d_info); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/Conv3D.h b/tests/validation/reference/Conv3D.h new file mode 100644 index 0000000000..e3674f4bfb --- /dev/null +++ b/tests/validation/reference/Conv3D.h @@ -0,0 +1,47 @@ +/* + * Copyright (c) 2021 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_TEST_CONV3D_LAYER_H +#define ARM_COMPUTE_TEST_CONV3D_LAYER_H + +#include "Utils.h" +#include "arm_compute/runtime/FunctionDescriptors.h" +#include "tests/SimpleTensor.h" +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T, typename TB> +SimpleTensor<T> conv3d(const SimpleTensor<T> &src, const SimpleTensor<T> &weights, const SimpleTensor<TB> &bias, SimpleTensor<T> &dst, + const Conv3dInfo &conv3d_info); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_CONV3D_LAYER_H */ diff --git a/tests/validation/reference/Convolution3d.h b/tests/validation/reference/Convolution3d.h index 1666e3857b..b67e88e839 100644 --- a/tests/validation/reference/Convolution3d.h +++ b/tests/validation/reference/Convolution3d.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,7 +25,7 @@ #define ARM_COMPUTE_TEST_VALIDATION_CONVOLUTION_H #include "arm_compute/core/utils/quantization/AsymmHelpers.h" -#include "support/Requires.h" +#include "support/AclRequires.h" #include "tests/validation/Helpers.h" #include "tests/validation/reference/UtilsQuantizedAsymm.h" diff --git a/tests/validation/reference/DFT.cpp b/tests/validation/reference/DFT.cpp index fd126c7d73..2b03c270ac 100644 --- a/tests/validation/reference/DFT.cpp +++ b/tests/validation/reference/DFT.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2020 Arm Limited. + * Copyright (c) 2019-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -400,10 +400,10 @@ SimpleTensor<T> conv2d_dft(const SimpleTensor<T> &src, const SimpleTensor<T> &w, auto padded_src = pad_layer(src, padding_in); // Flip weights - std::vector<uint32_t> axis_v = { 0, 1 }; - SimpleTensor<uint32_t> axis{ TensorShape(2U), DataType::U32 }; + std::vector<uint32_t> axis_v = { 0, 1 }; + SimpleTensor<int32_t> axis{ TensorShape(2U), DataType::S32 }; std::copy(axis_v.begin(), axis_v.begin() + axis.shape().x(), axis.data()); - auto flipped_w = reverse(w, axis); + auto flipped_w = reverse(w, axis, /* use_inverted_axis */ false); // Pad weights to have the same size as input const PaddingList paddings_w = { { 0, src.shape()[0] - 1 }, { 0, src.shape()[1] - 1 } }; diff --git a/tests/validation/reference/DepthConvertLayer.cpp b/tests/validation/reference/DepthConvertLayer.cpp index 94c719ade7..3f88897f8e 100644 --- a/tests/validation/reference/DepthConvertLayer.cpp +++ b/tests/validation/reference/DepthConvertLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -165,7 +165,7 @@ template SimpleTensor<half> depth_convert(const SimpleTensor<int32_t> &src, Data template SimpleTensor<float> depth_convert(const SimpleTensor<int32_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); // BFLOAT16 -template SimpleTensor<float> depth_convert(const SimpleTensor<bfloat16> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<bfloat16> depth_convert(const SimpleTensor<bfloat16> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); // F16 template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<half> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); @@ -186,6 +186,25 @@ template SimpleTensor<int32_t> depth_convert(const SimpleTensor<float> &src, Dat template SimpleTensor<half> depth_convert(const SimpleTensor<float> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); template SimpleTensor<bfloat16> depth_convert(const SimpleTensor<float> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +// S64 +template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<int8_t> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<uint16_t> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<int16_t> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<uint32_t> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<int32_t> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<half> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<float> depth_convert(const SimpleTensor<int64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); + +// U64 +template SimpleTensor<uint8_t> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<int8_t> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<uint16_t> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<int16_t> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<uint32_t> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<int32_t> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<half> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); +template SimpleTensor<float> depth_convert(const SimpleTensor<uint64_t> &src, DataType dt_out, ConvertPolicy policy, uint32_t shift); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/DequantizationLayer.cpp b/tests/validation/reference/DequantizationLayer.cpp index 64a89aa6a0..67d69c2c38 100644 --- a/tests/validation/reference/DequantizationLayer.cpp +++ b/tests/validation/reference/DequantizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -59,6 +59,12 @@ TOut dequantize(int16_t val, const UniformQuantizationInfo qinfo, DataType dt) ARM_COMPUTE_UNUSED(dt); return static_cast<TOut>(dequantize_qsymm16(val, qinfo)); } +template <typename TOut> +TOut dequantize(int32_t val, const UniformQuantizationInfo qinfo, DataType dt) +{ + ARM_COMPUTE_UNUSED(dt); + return static_cast<TOut>(dequantize_s32(val, qinfo)); +} } // namespace template <typename TOut, typename TIn> SimpleTensor<TOut> dequantization_layer(const SimpleTensor<TIn> &src) @@ -115,6 +121,7 @@ template SimpleTensor<half> dequantization_layer(const SimpleTensor<int8_t> &src template SimpleTensor<float> dequantization_layer(const SimpleTensor<int8_t> &src); template SimpleTensor<half> dequantization_layer(const SimpleTensor<int16_t> &src); template SimpleTensor<float> dequantization_layer(const SimpleTensor<int16_t> &src); +template SimpleTensor<float> dequantization_layer(const SimpleTensor<int32_t> &src); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ElementwiseOperations.cpp b/tests/validation/reference/ElementwiseOperations.cpp index f22c84e153..edbbab8600 100644 --- a/tests/validation/reference/ElementwiseOperations.cpp +++ b/tests/validation/reference/ElementwiseOperations.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -74,15 +74,6 @@ T arithm_op(ArithmeticOperation op, T src1, T src2, ConvertPolicy convert_policy case ArithmeticOperation::DIV: { val = (static_cast<intermediate_type>(src1) / static_cast<intermediate_type>(src2)); - if(std::is_integral<T>::value) - { - // Implement flooring division - val = (src2 == 0) ? 0 : val; - if(static_cast<int32_t>(src1) % static_cast<int32_t>(src2) != 0 && ((src1 < 0) != (src2 < 0))) - { - --val; - } - } break; } case ArithmeticOperation::POWER: diff --git a/tests/validation/reference/ElementwiseUnary.cpp b/tests/validation/reference/ElementwiseUnary.cpp index 5333b53c15..558f9d24fc 100644 --- a/tests/validation/reference/ElementwiseUnary.cpp +++ b/tests/validation/reference/ElementwiseUnary.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,7 +22,8 @@ * SOFTWARE. */ #include "ElementwiseUnary.h" - +#include "tests/validation/Helpers.h" +#include "utils/TypePrinter.h" namespace arm_compute { namespace test @@ -32,10 +33,8 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> elementwise_unary(const SimpleTensor<T> &src, ElementWiseUnary op) +SimpleTensor<T> elementwise_unary(const SimpleTensor<T> &src, SimpleTensor<T> &dst, ElementWiseUnary op) { - SimpleTensor<T> dst(src.shape(), src.data_type()); - for(int i = 0; i < src.num_elements(); ++i) { switch(op) @@ -65,13 +64,107 @@ SimpleTensor<T> elementwise_unary(const SimpleTensor<T> &src, ElementWiseUnary o ARM_COMPUTE_ERROR("Not implemented"); } } + return dst; +} +template <> +SimpleTensor<int8_t> elementwise_unary(const SimpleTensor<int8_t> &src, SimpleTensor<int8_t> &dst, ElementWiseUnary op) +{ + if(dst.data_type() == DataType::QASYMM8_SIGNED) + { + SimpleTensor<float> src_tmp = convert_from_asymmetric(src); + SimpleTensor<float> dst_tmp(src.shape(), DataType::F32); + for(int i = 0; i < src.num_elements(); ++i) + { + switch(op) + { + case ElementWiseUnary::RSQRT: + if(src_tmp[i] != 0) + { + dst_tmp[i] = 1.f / std::sqrt(src_tmp[i]); + } + else + { + // rsqrt(0) give 'inf' so set to the maximum in int8: 127 + dst_tmp[i] = (127.0f - dst.quantization_info().uniform().offset) * dst.quantization_info().uniform().scale ; + } + break; + + case ElementWiseUnary::LOG: + if(src_tmp[i] != 0) + { + dst_tmp[i] = std::log(src_tmp[i]); + } + else + { + dst_tmp[i] = (-128.0f - dst.quantization_info().uniform().offset) * dst.quantization_info().uniform().scale ; + } + break; + + default: + elementwise_unary(src_tmp, dst_tmp, op); + break; + } + } + dst = convert_to_asymmetric<int8_t>(dst_tmp, dst.quantization_info()); + } + else + { + ARM_COMPUTE_ERROR("Not implemented"); + } + return dst; +} +template <> +SimpleTensor<uint8_t> elementwise_unary(const SimpleTensor<uint8_t> &src, SimpleTensor<uint8_t> &dst, ElementWiseUnary op) +{ + if(dst.data_type() == DataType::QASYMM8) + { + SimpleTensor<float> src_tmp = convert_from_asymmetric(src); + SimpleTensor<float> dst_tmp(src.shape(), DataType::F32); + for(int i = 0; i < src.num_elements(); ++i) + { + switch(op) + { + case ElementWiseUnary::RSQRT: + if(src_tmp[i] != 0) + { + dst_tmp[i] = 1.f / std::sqrt(src_tmp[i]); + } + else + { + // rsqrt(0) give 'inf' so set to the maximum in uint8: 255 + dst_tmp[i] = (255.0f - dst.quantization_info().uniform().offset)* dst.quantization_info().uniform().scale; + } + break; + case ElementWiseUnary::LOG: + if(src_tmp[i] != 0) + { + dst_tmp[i] = std::log(src_tmp[i]); + } + else + { + dst_tmp[i] = -dst.quantization_info().uniform().offset * dst.quantization_info().uniform().scale; + } + break; + + default: + elementwise_unary(src_tmp, dst_tmp, op); + break; + } + } + dst = convert_to_asymmetric<uint8_t>(dst_tmp, dst.quantization_info()); + } + else + { + ARM_COMPUTE_ERROR("Not implemented"); + } return dst; } -template SimpleTensor<float> elementwise_unary(const SimpleTensor<float> &src, ElementWiseUnary op); -template SimpleTensor<half> elementwise_unary(const SimpleTensor<half> &src, ElementWiseUnary op); -template SimpleTensor<int32_t> elementwise_unary(const SimpleTensor<int32_t> &src, ElementWiseUnary op); +template SimpleTensor<float> elementwise_unary(const SimpleTensor<float> &src, SimpleTensor<float> &dst, ElementWiseUnary op); +template SimpleTensor<half> elementwise_unary(const SimpleTensor<half> &src, SimpleTensor<half> &dst, ElementWiseUnary op); +template SimpleTensor<int32_t> elementwise_unary(const SimpleTensor<int32_t> &src, SimpleTensor<int32_t> &dst, ElementWiseUnary op); + } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ElementwiseUnary.h b/tests/validation/reference/ElementwiseUnary.h index be4a229a5b..ae7a49bce4 100644 --- a/tests/validation/reference/ElementwiseUnary.h +++ b/tests/validation/reference/ElementwiseUnary.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 Arm Limited. + * Copyright (c) 2018-2019, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -35,7 +35,7 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> elementwise_unary(const SimpleTensor<T> &src, ElementWiseUnary op); +SimpleTensor<T> elementwise_unary(const SimpleTensor<T> &src, SimpleTensor<T> &dst, ElementWiseUnary op); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/FullyConnectedLayer.cpp b/tests/validation/reference/FullyConnectedLayer.cpp index 21333958f8..af30e9ee54 100644 --- a/tests/validation/reference/FullyConnectedLayer.cpp +++ b/tests/validation/reference/FullyConnectedLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -123,7 +123,7 @@ SimpleTensor<T> fully_connected_layer(const SimpleTensor<T> &src, const SimpleTe // Create reference SimpleTensor<T> dst{ TensorShape{ dst_shape }, src.data_type(), 1, out_quant_info }; - // Sanity checks + // Health checks const int num_batch_dimensions = std::max(0, static_cast<int>(dst_shape.num_dimensions()) - 1); const int num_input_dimensions = src.shape().num_dimensions() - num_batch_dimensions; const unsigned int linear_input_size = src.shape().total_size_lower(num_input_dimensions); diff --git a/tests/validation/reference/GEMM.cpp b/tests/validation/reference/GEMM.cpp index 6b3aa390f0..d513343796 100644 --- a/tests/validation/reference/GEMM.cpp +++ b/tests/validation/reference/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,6 +25,7 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/Types.h" +#include "tests/validation/reference/ArithmeticOperations.h" namespace arm_compute { @@ -35,10 +36,11 @@ namespace validation namespace reference { template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> -SimpleTensor<T> gemm(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta) +SimpleTensor<T> +gemm(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta) { // Create reference - SimpleTensor<T> dst{ c.shape(), c.data_type(), 1 }; + SimpleTensor<T> dst{c.shape(), c.data_type(), 1}; // Compute reference const int M = a.shape().y(); @@ -50,30 +52,47 @@ SimpleTensor<T> gemm(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const S const int a_stride_z = K * M; const int a_stride_w = K * M * D; - const int b_stride_z = b.shape().num_dimensions() > 2 ? N * K : 0; // Do not slide the matrix B along the 3th dimension in case matrix B has less than 3 dimensions - const int b_stride_w = b.shape().num_dimensions() > 3 ? K * N * D : 0; // Do not slide the matrix B along the 4th dimension in case matrix B has less than 4 dimensions + const int b_stride_z = + b.shape().num_dimensions() > 2 + ? N * K + : 0; // Do not slide the matrix B along the 3th dimension in case matrix B has less than 3 dimensions + int b_stride_w = + b.shape().num_dimensions() > 3 + ? K * N * D + : 0; // Do not slide the matrix B along the 4th dimension in case matrix B has less than 4 dimensions + + // Note: There are 3 gemm types: batched-gemm, multi-gemm, and batched of multi-gemms. The third dimension of tensor b is overloaded when tensor b has exactly 3 dimensions: + // it can be either number of batches or multis. Batched-GEMM computation is detected only when the third dimension of "a" and "c" tensors is 1 and the number of dimensions is 4 + const bool is_batched_gemm = b.shape().num_dimensions() == 3 && a.shape().num_dimensions() == 4 && + c.shape().num_dimensions() == 4 && a.shape()[2] == 1 && c.shape()[2] == 1; + + // Batched-GEMM + if (is_batched_gemm) + { + b_stride_w = b_stride_z; + } const int c_stride_z = N * M; const int c_stride_w = N * M * D; -#if defined(_OPENMP) && !( defined(__arm__) && defined(__ANDROID__)) +#if defined(_OPENMP) && !(defined(__arm__) && defined(__ANDROID__)) #pragma omp parallel for collapse(2) #endif /* _OPENMP */ - for(int w = 0; w < W; ++w) + for (int w = 0; w < W; ++w) { - for(int depth = 0; depth < D; ++depth) + for (int depth = 0; depth < D; ++depth) { const int base_addr_a = depth * a_stride_z + w * a_stride_w; const int base_addr_b = depth * b_stride_z + w * b_stride_w; const int base_addr_c = depth * c_stride_z + w * c_stride_w; - for(int row = 0; row < M; ++row) + for (int row = 0; row < M; ++row) { - for(int col = 0; col < N; ++col) + for (int col = 0; col < N; ++col) { T acc(0); - for(int k = 0; k < K; ++k) + for (int k = 0; k < K; ++k) { acc += a[base_addr_a + k + row * K] * b[base_addr_b + col + k * N]; } @@ -89,11 +108,12 @@ SimpleTensor<T> gemm(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const S } template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> -SimpleTensor<T> gemm_mixed_precision(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta) +SimpleTensor<T> gemm_mixed_precision( + const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta) { // GEMM mixed-precision combines F32 accumulators with F16 multiplications // Create reference - SimpleTensor<T> dst{ c.shape(), c.data_type(), 1 }; + SimpleTensor<T> dst{c.shape(), c.data_type(), 1}; // Compute reference const int M = a.shape().y(); @@ -105,36 +125,54 @@ SimpleTensor<T> gemm_mixed_precision(const SimpleTensor<T> &a, const SimpleTenso const int a_stride_z = K * M; const int a_stride_w = K * M * D; - const int b_stride_z = b.shape().num_dimensions() > 2 ? N * K : 0; // Do not slide the matrix B along the 3th dimension in case matrix B has less than 3 dimensions - const int b_stride_w = b.shape().num_dimensions() > 3 ? K * N * D : 0; // Do not slide the matrix B along the 4th dimension in case matrix B has less than 4 dimensions + const int b_stride_z = + b.shape().num_dimensions() > 2 + ? N * K + : 0; // Do not slide the matrix B along the 3th dimension in case matrix B has less than 3 dimensions + int b_stride_w = + b.shape().num_dimensions() > 3 + ? K * N * D + : 0; // Do not slide the matrix B along the 4th dimension in case matrix B has less than 4 dimensions + + // Note: There are 3 gemm types: batched-gemm, multi-gemm, and batched of multi-gemms. The third dimension of tensor b is overloaded when tensor b has exactly 3 dimensions: + // it can be either number of batches or multis. Batched-GEMM computation is detected only when the third dimension of "a" and "c" tensors is 1 and the number of dimensions is 4 + const bool is_batched_gemm = b.shape().num_dimensions() == 3 && a.shape().num_dimensions() == 4 && + c.shape().num_dimensions() == 4 && a.shape()[2] == 1 && c.shape()[2] == 1; + + // Batched-GEMM + if (is_batched_gemm) + { + b_stride_w = b_stride_z; + } const int c_stride_z = N * M; const int c_stride_w = N * M * D; -#if defined(_OPENMP) && !( defined(__arm__) && defined(__ANDROID__)) +#if defined(_OPENMP) && !(defined(__arm__) && defined(__ANDROID__)) #pragma omp parallel for collapse(2) #endif /* _OPENMP */ - for(int w = 0; w < W; ++w) + for (int w = 0; w < W; ++w) { - for(int depth = 0; depth < D; ++depth) + for (int depth = 0; depth < D; ++depth) { const int base_addr_a = depth * a_stride_z + w * a_stride_w; const int base_addr_b = depth * b_stride_z + w * b_stride_w; const int base_addr_c = depth * c_stride_z + w * c_stride_w; - for(int row = 0; row < M; ++row) + for (int row = 0; row < M; ++row) { - for(int col = 0; col < N; ++col) + for (int col = 0; col < N; ++col) { float acc(0); - for(int k = 0; k < K; ++k) + for (int k = 0; k < K; ++k) { acc += static_cast<float>(a[base_addr_a + k + row * K] * b[base_addr_b + col + k * N]); } // Finalize the result: alpha * A * B + beta * C - dst[base_addr_c + col + row * N] = static_cast<T>(alpha * acc + beta * c[base_addr_c + col + row * N]); + dst[base_addr_c + col + row * N] = + static_cast<T>(alpha * acc + beta * c[base_addr_c + col + row * N]); } } } @@ -143,8 +181,21 @@ SimpleTensor<T> gemm_mixed_precision(const SimpleTensor<T> &a, const SimpleTenso return dst; } +template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> +void gemm_accumulate(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta, SimpleTensor<T> &dst) +{ + // Compute reference + SimpleTensor<T> dst_gemm = gemm(a, b, c, alpha, beta); + reference::arithmetic_operation<T>(reference::ArithmeticOperation::ADD, dst, dst_gemm, dst, ConvertPolicy::SATURATE); +} + +template SimpleTensor<bfloat16> gemm(const SimpleTensor<bfloat16> &a, const SimpleTensor<bfloat16> &b, const SimpleTensor<bfloat16> &c, float alpha, float beta); template SimpleTensor<float> gemm(const SimpleTensor<float> &a, const SimpleTensor<float> &b, const SimpleTensor<float> &c, float alpha, float beta); template SimpleTensor<half> gemm(const SimpleTensor<half> &a, const SimpleTensor<half> &b, const SimpleTensor<half> &c, float alpha, float beta); + +template void gemm_accumulate(const SimpleTensor<float> &a, const SimpleTensor<float> &b, const SimpleTensor<float> &c, float alpha, float beta, SimpleTensor<float> &dst); +template void gemm_accumulate(const SimpleTensor<half> &a, const SimpleTensor<half> &b, const SimpleTensor<half> &c, float alpha, float beta, SimpleTensor<half> &dst); + template SimpleTensor<half> gemm_mixed_precision(const SimpleTensor<half> &a, const SimpleTensor<half> &b, const SimpleTensor<half> &c, float alpha, float beta); } // namespace reference } // namespace validation diff --git a/tests/validation/reference/GEMM.h b/tests/validation/reference/GEMM.h index 5feaeda584..1b97570122 100644 --- a/tests/validation/reference/GEMM.h +++ b/tests/validation/reference/GEMM.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 Arm Limited. + * Copyright (c) 2017-2019, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_GEMM_H -#define ARM_COMPUTE_TEST_GEMM_H +#ifndef ACL_TESTS_VALIDATION_REFERENCE_GEMM_H +#define ACL_TESTS_VALIDATION_REFERENCE_GEMM_H #include "tests/SimpleTensor.h" #include "tests/validation/Helpers.h" @@ -41,8 +41,11 @@ SimpleTensor<T> gemm(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const S template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0> SimpleTensor<T> gemm_mixed_precision(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta); +template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0> +void gemm_accumulate(const SimpleTensor<T> &a, const SimpleTensor<T> &b, const SimpleTensor<T> &c, float alpha, float beta, SimpleTensor<T> &dst); + } // namespace reference } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_GEMM_H */ +#endif // ACL_TESTS_VALIDATION_REFERENCE_GEMM_H diff --git a/tests/validation/reference/GEMMLowp.cpp b/tests/validation/reference/GEMMLowp.cpp index 1615b51e73..30c577d850 100644 --- a/tests/validation/reference/GEMMLowp.cpp +++ b/tests/validation/reference/GEMMLowp.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #include "GEMMLowp.h" #include "arm_compute/core/Types.h" +#include "tests/validation/reference/ArithmeticOperations.h" #include "tests/validation/reference/UtilsQuantizedAsymm.h" #include "support/ToolchainSupport.h" @@ -230,6 +231,13 @@ SimpleTensor<T_out> gemmlowp_matrix_multiply_core(const SimpleTensor<T_in> &a, c return c; } +template <typename T_out, typename T_in, typename T_in_1> +void gemmlowp_matrix_multiply_core_accumulate(const SimpleTensor<T_in> &a, const SimpleTensor<T_in_1> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset, SimpleTensor<T_out> &dst) +{ + SimpleTensor<T_out> dst_gemm = gemmlowp_matrix_multiply_core<T_out, T_in, T_in_1>(a, b, shape_c, a_offset, b_offset); + reference::arithmetic_operation<T_out>(reference::ArithmeticOperation::ADD, dst, dst_gemm, dst, ConvertPolicy::SATURATE); +} + // used to validate assembly kernels which don't know anything about offsets template <typename T1, typename T2, typename T3> SimpleTensor<T1> gemmlowp(const SimpleTensor<T2> &a, const SimpleTensor<T3> &b, TensorShape shape_c) @@ -336,6 +344,8 @@ template SimpleTensor<int8_t> gemmlowp_quantize_down_scale(const SimpleTensor<in std::vector<int32_t> result_shift, int32_t min, int32_t max); template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset); template SimpleTensor<int32_t> gemmlowp_matrix_multiply_core(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset); +template void gemmlowp_matrix_multiply_core_accumulate(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset, SimpleTensor<int32_t> &dst); +template void gemmlowp_matrix_multiply_core_accumulate(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset, SimpleTensor<int32_t> &dst); template SimpleTensor<int32_t> gemmlowp<int32_t, int8_t, int8_t>(const SimpleTensor<int8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c); template SimpleTensor<int32_t> gemmlowp<int32_t, uint8_t, uint8_t>(const SimpleTensor<uint8_t> &a, const SimpleTensor<uint8_t> &b, TensorShape shape_c); template SimpleTensor<int32_t> gemmlowp<int32_t, uint8_t, int8_t>(const SimpleTensor<uint8_t> &a, const SimpleTensor<int8_t> &b, TensorShape shape_c); diff --git a/tests/validation/reference/GEMMLowp.h b/tests/validation/reference/GEMMLowp.h index 99015d71fb..6e471fdad1 100644 --- a/tests/validation/reference/GEMMLowp.h +++ b/tests/validation/reference/GEMMLowp.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_GEMMLOWP_H -#define ARM_COMPUTE_TEST_GEMMLOWP_H +#ifndef ACL_TESTS_VALIDATION_REFERENCE_GEMMLOWP_H +#define ACL_TESTS_VALIDATION_REFERENCE_GEMMLOWP_H #include "tests/SimpleTensor.h" #include "tests/validation/Helpers.h" @@ -38,6 +38,9 @@ namespace reference template <typename T1, typename T2, typename T3> SimpleTensor<T1> gemmlowp_matrix_multiply_core(const SimpleTensor<T2> &a, const SimpleTensor<T3> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset); +template <typename T1, typename T2, typename T3> +void gemmlowp_matrix_multiply_core_accumulate(const SimpleTensor<T2> &a, const SimpleTensor<T3> &b, TensorShape shape_c, int32_t a_offset, int32_t b_offset, SimpleTensor<T1> &dst_); + template <typename T1, typename T2, typename T3 = T2> SimpleTensor<T1> gemmlowp(const SimpleTensor<T2> &a, const SimpleTensor<T3> &b, TensorShape shape_c); @@ -71,4 +74,4 @@ SimpleTensor<TOut> gemmlowp_quantize_down_scale_by_float(const SimpleTensor<TIn> } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_GEMMLOWP_H */ +#endif // ACL_TESTS_VALIDATION_REFERENCE_GEMMLOWP_H diff --git a/tests/validation/reference/Gather.cpp b/tests/validation/reference/Gather.cpp index 93ac09cf95..c90c04f8cc 100644 --- a/tests/validation/reference/Gather.cpp +++ b/tests/validation/reference/Gather.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 Arm Limited. + * Copyright (c) 2018-2019, 2022-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -39,27 +39,56 @@ namespace reference template <typename T> SimpleTensor<T> gather(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> &indices, uint32_t actual_axis) { - const auto *indices_ptr = static_cast<const uint32_t *>(indices.data()); const TensorShape dst_shape = arm_compute::misc::shape_calculator::compute_gather_shape(src.shape(), indices.shape(), actual_axis); SimpleTensor<T> dst(dst_shape, src.data_type()); + const auto src_ptr = static_cast<const T *>(src.data()); + const auto indices_ptr = static_cast<const uint32_t *>(indices.data()); + const auto dst_ptr = static_cast<T *>(dst.data()); + + const uint32_t index_limit = src.shape()[actual_axis]; + Window win; win.use_tensor_dimensions(dst_shape); - execute_window_loop(win, [&](const Coordinates & id) - { - Coordinates offset; - for(unsigned int dim = 0; dim < id.num_dimensions(); ++dim) + + execute_window_loop(win, [&](const Coordinates &dst_coords) { + const auto dst_addr = coords2index(dst.shape(), dst_coords); + + // Calculate the coordinates of the index value. + Coordinates idx_coords; + + for(size_t i = 0; i < indices.shape().num_dimensions(); ++i) { - if(dim == actual_axis) + idx_coords.set(i, dst_coords[i + actual_axis]); + } + + const auto index = indices_ptr[coords2index(indices.shape(), idx_coords)]; + + if(index < index_limit) + { + // Calculate the coordinates of the source data. + Coordinates src_coords; + + for(size_t i = 0; i < actual_axis; ++i) { - offset.set(dim, indices_ptr[id[dim]]); + src_coords.set(i, dst_coords[i]); } - else + + src_coords.set(actual_axis, index); + + for(size_t i = actual_axis + 1; i < src.shape().num_dimensions(); ++i) { - offset.set(dim, id[dim]); + src_coords.set(i, dst_coords[i + indices.shape().num_dimensions() - 1]); } + + // Copy the data. + const auto src_addr = coords2index(src.shape(), src_coords); + dst_ptr[dst_addr] = src_ptr[src_addr]; + } + else + { + dst_ptr[dst_addr] = 0; } - *reinterpret_cast<T *>(dst(id)) = *reinterpret_cast<const T *>(src(offset)); }); return dst; @@ -72,4 +101,4 @@ template SimpleTensor<uint8_t> gather(const SimpleTensor<uint8_t> &src, const Si } // namespace reference } // namespace validation } // namespace test -} // namespace arm_compute
\ No newline at end of file +} // namespace arm_compute diff --git a/tests/validation/reference/IndirectConv2dAddressPrecalculation.cpp b/tests/validation/reference/IndirectConv2dAddressPrecalculation.cpp new file mode 100644 index 0000000000..7500560c91 --- /dev/null +++ b/tests/validation/reference/IndirectConv2dAddressPrecalculation.cpp @@ -0,0 +1,98 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "IndirectConv2dAddressPrecalculation.h" + +#include "arm_compute/core/Types.h" + +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +SimpleTensor<int32_t> indirect_conv2d_addr_precalculation(const TensorShape &shape_conv_src, const TensorShape &shape_conv_wei, const TensorShape &shape_conv_dst, const TensorShape &shape_dst, + const PadStrideInfo &conv_info) +{ + SimpleTensor<int32_t> out{ shape_dst, DataType::S32 }; + + constexpr unsigned int width_idx = 1; + constexpr unsigned int heigh_idx = 2; + + const int src_conv_width = static_cast<int32_t>(shape_conv_src[width_idx]); // NHWC + const int src_conv_height = static_cast<int32_t>(shape_conv_src[heigh_idx]); // NHWC + const int dst_conv_width = static_cast<int32_t>(shape_conv_dst[width_idx]); // NHWC + const int wei_conv_width = static_cast<int32_t>(shape_conv_wei[width_idx]); // NHWC + const int wei_conv_height = static_cast<int32_t>(shape_conv_wei[heigh_idx]); // NHWC + const int dst_width = static_cast<int32_t>(shape_dst[0]); + const int dst_height = static_cast<int32_t>(shape_dst[1]); + const int dst_batch = static_cast<int32_t>(shape_dst[2]); + const int ks = wei_conv_width * wei_conv_height; + const int stride_x = static_cast<int32_t>(conv_info.stride().first); + const int stride_y = static_cast<int32_t>(conv_info.stride().second); + const int pad_left = static_cast<int32_t>(conv_info.pad_left()); + const int pad_top = static_cast<int32_t>(conv_info.pad_top()); + + const int m0 = dst_width / ks; + + for(int z = 0; z < dst_batch; ++z) + { + for(int y = 0; y < dst_height; ++y) + { + const int mout = y * m0; + for(int ki = 0; ki < ks; ++ki) + { + const int xk = ki % wei_conv_width; + const int yk = ki / wei_conv_width; + for(int mi = 0; mi < m0; ++mi) + { + int xi = ((mout + mi) % dst_conv_width) * stride_x; + int yi = ((mout + mi) / dst_conv_width) * stride_y; + xi -= pad_left; + yi -= pad_top; + const int x_s = xi + xk; + const int y_s = yi + yk; + int my = x_s + y_s * src_conv_width; + my = my + z * src_conv_width * src_conv_height; + my = x_s >= 0 ? my : -1; + my = x_s < src_conv_width ? my : -1; + my = y_s >= 0 ? my : -1; + my = y_s < src_conv_height ? my : -1; + + const unsigned int addr_out = mi + ki * m0 + y * (dst_width) + z * (dst_width * dst_height); + out[addr_out] = my; + } + } + } + } + + return out; +} +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute
\ No newline at end of file diff --git a/tests/validation/reference/IndirectConv2dAddressPrecalculation.h b/tests/validation/reference/IndirectConv2dAddressPrecalculation.h new file mode 100644 index 0000000000..f4a90dfd9f --- /dev/null +++ b/tests/validation/reference/IndirectConv2dAddressPrecalculation.h @@ -0,0 +1,44 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_H +#define ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_H + +#include "tests/SimpleTensor.h" +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +SimpleTensor<int32_t> indirect_conv2d_addr_precalculation(const TensorShape &shape_conv_src, const TensorShape &shape_conv_wei, const TensorShape &shape_conv_out, const TensorShape &shape_out, + const PadStrideInfo &conv_info); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_INDIRECT_CONV2D_ADDRESS_PRECALCULATION_H */
\ No newline at end of file diff --git a/tests/validation/reference/MeanStdDevNormalizationLayer.cpp b/tests/validation/reference/MeanStdDevNormalizationLayer.cpp index 0a23fa19bb..a7c8a784d9 100644 --- a/tests/validation/reference/MeanStdDevNormalizationLayer.cpp +++ b/tests/validation/reference/MeanStdDevNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019 Arm Limited. + * Copyright (c) 2019, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -63,6 +63,15 @@ SimpleTensor<T> mean_std_normalization_layer(const SimpleTensor<T> &src, float e return dst; } +template <> +SimpleTensor<uint8_t> mean_std_normalization_layer(const SimpleTensor<uint8_t> &src, float epsilon) +{ + SimpleTensor<float> src_tmp = convert_from_asymmetric(src); + SimpleTensor<float> dst_tmp = mean_std_normalization_layer<float>(src_tmp, epsilon); + SimpleTensor<uint8_t> dst = convert_to_asymmetric<uint8_t>(dst_tmp, src.quantization_info()); + return dst; +} + template SimpleTensor<float> mean_std_normalization_layer(const SimpleTensor<float> &src, float epsilon); template SimpleTensor<half> mean_std_normalization_layer(const SimpleTensor<half> &src, float epsilon); } // namespace reference diff --git a/tests/validation/reference/Permute.cpp b/tests/validation/reference/Permute.cpp index 6f122b1bf5..7aa3011d8f 100644 --- a/tests/validation/reference/Permute.cpp +++ b/tests/validation/reference/Permute.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 Arm Limited. + * Copyright (c) 2017-2019,2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #include "Permute.h" #include "arm_compute/core/Types.h" + #include "tests/validation/Helpers.h" namespace arm_compute @@ -42,11 +43,11 @@ SimpleTensor<T> permute(const SimpleTensor<T> &src, PermutationVector perm) permute(dst_shape, perm); // Create reference - SimpleTensor<T> dst{ dst_shape, src.data_type(), src.num_channels(), src.quantization_info() }; + SimpleTensor<T> dst{dst_shape, src.data_type(), src.num_channels(), src.quantization_info()}; // Compute reference const uint32_t num_elements = src.num_elements(); - for(uint32_t i = 0; i < num_elements; ++i) + for (uint32_t i = 0; i < num_elements; ++i) { const Coordinates src_coords = index2coord(src.shape(), i); Coordinates dst_coords = src_coords; @@ -58,13 +59,14 @@ SimpleTensor<T> permute(const SimpleTensor<T> &src, PermutationVector perm) return dst; } -template SimpleTensor<int8_t> permute(const SimpleTensor<int8_t> &src, PermutationVector perm); -template SimpleTensor<uint8_t> permute(const SimpleTensor<uint8_t> &src, PermutationVector perm); -template SimpleTensor<int16_t> permute(const SimpleTensor<int16_t> &src, PermutationVector perm); +template SimpleTensor<int8_t> permute(const SimpleTensor<int8_t> &src, PermutationVector perm); +template SimpleTensor<uint8_t> permute(const SimpleTensor<uint8_t> &src, PermutationVector perm); +template SimpleTensor<int16_t> permute(const SimpleTensor<int16_t> &src, PermutationVector perm); template SimpleTensor<uint16_t> permute(const SimpleTensor<uint16_t> &src, PermutationVector perm); template SimpleTensor<uint32_t> permute(const SimpleTensor<uint32_t> &src, PermutationVector perm); -template SimpleTensor<float> permute(const SimpleTensor<float> &src, PermutationVector perm); -template SimpleTensor<half> permute(const SimpleTensor<half> &src, PermutationVector perm); +template SimpleTensor<float> permute(const SimpleTensor<float> &src, PermutationVector perm); +template SimpleTensor<half> permute(const SimpleTensor<half> &src, PermutationVector perm); +template SimpleTensor<bfloat16> permute(const SimpleTensor<bfloat16> &src, PermutationVector perm); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Pooling3dLayer.cpp b/tests/validation/reference/Pooling3dLayer.cpp new file mode 100644 index 0000000000..2e8f3a0b92 --- /dev/null +++ b/tests/validation/reference/Pooling3dLayer.cpp @@ -0,0 +1,220 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ + +#include "Pooling3dLayer.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +using namespace arm_compute::misc::shape_calculator; + +template <typename T> +SimpleTensor<T> pooling_3d_layer_internal(const SimpleTensor<T> &src, const Pooling3dLayerInfo &pool3d_info, SimpleTensor<uint32_t> *indices) +{ + TensorShape pooled_shape = compute_pool3d_shape(src.shape(), pool3d_info); + SimpleTensor<T> dst{ pooled_shape, src.data_type(), 1 }; + + if(indices != nullptr) + { + *indices = SimpleTensor<uint32_t> { pooled_shape, DataType::U32, 1 }; + } + + const int idx_channel = 0; + const int idx_width = 1; + const int idx_height = 2; + const int idx_depth = 3; + const int idx_batch = 4; + + const int pool_size_width = pool3d_info.is_global_pooling ? src.shape()[idx_width] : pool3d_info.pool_size.width; + const int pool_size_height = pool3d_info.is_global_pooling ? src.shape()[idx_height] : pool3d_info.pool_size.height; + const int pool_size_depth = pool3d_info.is_global_pooling ? src.shape()[idx_depth] : pool3d_info.pool_size.depth; + + const int pool_stride_width = static_cast<int>(pool3d_info.stride.width); + const int pool_stride_height = static_cast<int>(pool3d_info.stride.height); + const int pool_stride_depth = static_cast<int>(pool3d_info.stride.depth); + + const int pad_left = static_cast<int>(pool3d_info.padding.left); + const int pad_top = static_cast<int>(pool3d_info.padding.top); + const int pad_front = static_cast<int>(pool3d_info.padding.front); + + const int pad_right = static_cast<int>(pool3d_info.padding.right); + const int pad_bottom = static_cast<int>(pool3d_info.padding.bottom); + const int pad_back = static_cast<int>(pool3d_info.padding.back); + + const int num_channels = static_cast<int>(src.shape()[idx_channel]); + const int num_batches = static_cast<int>(src.shape()[idx_batch]); + + ARM_COMPUTE_ERROR_ON(num_channels != static_cast<int>(dst.shape()[idx_channel])); + ARM_COMPUTE_ERROR_ON(num_batches != static_cast<int>(dst.shape()[idx_batch])); + + const int w_src = static_cast<int>(src.shape()[idx_width]); + const int h_src = static_cast<int>(src.shape()[idx_height]); + const int d_src = static_cast<int>(src.shape()[idx_depth]); + const int w_dst = static_cast<int>(dst.shape()[idx_width]); + const int h_dst = static_cast<int>(dst.shape()[idx_height]); + const int d_dst = static_cast<int>(dst.shape()[idx_depth]); + + const bool exclude_padding = pool3d_info.exclude_padding; + + const int height_stride_src = num_channels * w_src; + const int depth_stride_src = height_stride_src * h_src; + const int batch_stride_src = depth_stride_src * d_src; + const int height_stride_dst = num_channels * w_dst; + const int depth_stride_dst = height_stride_dst * h_dst; + const int batch_stride_dst = depth_stride_dst * d_dst; + + for(int b = 0; b < num_batches; ++b) + { + const int batch_offset_dst = b * batch_stride_dst; + const int batch_offset_src = b * batch_stride_src; + for(int c = 0; c < num_channels; ++c) + { + for(int d = 0; d < d_dst; ++d) + { + const int depth_offset_dst = d * depth_stride_dst; + for(int h = 0; h < h_dst; ++h) + { + const int height_offset_dst = h * height_stride_dst; + for(int w = 0; w < w_dst; ++w) + { + int wstart = w * pool_stride_width - pad_left; + int hstart = h * pool_stride_height - pad_top; + int dstart = d * pool_stride_depth - pad_front; + int wend = std::min(wstart + pool_size_width, w_src + pad_right); + int hend = std::min(hstart + pool_size_height, h_src + pad_bottom); + int dend = std::min(dstart + pool_size_depth, d_src + pad_back); + + // this may not be equal to pool_w * pool_h * pool_d because of + // DimensionRoundingType choice (CEIL) + int pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + + // limit [start, end) to [0, w_src) + wstart = std::max(wstart, 0); + hstart = std::max(hstart, 0); + dstart = std::max(dstart, 0); + wend = std::min(wend, w_src); + hend = std::min(hend, h_src); + dend = std::min(dend, d_src); + + auto max_val = -std::numeric_limits<T>::infinity(); + int max_index{ 0 }; + T avg_val = static_cast<T>(0.f); + T l2_val = static_cast<T>(0.f); + + if(exclude_padding) + { + pool_size = (dend - dstart) * (hend - hstart) * (wend - wstart); + } + + for(int z = dstart; z < dend; ++z) + { + const int depth_offset_src = z * depth_stride_src; + for(int y = hstart; y < hend; ++y) + { + const int height_offset_src = y * height_stride_src; + for(int x = wstart; x < wend; ++x) + { + const auto val = static_cast<T>( + src[batch_offset_src + depth_offset_src + height_offset_src + x * num_channels + c]); + if(val > max_val) + { + max_val = val; + max_index = coord2index(src.shape(), Coordinates(c, x, y, z, 0)); + } + + avg_val += val; + l2_val += val * val; + } + } + } + + avg_val /= pool_size; + l2_val = static_cast<T>(std::sqrt(l2_val / pool_size)); + + int dst_index = batch_offset_dst + depth_offset_dst + height_offset_dst + w * num_channels + c; + switch(pool3d_info.pool_type) + { + case PoolingType::MAX: + dst[dst_index] = static_cast<T>(max_val); + break; + case PoolingType::AVG: + dst[dst_index] = static_cast<T>(avg_val); + break; + case PoolingType::L2: + dst[dst_index] = static_cast<T>(l2_val); + break; + default: + ARM_COMPUTE_ERROR("Pooling Type should be either MAX, AVG or L2"); + } + + if(indices != nullptr) + { + (*indices)[dst_index] = max_index; + } + } + } + } + } + } + + return dst; +} + +template SimpleTensor<float> pooling_3d_layer(const SimpleTensor<float> &src, const Pooling3dLayerInfo &pool3d_info, const QuantizationInfo &output_qinfo, SimpleTensor<uint32_t> *indices); +template SimpleTensor<half> pooling_3d_layer(const SimpleTensor<half> &src, const Pooling3dLayerInfo &pool3d_info, const QuantizationInfo &output_qinfo, SimpleTensor<uint32_t> *indices); + +template <typename T> +SimpleTensor<T> pooling_3d_layer(const SimpleTensor<T> &src, const Pooling3dLayerInfo &pool3d_info, const QuantizationInfo &output_qinfo, SimpleTensor<uint32_t> *indices) +{ + ARM_COMPUTE_UNUSED(output_qinfo); + return pooling_3d_layer_internal<T>(src, pool3d_info, indices); +} + +template <> +SimpleTensor<int8_t> pooling_3d_layer<int8_t>(const SimpleTensor<int8_t> &src, const Pooling3dLayerInfo &pool3d_info, const QuantizationInfo &output_qinfo, SimpleTensor<uint32_t> *indices) +{ + SimpleTensor<float> src_tmp = convert_from_asymmetric(src); + SimpleTensor<float> dst_tmp = pooling_3d_layer_internal<float>(src_tmp, pool3d_info, indices); + return convert_to_asymmetric<int8_t>(dst_tmp, output_qinfo); +} + +template <> +SimpleTensor<uint8_t> pooling_3d_layer<uint8_t>(const SimpleTensor<uint8_t> &src, const Pooling3dLayerInfo &pool3d_info, const QuantizationInfo &output_qinfo, SimpleTensor<uint32_t> *indices) +{ + SimpleTensor<float> src_tmp = convert_from_asymmetric(src); + SimpleTensor<float> dst_tmp = pooling_3d_layer_internal<float>(src_tmp, pool3d_info, indices); + return convert_to_asymmetric<uint8_t>(dst_tmp, output_qinfo); +} + +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/Pooling3dLayer.h b/tests/validation/reference/Pooling3dLayer.h new file mode 100644 index 0000000000..481a0d3024 --- /dev/null +++ b/tests/validation/reference/Pooling3dLayer.h @@ -0,0 +1,50 @@ +/* + * Copyright (c) 2022 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ARM_COMPUTE_TEST_POOL3D_LAYER_H +#define ARM_COMPUTE_TEST_POOL3D_LAYER_H + +#include "Utils.h" +#include "arm_compute/core/Types.h" +#include "tests/SimpleTensor.h" +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<T> pooling_3d_layer_internal(const SimpleTensor<T> &src, const Pooling3dLayerInfo &pool3d_info, SimpleTensor<uint32_t> *indices = nullptr); + +template <typename T> +SimpleTensor<T> pooling_3d_layer(const SimpleTensor<T> &src, const Pooling3dLayerInfo &pool3d_info, const QuantizationInfo &output_qinfo = QuantizationInfo(), + SimpleTensor<uint32_t> *indices = nullptr); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_POOL3D_LAYER_H */ diff --git a/tests/validation/reference/PoolingLayer.cpp b/tests/validation/reference/PoolingLayer.cpp index 5f4edfe49c..bf7bd0c1df 100644 --- a/tests/validation/reference/PoolingLayer.cpp +++ b/tests/validation/reference/PoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -40,7 +40,6 @@ using namespace arm_compute::misc::shape_calculator; template <typename T, typename ACC_T, typename std::enable_if<is_floating_point<T>::value, int>::type> SimpleTensor<T> pooling_layer_internal(const SimpleTensor<T> &src, const PoolingLayerInfo &info, SimpleTensor<uint32_t> *indices, DataLayout data_layout) { - ARM_COMPUTE_ERROR_ON(info.is_global_pooling && (src.shape().x() != src.shape().y())); // Create reference SimpleTensor<T> dst{ compute_pool_shape(TensorInfo(src.shape(), 1, src.data_type()), info), src.data_type(), 1 }; auto pooled_shape = compute_pool_shape(TensorInfo(src.shape(), 1, src.data_type()), info); @@ -84,20 +83,28 @@ SimpleTensor<T> pooling_layer_internal(const SimpleTensor<T> &src, const Pooling { int wstart = w * pool_stride_x - pad_left; int hstart = h * pool_stride_y - pad_top; + + // Used to calculate kernel indices + int kh_start = std::max(0, -hstart); + int kw_start = std::max(0, -wstart); + int max_ker_index{ 0 }; + int wend = std::min(wstart + pool_size_x, w_src); int hend = std::min(hstart + pool_size_y, h_src); wstart = std::max(wstart, 0); hstart = std::max(hstart, 0); - auto max_val = std::numeric_limits<ACC_T>::lowest(); + auto max_val = info.use_inf_as_limit ? -std::numeric_limits<ACC_T>::infinity() : std::numeric_limits<ACC_T>::lowest(); int max_index{ 0 }; - for(int y = hstart; y < hend; ++y) + + for(int y = hstart, kh = kh_start; y < hend; ++y, ++kh) { - for(int x = wstart; x < wend; ++x) + for(int x = wstart, kw = kw_start; x < wend; ++x, ++kw) { const auto val = static_cast<ACC_T>(src[b * z_src * h_src * w_src + r * h_src * w_src + y * w_src + x]); if(val > max_val) { - max_val = val; + max_val = val; + max_ker_index = pool_size_x * (kh) + (kw); if(data_layout == DataLayout::NCHW) { max_index = coord2index(src.shape(), Coordinates(x, y, r, 0)); @@ -113,7 +120,7 @@ SimpleTensor<T> pooling_layer_internal(const SimpleTensor<T> &src, const Pooling dst[b * z_dst * h_dst * w_dst + r * h_dst * w_dst + h * w_dst + w] = static_cast<T>(max_val); if(indices) { - (*indices)[b * z_dst * h_dst * w_dst + r * h_dst * w_dst + h * w_dst + w] = max_index; + (*indices)[b * z_dst * h_dst * w_dst + r * h_dst * w_dst + h * w_dst + w] = (info.use_kernel_indices) ? max_ker_index : max_index; } } } diff --git a/tests/validation/reference/QuantizationLayer.cpp b/tests/validation/reference/QuantizationLayer.cpp index 27665375c3..ad7ba7ac43 100644 --- a/tests/validation/reference/QuantizationLayer.cpp +++ b/tests/validation/reference/QuantizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2024 Arm Limited. * * SPDX-License-Identifier: MIT * diff --git a/tests/validation/reference/ReductionOperation.cpp b/tests/validation/reference/ReductionOperation.cpp index ffb79f86c5..c189bc2d47 100644 --- a/tests/validation/reference/ReductionOperation.cpp +++ b/tests/validation/reference/ReductionOperation.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,7 +22,6 @@ * SOFTWARE. */ #include "ReductionOperation.h" - #include "tests/validation/Helpers.h" #include <algorithm> @@ -39,7 +38,7 @@ namespace reference namespace { template <typename T, typename OT> -OT reduce_operation(const T *ptr, int reduce_elements, ReductionOperation op, int stride) +OT reduce_operation(const T *ptr, int reduce_elements, ReductionOperation op, int stride, RoundingPolicy policy) { using type = typename std::remove_cv<OT>::type; T res; @@ -99,7 +98,14 @@ OT reduce_operation(const T *ptr, int reduce_elements, ReductionOperation op, in } if(op == ReductionOperation::MEAN_SUM && reduce_elements > 0) { - int_res /= reduce_elements; + // Only use rounding in aarch64 to be consistent with kernel +#ifdef __aarch64__ + // Divide in float format, then rounded to nearest and implicitly cast back to int + int_res = round(static_cast<float>(int_res) / static_cast<float>(reduce_elements), policy); +#else // defined(__aarch64__) + ARM_COMPUTE_UNUSED(policy); + int_res /= reduce_elements; // Legacy compatibility +#endif // __aarch64 } res = static_cast<type>(int_res); } @@ -175,12 +181,12 @@ OT reduce_operation_arg_min_max(const T *ptr, int reduce_elements, ReductionOper } // namespace template <typename T, typename OT> -SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op) +SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, + DataType output_type, RoundingPolicy policy) { // Create reference - const bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX); - DataType output_data_type = is_arg_min_max ? DataType::S32 : src.data_type(); - SimpleTensor<OT> dst{ dst_shape, output_data_type, 1, src.quantization_info() }; + const bool is_arg_min_max = (op == ReductionOperation::ARG_IDX_MIN || op == ReductionOperation::ARG_IDX_MAX); + SimpleTensor<OT> dst{ dst_shape, output_type, 1, src.quantization_info() }; const unsigned int src_width = src.shape().x(); const unsigned int src_height = src.shape().y(); const unsigned int src_depth = src.shape().z(); @@ -197,7 +203,7 @@ SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const T const T *src_row_ptr = src.data() + du * reduce_elems; dst[du] = is_arg_min_max ? reduce_operation_arg_min_max<T, OT>(src_row_ptr, reduce_elems, op, 1) : - reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, 1); + reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, 1, policy); } } break; @@ -213,7 +219,7 @@ SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const T const T *src_row_ptr = src.data() + in_offset; dst[out_offset] = is_arg_min_max ? reduce_operation_arg_min_max<T, OT>(src_row_ptr, reduce_elems, op, src_width) : - reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, src_width); + reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, src_width, policy); } } } @@ -232,7 +238,7 @@ SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const T const T *src_row_ptr = src.data() + in_offset; dst[out_offset] = is_arg_min_max ? reduce_operation_arg_min_max<T, OT>(src_row_ptr, reduce_elems, op, src_width * src_height) : - reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, src_width * src_height); + reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, src_width * src_height, policy); } } } @@ -254,7 +260,7 @@ SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const T const T *src_row_ptr = src.data() + in_offset; dst[out_offset] = is_arg_min_max ? reduce_operation_arg_min_max<T, OT>(src_row_ptr, reduce_elems, op, src_width * src_height * src_depth) : - reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, src_width * src_height * src_depth); + reduce_operation<T, OT>(src_row_ptr, reduce_elems, op, src_width * src_height * src_depth, policy); } } } @@ -269,74 +275,89 @@ SimpleTensor<OT> compute_reduction_operation(const SimpleTensor<T> &src, const T } template <typename T, typename OT> -SimpleTensor<OT> reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info_output) +SimpleTensor<OT> reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, + DataType output_type, QuantizationInfo quantization_info_output, RoundingPolicy policy) { ARM_COMPUTE_UNUSED(quantization_info_output); - return compute_reduction_operation<T, OT>(src, dst_shape, axis, op); + return compute_reduction_operation<T, OT>(src, dst_shape, axis, op, output_type, policy); } template <> -SimpleTensor<uint8_t> reduction_operation(const SimpleTensor<uint8_t> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info_output) +SimpleTensor<uint8_t> reduction_operation(const SimpleTensor<uint8_t> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, + DataType output_type, QuantizationInfo quantization_info_output, RoundingPolicy policy) { if(src.data_type() == DataType::QASYMM8) { // If the operation is MEAN_SUM, we can directly use the uint8 implementation without taking into account scale and offset if(op == ReductionOperation::MEAN_SUM && src.quantization_info() == quantization_info_output) { - return compute_reduction_operation<uint8_t, uint8_t>(src, dst_shape, axis, op); + return compute_reduction_operation<uint8_t, uint8_t>(src, dst_shape, axis, op, output_type, policy); } else { SimpleTensor<float> src_f = convert_from_asymmetric(src); - SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op); + SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op, output_type); return convert_to_asymmetric<uint8_t>(dst_f, quantization_info_output); } } else { - return compute_reduction_operation<uint8_t, uint8_t>(src, dst_shape, axis, op); + return compute_reduction_operation<uint8_t, uint8_t>(src, dst_shape, axis, op, output_type, policy); } } template <> -SimpleTensor<int8_t> reduction_operation(const SimpleTensor<int8_t> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, QuantizationInfo quantization_info_output) +SimpleTensor<int8_t> reduction_operation(const SimpleTensor<int8_t> &src, const TensorShape &dst_shape, unsigned int axis, + ReductionOperation op, DataType output_type, QuantizationInfo quantization_info_output, RoundingPolicy policy) { if(src.data_type() == DataType::QASYMM8_SIGNED) { // If the operation is MEAN_SUM, we can directly use the int8 implementation without taking into account scale and offset if(op == ReductionOperation::MEAN_SUM && src.quantization_info() == quantization_info_output) { - return compute_reduction_operation<int8_t, int8_t>(src, dst_shape, axis, op); + return compute_reduction_operation<int8_t, int8_t>(src, dst_shape, axis, op, output_type, policy); } else { SimpleTensor<float> src_f = convert_from_asymmetric(src); - SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op); + SimpleTensor<float> dst_f = reference::reduction_operation<float, float>(src_f, dst_shape, axis, op, output_type); return convert_to_asymmetric<int8_t>(dst_f, quantization_info_output); } } else { - return compute_reduction_operation<int8_t, int8_t>(src, dst_shape, axis, op); + return compute_reduction_operation<int8_t, int8_t>(src, dst_shape, axis, op, output_type, policy); } } template SimpleTensor<float> reduction_operation(const SimpleTensor<float> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, QuantizationInfo quantization_info_output = QuantizationInfo(), + RoundingPolicy policy = RoundingPolicy::TO_ZERO); + template SimpleTensor<half> reduction_operation(const SimpleTensor<half> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); template SimpleTensor<int32_t> reduction_operation(const SimpleTensor<float> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); + template SimpleTensor<int32_t> reduction_operation(const SimpleTensor<int32_t> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); template SimpleTensor<int32_t> reduction_operation(const SimpleTensor<half> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); template SimpleTensor<int32_t> reduction_operation(const SimpleTensor<uint8_t> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); template SimpleTensor<int32_t> reduction_operation(const SimpleTensor<int8_t> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); + DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); +template SimpleTensor<int64_t> reduction_operation(const SimpleTensor<float> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, + DataType output_type = DataType::S32, QuantizationInfo quantization_info_output = QuantizationInfo(), + RoundingPolicy policy = RoundingPolicy::TO_ZERO); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ReductionOperation.h b/tests/validation/reference/ReductionOperation.h index 9c9e721b29..fb2e7a7093 100644 --- a/tests/validation/reference/ReductionOperation.h +++ b/tests/validation/reference/ReductionOperation.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #ifndef ARM_COMPUTE_TEST_REDUCTION_OPERATION_H #define ARM_COMPUTE_TEST_REDUCTION_OPERATION_H +#include "arm_compute/core/Rounding.h" #include "tests/SimpleTensor.h" #include "tests/validation/Helpers.h" @@ -36,8 +37,8 @@ namespace validation namespace reference { template <typename T, typename OT> -SimpleTensor<OT> reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, - QuantizationInfo quantization_info_output = QuantizationInfo()); +SimpleTensor<OT> reduction_operation(const SimpleTensor<T> &src, const TensorShape &dst_shape, unsigned int axis, ReductionOperation op, DataType output_type = DataType::S32, + QuantizationInfo quantization_info_output = QuantizationInfo(), RoundingPolicy policy = RoundingPolicy::TO_ZERO); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Remap.cpp b/tests/validation/reference/Remap.cpp deleted file mode 100644 index 33c5a7de68..0000000000 --- a/tests/validation/reference/Remap.cpp +++ /dev/null @@ -1,112 +0,0 @@ -/* - * Copyright (c) 2017-2021 Arm Limited. - * - * SPDX-License-Identifier: MIT - * - * Permission is hereby granted, free of charge, to any person obtaining a copy - * of this software and associated documentation files (the "Software"), to - * deal in the Software without restriction, including without limitation the - * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or - * sell copies of the Software, and to permit persons to whom the Software is - * furnished to do so, subject to the following conditions: - * - * The above copyright notice and this permission notice shall be included in all - * copies or substantial portions of the Software. - * - * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR - * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, - * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE - * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER - * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, - * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE - * SOFTWARE. - */ -#include "Remap.h" - -#include "Utils.h" -#include "tests/validation/Helpers.h" - -#include <algorithm> -#include <array> - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -namespace reference -{ -template <typename T> -SimpleTensor<T> remap(const SimpleTensor<T> &in, SimpleTensor<float> &map_x, SimpleTensor<float> &map_y, SimpleTensor<T> &valid_mask, InterpolationPolicy policy, BorderMode border_mode, - T constant_border_value) -{ - ARM_COMPUTE_ERROR_ON_MSG(border_mode == BorderMode::REPLICATE, "BorderMode not supported"); - SimpleTensor<T> out(in.shape(), in.data_type()); - ARM_COMPUTE_ERROR_ON(out.num_elements() != map_x.num_elements()); - const int width = in.shape().x(); - const int height = in.shape().y(); - const uint32_t num_elements = out.num_elements(); - for(uint32_t idx = 0; idx < num_elements; idx++) - { - const Coordinates id_out = index2coord(out.shape(), idx); - valid_mask[idx] = 1; - Coordinates src_idx = id_out; // need to setup all coordinates and not just xy - if((0 <= map_y[idx]) && (map_y[idx] < height) && (0 <= map_x[idx]) && (map_x[idx] < width)) - { - switch(policy) - { - case InterpolationPolicy::NEAREST_NEIGHBOR: - { - src_idx.set(0, static_cast<int>(std::floor(map_x[idx]))); - src_idx.set(1, static_cast<int>(std::floor(map_y[idx]))); - out[idx] = in[coord2index(in.shape(), src_idx)]; - break; - } - case InterpolationPolicy::BILINEAR: - { - (valid_bilinear_policy(map_x[idx], map_y[idx], width, height, border_mode)) ? - out[idx] = bilinear_policy(in, src_idx, map_x[idx], map_y[idx], border_mode, constant_border_value) : - valid_mask[idx] = 0; - break; - } - case InterpolationPolicy::AREA: - default: - ARM_COMPUTE_ERROR("Interpolation not supported"); - break; - } - } - else - { - if(border_mode == BorderMode::UNDEFINED) - { - valid_mask[idx] = 0; - } - else - { - switch(policy) - { - case InterpolationPolicy::NEAREST_NEIGHBOR: - out[idx] = constant_border_value; - break; - case InterpolationPolicy::BILINEAR: - out[idx] = bilinear_policy(in, src_idx, map_x[idx], map_y[idx], border_mode, constant_border_value); - break; - case InterpolationPolicy::AREA: - default: - break; - } - } - } - } - - return out; -} - -template SimpleTensor<uint8_t> remap(const SimpleTensor<uint8_t> &src, SimpleTensor<float> &map_x, SimpleTensor<float> &map_y, SimpleTensor<uint8_t> &valid_mask, InterpolationPolicy policy, - BorderMode border_mode, - uint8_t constant_border_value); -} // namespace reference -} // namespace validation -} // namespace test -} // namespace arm_compute diff --git a/tests/validation/reference/Reorder.cpp b/tests/validation/reference/Reorder.cpp new file mode 100644 index 0000000000..8abb372596 --- /dev/null +++ b/tests/validation/reference/Reorder.cpp @@ -0,0 +1,156 @@ +/* + * Copyright (c) 2023 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "Reorder.h" +#include "src/core/NEON/kernels/arm_gemm/utils.hpp" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ + +/* + * Generic transform. + * + * Assuming the untransposed case, this works by first reading <BlockBy> + * consecutive values from the first input row. This same number of values + * are then read from the next <IntBy-1> rows. Now return to the first + * input row and repeat. + * + * Need to cope with the work requested in either dimension not actually + * being a multiple of the block sizes. + */ +template <unsigned int tIntBy, unsigned int BlockBy, bool Transposed, size_t TOutSize, size_t TInSize, typename d_type, arm_gemm::VLType vlt> +struct Transform_ref +{ + template <typename TOut, typename TIn> + static void Transform(TOut &out, const TIn in, const int stride, + const int y0, const int ymax, const int x0, const int xmax) + { + // NOTE: This code is disabled to avoid the call to get_vector_length(), so templated transforms will not be + // correct for SVE. This is not an issue as we have specializations for all SVE cases. + // For SVE cases we multiply the interleave factor by the vector length. + // const unsigned int IntBy = tIntBy * (vlt == VLType::SVE ? get_vector_length<TOut>() / BlockBy : 1); + const unsigned int IntBy = tIntBy; + int out_index = 0; + + const int n_whole_y_blocks = (ymax - y0) / IntBy; + const int y_remainders = (ymax - y0) % IntBy; + const int n_y_blocks = n_whole_y_blocks + (y_remainders ? 1 : 0); + + const int n_whole_x_blocks = (xmax - x0) / BlockBy; + const int x_remainders = (xmax - x0) % BlockBy; + const int n_x_blocks = n_whole_x_blocks + (x_remainders ? 1 : 0); + + // "Y" loop: advance down the rows of the source IntBy rows at a time. + // Set up fill_rows to show the number rows to copy from, and blank_rows + // for the number of blank rows to add. + for(int y_block = 0; y_block < n_y_blocks; y_block++) + { + const int fill_rows = (y_block < n_whole_y_blocks) ? IntBy : y_remainders; + const int blank_rows = IntBy - fill_rows; + + const int y_base = y0 + (y_block * IntBy); + + // So now advance along this block of rows, BlockBy columns at a time. + for(int x_block = 0; x_block < n_x_blocks; x_block++) + { + const int fill_cols = (x_block < n_whole_x_blocks) ? BlockBy : x_remainders; + const int blank_cols = BlockBy - fill_cols; + + const int x_base = x0 + (x_block * BlockBy); + + for(int row = 0; row < fill_rows; row++) + { + for(int col = 0; col < fill_cols; col++) + { + // In-range copy. If it's transposed, we reverse the sense of rows and columns here. + if(Transposed) + { + out[out_index] = in[(x_base + col) * stride + y_base + row]; + out_index++; + } + else + { + out[out_index] = in[(y_base + row) * stride + x_base + col]; + out_index++; + } + } + // "col" tail - row is in range but column is out of range. + for(int col = 0; col < blank_cols; col++) + { + out[out_index] = 0; + out_index++; + } + } + // "row" tail - row is out of range so fill with zeros always. + const d_type zeroval = 0; + const int pads = blank_rows * (fill_cols + blank_cols); + + for(int i = 0; i < pads; i++) + { + out[out_index] = zeroval; + } + + out_index += pads; + } + } + } +}; + +template <typename T> +SimpleTensor<T> reorder_layer(const SimpleTensor<T> &src, const TensorShape &output_shape, WeightFormat output_wf) +{ + SimpleTensor<T> dst{ output_shape, src.data_type() }; + const int cols = src.shape()[0]; + const int rows = src.shape()[1]; + + switch(output_wf) + { + case WeightFormat::OHWIo4: + { + Transform_ref<4, 1, true, sizeof(float), sizeof(float), float, arm_gemm::VLType::None>::Transform<SimpleTensor<T> &, SimpleTensor<T>>(dst, src, rows, 0, rows, 0, cols); + break; + } + case WeightFormat::OHWIo8: + { + Transform_ref<8, 1, true, sizeof(float), sizeof(float), float, arm_gemm::VLType::None>::Transform<SimpleTensor<T> &, SimpleTensor<T>>(dst, src, rows, 0, rows, 0, cols); + break; + } + default: + break; + } + + return dst; +} + +template SimpleTensor<float> reorder_layer(const SimpleTensor<float> &src, const TensorShape &output_shape, WeightFormat output_wf); + +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/Remap.h b/tests/validation/reference/Reorder.h index 0726f75965..94ee5078f8 100644 --- a/tests/validation/reference/Remap.h +++ b/tests/validation/reference/Reorder.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2019 Arm Limited. + * Copyright (c) 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,10 +21,11 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_REMAP_H -#define ARM_COMPUTE_TEST_REMAP_H +#ifndef ACL_TESTS_VALIDATION_REFERENCE_REORDER +#define ACL_TESTS_VALIDATION_REFERENCE_REORDER #include "tests/SimpleTensor.h" +#include "tests/Types.h" namespace arm_compute { @@ -35,10 +36,9 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> remap(const SimpleTensor<T> &in, SimpleTensor<float> &map_x, SimpleTensor<float> &map_y, SimpleTensor<T> &valid_mask, InterpolationPolicy policy, BorderMode border_mode, - T constant_border_value = 0); +SimpleTensor<T> reorder_layer(const SimpleTensor<T> &src, const TensorShape &output_shape, WeightFormat output_wf); } // namespace reference } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_REMAP_H */ +#endif /* ACL_TESTS_VALIDATION_REFERENCE_REORDER */ diff --git a/tests/validation/reference/ReshapeLayer.cpp b/tests/validation/reference/ReshapeLayer.cpp index daea001be6..30a58dd65b 100644 --- a/tests/validation/reference/ReshapeLayer.cpp +++ b/tests/validation/reference/ReshapeLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 Arm Limited. + * Copyright (c) 2017,2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -44,14 +44,15 @@ SimpleTensor<T> reshape_layer(const SimpleTensor<T> &src, const TensorShape &out return dst; } -template SimpleTensor<uint8_t> reshape_layer(const SimpleTensor<uint8_t> &src, const TensorShape &output_shape); -template SimpleTensor<int8_t> reshape_layer(const SimpleTensor<int8_t> &src, const TensorShape &output_shape); +template SimpleTensor<uint8_t> reshape_layer(const SimpleTensor<uint8_t> &src, const TensorShape &output_shape); +template SimpleTensor<int8_t> reshape_layer(const SimpleTensor<int8_t> &src, const TensorShape &output_shape); template SimpleTensor<uint16_t> reshape_layer(const SimpleTensor<uint16_t> &src, const TensorShape &output_shape); -template SimpleTensor<int16_t> reshape_layer(const SimpleTensor<int16_t> &src, const TensorShape &output_shape); +template SimpleTensor<int16_t> reshape_layer(const SimpleTensor<int16_t> &src, const TensorShape &output_shape); template SimpleTensor<uint32_t> reshape_layer(const SimpleTensor<uint32_t> &src, const TensorShape &output_shape); -template SimpleTensor<int32_t> reshape_layer(const SimpleTensor<int32_t> &src, const TensorShape &output_shape); -template SimpleTensor<half> reshape_layer(const SimpleTensor<half> &src, const TensorShape &output_shape); -template SimpleTensor<float> reshape_layer(const SimpleTensor<float> &src, const TensorShape &output_shape); +template SimpleTensor<int32_t> reshape_layer(const SimpleTensor<int32_t> &src, const TensorShape &output_shape); +template SimpleTensor<half> reshape_layer(const SimpleTensor<half> &src, const TensorShape &output_shape); +template SimpleTensor<float> reshape_layer(const SimpleTensor<float> &src, const TensorShape &output_shape); +template SimpleTensor<bfloat16> reshape_layer(const SimpleTensor<bfloat16> &src, const TensorShape &output_shape); /** [ReshapeLayer] **/ } // namespace reference } // namespace validation diff --git a/tests/validation/reference/Reverse.cpp b/tests/validation/reference/Reverse.cpp index c6c4614278..7924f900d1 100644 --- a/tests/validation/reference/Reverse.cpp +++ b/tests/validation/reference/Reverse.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2020, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -35,8 +35,9 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> reverse(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> &axis) +SimpleTensor<T> reverse(const SimpleTensor<T> &src, const SimpleTensor<int32_t> &axis, bool use_inverted_axis) { + ARM_COMPUTE_ERROR_ON(src.shape().num_dimensions() > 4); ARM_COMPUTE_ERROR_ON(axis.shape().num_dimensions() > 1); ARM_COMPUTE_ERROR_ON(axis.shape().x() > 4); @@ -48,10 +49,32 @@ SimpleTensor<T> reverse(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> const unsigned int depth = src.shape()[2]; const unsigned int batches = src.shape()[3]; + const int rank = src.shape().num_dimensions(); + std::array<bool, 4> to_reverse = { { false, false, false, false } }; for(int i = 0; i < axis.num_elements(); ++i) { - to_reverse[axis[i]] = true; + int axis_i = axis[i]; + + // The values of axis tensor must be between [-rank, rank-1]. + if((axis_i < -rank) || (axis_i >= rank)) + { + ARM_COMPUTE_ERROR("the values of the axis tensor must be within [-rank, rank-1]."); + } + + // In case of negative axis value i.e targeted axis(i) = rank + axis(i) + if(axis_i < 0) + { + axis_i = rank + axis_i; + } + + // Reverse ACL axis indices convention i.e. (inverted)axis = (tensor_rank - 1) - axis + if(use_inverted_axis) + { + axis_i = (rank - 1) - axis_i; + } + + to_reverse[axis_i] = true; } const uint32_t num_elements = src.num_elements(); @@ -73,9 +96,9 @@ SimpleTensor<T> reverse(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> return dst; } -template SimpleTensor<uint8_t> reverse(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint32_t> &axis); -template SimpleTensor<half> reverse(const SimpleTensor<half> &src, const SimpleTensor<uint32_t> &axis); -template SimpleTensor<float> reverse(const SimpleTensor<float> &src, const SimpleTensor<uint32_t> &axis); +template SimpleTensor<uint8_t> reverse(const SimpleTensor<uint8_t> &src, const SimpleTensor<int32_t> &axis, bool use_inverted_axis); +template SimpleTensor<half> reverse(const SimpleTensor<half> &src, const SimpleTensor<int32_t> &axis, bool use_inverted_axis); +template SimpleTensor<float> reverse(const SimpleTensor<float> &src, const SimpleTensor<int32_t> &axis, bool use_inverted_axis); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Reverse.h b/tests/validation/reference/Reverse.h index 4a28da7270..30926b05a5 100644 --- a/tests/validation/reference/Reverse.h +++ b/tests/validation/reference/Reverse.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2019 Arm Limited. + * Copyright (c) 2018-2019, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,8 +21,8 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_TEST_REVERSE_H -#define ARM_COMPUTE_TEST_REVERSE_H +#ifndef ACL_TESTS_VALIDATION_REFERENCE_REVERSE_H +#define ACL_TESTS_VALIDATION_REFERENCE_REVERSE_H #include "tests/SimpleTensor.h" @@ -35,9 +35,9 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> reverse(const SimpleTensor<T> &src, const SimpleTensor<uint32_t> &axis); +SimpleTensor<T> reverse(const SimpleTensor<T> &src, const SimpleTensor<int32_t> &axis, bool use_inverted_axis = false); } // namespace reference } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_REVERSE_H */ +#endif // ACL_TESTS_VALIDATION_REFERENCE_REVERSE_H diff --git a/tests/validation/reference/Scale.cpp b/tests/validation/reference/Scale.cpp index 71e98fd776..2f429cb29b 100644 --- a/tests/validation/reference/Scale.cpp +++ b/tests/validation/reference/Scale.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -25,7 +25,6 @@ #include "Scale.h" #include "Utils.h" -#include "arm_compute/core/utils/misc/Utility.h" #include "src/core/utils/ScaleUtils.h" #include "support/Rounding.h" @@ -183,14 +182,15 @@ SimpleTensor<T> scale_core(const SimpleTensor<T> &in, float scale_x, float scale template <typename T> SimpleTensor<T> scale(const SimpleTensor<T> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, T constant_border_value, - SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners) + SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info) { + ARM_COMPUTE_UNUSED(output_quantization_info); return scale_core<T>(src, scale_x, scale_y, policy, border_mode, constant_border_value, sampling_policy, ceil_policy_scale, align_corners); } template <> SimpleTensor<uint8_t> scale(const SimpleTensor<uint8_t> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, uint8_t constant_border_value, - SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners) + SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info) { SimpleTensor<uint8_t> dst; if(src.quantization_info().uniform().scale != 0.f) @@ -198,7 +198,7 @@ SimpleTensor<uint8_t> scale(const SimpleTensor<uint8_t> &src, float scale_x, flo SimpleTensor<float> src_tmp = convert_from_asymmetric(src); float constant_border_value_f = dequantize_qasymm8(constant_border_value, src.quantization_info()); SimpleTensor<float> dst_tmp = scale_core<float>(src_tmp, scale_x, scale_y, policy, border_mode, constant_border_value_f, sampling_policy, ceil_policy_scale, align_corners); - dst = convert_to_asymmetric<uint8_t>(dst_tmp, src.quantization_info()); + dst = convert_to_asymmetric<uint8_t>(dst_tmp, output_quantization_info); } else { @@ -209,7 +209,7 @@ SimpleTensor<uint8_t> scale(const SimpleTensor<uint8_t> &src, float scale_x, flo template <> SimpleTensor<int8_t> scale(const SimpleTensor<int8_t> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, int8_t constant_border_value, - SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners) + SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info) { SimpleTensor<int8_t> dst; if(src.quantization_info().uniform().scale != 0.f) @@ -217,7 +217,7 @@ SimpleTensor<int8_t> scale(const SimpleTensor<int8_t> &src, float scale_x, float SimpleTensor<float> src_tmp = convert_from_asymmetric(src); float constant_border_value_f = dequantize_qasymm8_signed(constant_border_value, src.quantization_info()); SimpleTensor<float> dst_tmp = scale_core<float>(src_tmp, scale_x, scale_y, policy, border_mode, constant_border_value_f, sampling_policy, ceil_policy_scale, align_corners); - dst = convert_to_asymmetric<int8_t>(dst_tmp, src.quantization_info()); + dst = convert_to_asymmetric<int8_t>(dst_tmp, output_quantization_info); } else { @@ -227,11 +227,11 @@ SimpleTensor<int8_t> scale(const SimpleTensor<int8_t> &src, float scale_x, float } template SimpleTensor<int16_t> scale(const SimpleTensor<int16_t> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, int16_t constant_border_value, - SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners); + SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info); template SimpleTensor<half> scale(const SimpleTensor<half> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, half constant_border_value, - SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners); + SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info); template SimpleTensor<float> scale(const SimpleTensor<float> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, float constant_border_value, - SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners); + SamplingPolicy sampling_policy, bool ceil_policy_scale, bool align_corners, QuantizationInfo output_quantization_info); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Scale.h b/tests/validation/reference/Scale.h index c66af8d94e..c32c07d1c0 100644 --- a/tests/validation/reference/Scale.h +++ b/tests/validation/reference/Scale.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 Arm Limited. + * Copyright (c) 2017-2020, 2022 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -37,7 +37,7 @@ namespace reference { template <typename T> SimpleTensor<T> scale(const SimpleTensor<T> &src, float scale_x, float scale_y, InterpolationPolicy policy, BorderMode border_mode, T constant_border_value = 0, - SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool ceil_policy_scale = false, bool align_corners = false); + SamplingPolicy sampling_policy = SamplingPolicy::CENTER, bool ceil_policy_scale = false, bool align_corners = false, QuantizationInfo output_quantization_info = QuantizationInfo()); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ScatterLayer.cpp b/tests/validation/reference/ScatterLayer.cpp new file mode 100644 index 0000000000..55c48a9002 --- /dev/null +++ b/tests/validation/reference/ScatterLayer.cpp @@ -0,0 +1,152 @@ +/* + * Copyright (c) 2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "ScatterLayer.h" +#include "tests/validation/Helpers.h" +#include "arm_compute/core/TensorShape.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +namespace +{ + +template <typename T> +T reduce_op(const T ¤t,const T &update,const ScatterFunction func) +{ + switch(func) + { + case ScatterFunction::Update: + return update; + break; + case ScatterFunction::Add: + return current + update; + break; + case ScatterFunction::Sub: + return current - update; + break; + case ScatterFunction::Max: + return std::max(current, update); + break; + case ScatterFunction::Min: + return std::min(current, update); + break; + default: + ARM_COMPUTE_ERROR("Unsupported Scatter function"); + break; + } +} + +template float reduce_op(const float ¤t,const float &update,const ScatterFunction func); +template half reduce_op(const half ¤t,const half &update,const ScatterFunction func); +} + +// NOTE: This function expects collapsed tensors as input. +// Batch dims for update/indices tensors should be collapsed into a single dim. +// Data dims should be collapsed into a single dim for both update and src tensors prior to calling this function. +template <typename T> +SimpleTensor<T> scatter_layer_internal(const SimpleTensor<T> &src, const SimpleTensor<T> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info) +{ + // 1. If zero initialization variable is false, copy src data to dst. + SimpleTensor<T> dst{ out_shape, src.data_type(), 1 }; + if(!info.zero_initialization) + { + std::copy_n(src.data(), src.num_elements(), dst.data()); + } + + // Number of elements between each value of the dim being iterated through + const unsigned int data_stride = updates.shape().total_size_lower(updates.shape().num_dimensions() - 1); + const unsigned int no_output_dims = out_shape.num_dimensions(); + + // Calculate output stride at given index for all output dims. + std::vector<unsigned int> out_stride_at_idx(no_output_dims); + for (unsigned int i = 0 ; i < no_output_dims; i++) + { + out_stride_at_idx[i] = out_shape.total_size_lower(i); + } + + const unsigned int indices_x_dim = static_cast<unsigned int>(indices.shape()[0]); + const unsigned int indices_y_dim = static_cast<unsigned int>(indices.shape()[1]); + + // 2. Iterate over indices tensor y-dim and replace sections of dst tensor with relevant areas of update tensor. + for(unsigned int i = 0; i < indices_y_dim; i++) + { + // NOTE : Currently, indices.shape() == [X, Y, 1, 1], where X is the indices dim and Y is the batch dim + // Starting index for both the update and indices tensors. + const unsigned int update_dim_start = i * data_stride; + const unsigned int indices_dim_start = i * indices_x_dim; + bool out_of_bounds = false; + unsigned int out_offset_acc = 0; + + // Iterate over each indices value for the relevant batch and accumulate the offset. + for(unsigned int j = 0; j < indices_x_dim; j++) + { + // Get first index value with i * indices_x_dim (iterating through y-dim/batch idx), then iterate through x dim by adding k + const int index_value = indices[indices_dim_start + j]; + const unsigned int out_dim = no_output_dims - (j+1); // Calculate corresponding output dim to current index value. + if(index_value < static_cast<int>(out_shape[out_dim]) && index_value >= 0) + { + out_offset_acc += (index_value * out_stride_at_idx[out_dim]); // offset accumulation + } + else + { + out_of_bounds = true; + break; + } + } + + // If not out of bounds, copy update tensor elements to output + if(!out_of_bounds) + { + for (unsigned int j = 0 ; j < data_stride; j++) + { + dst[out_offset_acc + j] = reduce_op(dst[out_offset_acc + j], updates[update_dim_start + j], info.func); + } + } + } + return dst; +} + +template <typename T> +SimpleTensor<T> scatter_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info) +{ + return scatter_layer_internal<T>(src, updates, indices, out_shape, info); +} + +template SimpleTensor<float> scatter_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<half> scatter_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<int32_t> scatter_layer(const SimpleTensor<int32_t> &src, const SimpleTensor<int32_t> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<uint32_t> scatter_layer(const SimpleTensor<uint32_t> &src, const SimpleTensor<uint32_t> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<int16_t> scatter_layer(const SimpleTensor<int16_t> &src, const SimpleTensor<int16_t> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<uint16_t> scatter_layer(const SimpleTensor<uint16_t> &src, const SimpleTensor<uint16_t> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<int8_t> scatter_layer(const SimpleTensor<int8_t> &src, const SimpleTensor<int8_t> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +template SimpleTensor<uint8_t> scatter_layer(const SimpleTensor<uint8_t> &src, const SimpleTensor<uint8_t> &updates, const SimpleTensor<int32_t> &indices, const TensorShape &out_shape, const ScatterInfo &info); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/ScatterLayer.h b/tests/validation/reference/ScatterLayer.h new file mode 100644 index 0000000000..97d5e70b0d --- /dev/null +++ b/tests/validation/reference/ScatterLayer.h @@ -0,0 +1,48 @@ +/* + * Copyright (c) 2024 Arm Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#ifndef ACL_TESTS_VALIDATION_REFERENCE_SCATTERLAYER_H +#define ACL_TESTS_VALIDATION_REFERENCE_SCATTERLAYER_H + +#include "Utils.h" +#include "arm_compute/function_info/ScatterInfo.h" +#include "tests/SimpleTensor.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<T> scatter_layer_internal(const SimpleTensor<T> &src, const SimpleTensor<T> &update, const SimpleTensor<int32_t> &indices, const TensorShape &shape, const ScatterInfo &info); + +template <typename T> +SimpleTensor<T> scatter_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &update, const SimpleTensor<int32_t> &indices, const TensorShape &shape, const ScatterInfo &info); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif // ACL_TESTS_VALIDATION_REFERENCE_SCATTERLAYER_H diff --git a/tests/validation/reference/UtilsQuantizedAsymm.h b/tests/validation/reference/UtilsQuantizedAsymm.h index 1f593bb696..e5ecc66545 100644 --- a/tests/validation/reference/UtilsQuantizedAsymm.h +++ b/tests/validation/reference/UtilsQuantizedAsymm.h @@ -32,6 +32,22 @@ namespace test { namespace validation { +namespace +{ +#if __clang__ +// This has been tested on clang 7.0.2 (__clang_major__ == 7 && __clang_minor__ == 0 && __clang_patchlevel__ == 2) +inline int64_t to_int64(int32_t val) +{ + return static_cast<int64_t>(val) | ((val < 0) ? (((1ll << 32) - 1) << 32) : 0); +} +#else // __clang__ +inline int64_t to_int64(int32_t val) +{ + return static_cast<int64_t>(val); +} +#endif // __clang__ +} // namespace + /** Rounded to nearest division by a power-of-two. */ inline int32_t asymm_rounding_divide_by_pow2(int32_t x, int exponent) { @@ -43,12 +59,12 @@ inline int32_t asymm_rounding_divide_by_pow2(int32_t x, int exponent) /** Multiplication of two integers. The same as ARMv7 Arm® Neon™ VQRDMULH instruction. */ inline int32_t asymm_int_mult(int32_t a, int32_t b) { - bool overflow = a == b && a == std::numeric_limits<int32_t>::min(); - int64_t a_64(a); - int64_t b_64(b); - int64_t ab_64 = a_64 * b_64; - int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30)); - int32_t ab_x2_high32 = static_cast<int32_t>((ab_64 + nudge) / (1ll << 31)); + const bool overflow = a == b && a == std::numeric_limits<int32_t>::min(); + const int64_t a_64 = to_int64(a); + const int64_t b_64 = to_int64(b); + const int64_t ab_64 = a_64 * b_64; + const int32_t nudge = ab_64 >= 0 ? (1 << 30) : (1 - (1 << 30)); + const int32_t ab_x2_high32 = static_cast<int32_t>((ab_64 + nudge) / (1ll << 31)); return overflow ? std::numeric_limits<int32_t>::max() : ab_x2_high32; } |