diff options
Diffstat (limited to 'tests')
| -rw-r--r-- | tests/datasets/ScatterDataset.h | 41 | ||||
| -rw-r--r-- | tests/validation/CL/ScatterLayer.cpp | 138 | ||||
| -rw-r--r-- | tests/validation/NEON/GEMMLowp.cpp | 12 | ||||
| -rw-r--r-- | tests/validation/NEON/ReorderLayer.cpp | 46 | ||||
| -rw-r--r-- | tests/validation/NEON/SoftmaxLayer.cpp | 2 | ||||
| -rw-r--r-- | tests/validation/NEON/UNIT/RuntimeContext.cpp | 15 | ||||
| -rw-r--r-- | tests/validation/fixtures/GEMMLowpFixture.h | 10 | ||||
| -rw-r--r-- | tests/validation/fixtures/ReorderFixture.h | 27 | ||||
| -rw-r--r-- | tests/validation/fixtures/ScatterLayerFixture.h | 59 | ||||
| -rw-r--r-- | tests/validation/reference/ScatterLayer.cpp | 9 |
10 files changed, 306 insertions, 53 deletions
diff --git a/tests/datasets/ScatterDataset.h b/tests/datasets/ScatterDataset.h index c0858941db..8fd4448d2d 100644 --- a/tests/datasets/ScatterDataset.h +++ b/tests/datasets/ScatterDataset.h @@ -179,10 +179,47 @@ public: // NOTE: Config is src, updates, indices, output. // NOTE: Updates/Indices tensors are now batched. // NOTE: indices.shape.x = (updates_batched) ? (src.num_dimensions - updates.num_dimensions) + 2 : (src.num_dimensions - updates.num_dimensions) + 1 + // k is the number of batch dimensions + // k = 2 add_config(TensorShape(6U, 5U), TensorShape(6U, 2U, 2U), TensorShape(1U, 2U, 2U), TensorShape(6U, 5U)); + add_config(TensorShape(5U, 5U, 4U, 2U, 2U), TensorShape(5U, 5U, 6U, 2U), TensorShape(3U, 6U, 2U), TensorShape(5U, 5U, 4U, 2U, 2U)); + + // k = 3 + add_config(TensorShape(6U, 5U), TensorShape(6U, 2U, 2U, 2U), TensorShape(1U, 2U, 2U, 2U), TensorShape(6U, 5U)); + add_config(TensorShape(5U, 5U, 4U, 2U, 2U), TensorShape(5U, 5U, 3U, 6U, 2U), TensorShape(3U, 3U, 6U, 2U), TensorShape(5U, 5U, 4U, 2U, 2U)); + + // k = 4 + add_config(TensorShape(5U, 5U, 4U, 2U, 2U), TensorShape(5U, 6U, 2U, 3U, 2U), TensorShape(4U, 6U, 2U, 3U, 2U), TensorShape(5U, 5U, 4U, 2U, 2U)); + + // k = 5 + add_config(TensorShape(5U, 5U, 4U, 2U, 2U), TensorShape(5U, 3U, 4U, 3U, 2U, 2U), TensorShape(4U, 3U, 4U, 3U, 2U, 2U), TensorShape(5U, 5U, 4U, 2U, 2U)); + } +}; + +class SmallScatterScalarDataset final : public ScatterDataset +{ +public: + // batched scalar case + SmallScatterScalarDataset() + { + add_config(TensorShape(6U, 5U), TensorShape(6U), TensorShape(2U, 6U), TensorShape(6U, 5U)); + add_config(TensorShape(6U, 5U), TensorShape(6U, 6U), TensorShape(2U, 6U, 6U), TensorShape(6U, 5U)); + add_config(TensorShape(3U, 3U, 6U, 5U), TensorShape(6U, 6U), TensorShape(4U, 6U, 6U), TensorShape(3U, 3U, 6U, 5U)); + } +}; + +// This dataset is for data types that does not require full testing. It contains selected tests from the above. +class SmallScatterMixedDataset final : public ScatterDataset +{ +public: + SmallScatterMixedDataset() + { + add_config(TensorShape(10U), TensorShape(2U), TensorShape(1U, 2U), TensorShape(10U)); + add_config(TensorShape(9U, 3U, 4U), TensorShape(9U, 3U, 2U), TensorShape(1U, 2U), TensorShape(9U, 3U, 4U)); + add_config(TensorShape(6U, 5U), TensorShape(6U, 6U), TensorShape(2U, 6U, 6U), TensorShape(6U, 5U)); + add_config(TensorShape(35U, 4U, 3U, 2U, 2U), TensorShape(35U, 4U), TensorShape(4U, 4U), TensorShape(35U, 4U, 3U, 2U, 2U)); + add_config(TensorShape(11U, 3U, 3U, 2U, 4U), TensorShape(11U, 3U, 3U, 4U), TensorShape(2U, 4U), TensorShape(11U, 3U, 3U, 2U, 4U)); add_config(TensorShape(6U, 5U, 2U), TensorShape(6U, 2U, 2U), TensorShape(2U, 2U, 2U), TensorShape(6U, 5U, 2U)); - add_config(TensorShape(6U, 5U, 2U, 2U), TensorShape(3U, 2U), TensorShape(4U, 3U, 2U), TensorShape(6U, 5U, 2U, 2U)); - add_config(TensorShape(5U, 5U, 4U, 2U, 2U), TensorShape(6U, 2U), TensorShape(5U, 6U, 2U), TensorShape(5U, 5U, 4U, 2U, 2U)); } }; } // namespace datasets diff --git a/tests/validation/CL/ScatterLayer.cpp b/tests/validation/CL/ScatterLayer.cpp index 4a2462c7d2..b1531eb64a 100644 --- a/tests/validation/CL/ScatterLayer.cpp +++ b/tests/validation/CL/ScatterLayer.cpp @@ -41,6 +41,8 @@ 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> @@ -53,6 +55,7 @@ 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. @@ -63,6 +66,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::PRECOMMIT, zip( }), 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), @@ -73,6 +77,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::PRECOMMIT, zip( }), 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), @@ -83,6 +88,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::PRECOMMIT, zip( }), 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), @@ -93,6 +99,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::PRECOMMIT, zip( }), 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), @@ -101,7 +108,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::PRECOMMIT, zip( ScatterInfo(ScatterFunction::Update, false), ScatterInfo(ScatterFunction::Update, false), }), - make("Expected", { false, true, true, false, false, false, false, false, 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); @@ -118,7 +125,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLScatterLayerFixture<float>, framework::Datase make("DataType", {DataType::F32}), allScatterFunctions, make("ZeroInit", {false}), - make("Inplace", {false}))) + make("Inplace", {false}), + make("Padding", {true}))) { validate(CLAccessor(_target), _reference, tolerance_f32); } @@ -129,7 +137,8 @@ FIXTURE_DATA_TEST_CASE(RunSmallZeroInit, CLScatterLayerFixture<float>, framework make("DataType", {DataType::F32}), make("ScatterFunction", {ScatterFunction::Add}), make("ZeroInit", {true}), - make("Inplace", {false}))) + make("Inplace", {false}), + make("Padding", {true}))) { validate(CLAccessor(_target), _reference, tolerance_f32); } @@ -140,7 +149,8 @@ FIXTURE_DATA_TEST_CASE(RunSmallMultiDim, CLScatterLayerFixture<float>, framework make("DataType", {DataType::F32}), allScatterFunctions, make("ZeroInit", {false}), - make("Inplace", {false}))) + make("Inplace", {false}), + make("Padding", {true}))) { validate(CLAccessor(_target), _reference, tolerance_f32); } @@ -151,24 +161,136 @@ FIXTURE_DATA_TEST_CASE(RunSmallMultiIndices, CLScatterLayerFixture<float>, frame make("DataType", {DataType::F32}), make("ScatterFunction", {ScatterFunction::Update, ScatterFunction::Add }), make("ZeroInit", {false}), - make("Inplace", {false, true}))) + 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::DISABLED, +FIXTURE_DATA_TEST_CASE(RunSmallBatchedMultiIndices, CLScatterLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::SmallScatterBatchedDataset(), make("DataType", {DataType::F32}), - make("ScatterFunction", {ScatterFunction::Update, ScatterFunction::Add }), + 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("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 diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp index 9b1da61ed7..d25f43a330 100644 --- a/tests/validation/NEON/GEMMLowp.cpp +++ b/tests/validation/NEON/GEMMLowp.cpp @@ -360,13 +360,21 @@ TEST_SUITE_END() // DynamicQuantization // 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, datasets::SmallGEMMLowpDataset()) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpDequantizedMatrixMultiplyValidationFixture, framework::DatasetMode::ALL, + combine( + datasets::SmallGEMMLowpDataset(), + make("accumulate", {true, false}) + )) { // Validate output validate(Accessor(_target), _reference, tolerance_dequantized); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpDequantizedMatrixMultiplyValidationFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpDataset()) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpDequantizedMatrixMultiplyValidationFixture, framework::DatasetMode::NIGHTLY, + combine( + datasets::LargeGEMMLowpDataset(), + make("accumulate", {false}) + )) { // Validate output validate(Accessor(_target), _reference, tolerance_dequantized); diff --git a/tests/validation/NEON/ReorderLayer.cpp b/tests/validation/NEON/ReorderLayer.cpp index 42fa0f8b00..839ad0ac92 100644 --- a/tests/validation/NEON/ReorderLayer.cpp +++ b/tests/validation/NEON/ReorderLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2023 Arm Limited. + * Copyright (c) 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -33,6 +33,7 @@ #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 { @@ -40,6 +41,8 @@ namespace test { namespace validation { +using framework::dataset::make; + TEST_SUITE(NEON) TEST_SUITE(ReorderLayer) @@ -48,13 +51,46 @@ using NEReorderLayerAlias = ReorderValidationFixture<Tensor, Accessor, NEReorder TEST_SUITE(FP32) #if defined(ARM_COMPUTE_ENABLE_SVE) -FIXTURE_DATA_TEST_CASE(RunBlock8, NEReorderLayerAlias<float>, framework::DatasetMode::ALL, combine(datasets::ReorderLayerDatasetBlock8(), framework::dataset::make("DataType", DataType::F32))) +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 - validate(Accessor(_target), _reference); + 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(), framework::dataset::make("DataType", DataType::F32))) + +FIXTURE_DATA_TEST_CASE(RunBlock4, NEReorderLayerAlias<float>, framework::DatasetMode::ALL, combine(datasets::ReorderLayerDatasetBlock4(), make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference); @@ -68,4 +104,4 @@ TEST_SUITE_END() // NEON } // namespace test } // namespace arm_compute -#endif // defined(__aarch64__)
\ No newline at end of file +#endif // defined(__aarch64__) diff --git a/tests/validation/NEON/SoftmaxLayer.cpp b/tests/validation/NEON/SoftmaxLayer.cpp index 8da5a0d953..94d0866c38 100644 --- a/tests/validation/NEON/SoftmaxLayer.cpp +++ b/tests/validation/NEON/SoftmaxLayer.cpp @@ -145,7 +145,7 @@ DATA_TEST_CASE(KernelSelection, framework::DatasetMode::ALL, cpu_isa.fp16 = (data_type == DataType::F16); const auto *selected_impl = CpuSoftmaxKernel::get_implementation( - SoftmaxKernelDataTypeISASelectorData{ data_type, cpu_isa, false /* is_log */, 0 /* axis */}, + 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); diff --git a/tests/validation/NEON/UNIT/RuntimeContext.cpp b/tests/validation/NEON/UNIT/RuntimeContext.cpp index 819811943d..e0d45c639a 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 +48,19 @@ 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 nonlittle_num_cpus = ci.get_cpu_num_excluding_little(); + const uint32_t num_threads = arm_compute::Scheduler::get().num_threads(); + + 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/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h index 6b7cbba92e..aa4eedb75d 100644 --- a/tests/validation/fixtures/GEMMLowpFixture.h +++ b/tests/validation/fixtures/GEMMLowpFixture.h @@ -472,15 +472,9 @@ template <typename TensorType, typename AccessorType, typename FunctionType, boo 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) + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, bool accumulate) { - // Accumulation is supported for Int8/UInt8 only in aarch64 - bool accumulate = true; - // Accumulation is not supported for Int8/UInt8 in aarch32 -#ifdef __arm__ - accumulate = false; -#endif //__arm__ - bool dynamic_qinfo = false; + 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; diff --git a/tests/validation/fixtures/ReorderFixture.h b/tests/validation/fixtures/ReorderFixture.h index 36e62696bc..8e28484c48 100644 --- a/tests/validation/fixtures/ReorderFixture.h +++ b/tests/validation/fixtures/ReorderFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2023 Arm Limited. + * Copyright (c) 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 ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE -#define ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE +#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" @@ -32,6 +32,7 @@ #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 { @@ -44,10 +45,23 @@ template <typename TensorType, typename AccessorType, typename FunctionType, typ 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) { - _target = compute_target(input_shape, output_shape, input_wf, output_wf, data_type); - _reference = compute_reference(input_shape, output_shape, output_wf, 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: @@ -98,6 +112,7 @@ public: return reference::reorder_layer<T>(src, output_shape, output_wf); } + bool _hardware_supports = true; TensorType _target{}; SimpleTensor<T> _reference{}; }; @@ -105,4 +120,4 @@ public: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE */ +#endif // ACL_TESTS_VALIDATION_FIXTURES_REORDERFIXTURE_H diff --git a/tests/validation/fixtures/ScatterLayerFixture.h b/tests/validation/fixtures/ScatterLayerFixture.h index 4fb2d7f127..af161ef98b 100644 --- a/tests/validation/fixtures/ScatterLayerFixture.h +++ b/tests/validation/fixtures/ScatterLayerFixture.h @@ -48,7 +48,7 @@ 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, + 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 @@ -57,19 +57,36 @@ public: + 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, src_qinfo, o_qinfo); + _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, float lo = -10.f, float hi = 10.f) + void fill(U &&tensor, int i) { switch(tensor.data_type()) { case DataType::F32: + case DataType::F16: { - std::uniform_real_distribution<float> distribution(lo, hi); + 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; } @@ -86,12 +103,12 @@ protected: 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::max({shape[0] , shape[1], shape[2]}); - library->fill_tensor_uniform(tensor, i, static_cast<int32_t>(-2), static_cast<int32_t>(max)); + 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, + 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. @@ -129,11 +146,14 @@ protected: ARM_COMPUTE_ASSERT(indices.info()->is_resizable()); ARM_COMPUTE_ASSERT(dst.info()->is_resizable()); - add_padding_x({ &src, &updates, &indices}); - - if(!inplace) + if(padding) { - add_padding_x({ &dst }); + add_padding_x({ &src, &updates, &indices}); + + if(!inplace) + { + add_padding_x({ &dst }); + } } // Allocate tensors @@ -180,12 +200,13 @@ protected: 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(c_shape.num_dimensions() >= 3) + if(num_ind_dims >= 3) { indices_shape = indices_shape.collapsed_from(1); - updates_shape = updates_shape.collapsed_from(updates_shape.num_dimensions() - 2); // Collapses from last 2 dims + 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. @@ -195,16 +216,16 @@ protected: 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{ a_shape, data_type, 1, a_qinfo }; - SimpleTensor<T> updates{b_shape, data_type, 1, QuantizationInfo() }; - SimpleTensor<int32_t> indices{ c_shape, DataType::S32, 1, QuantizationInfo() }; + 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. + // Calculate individual reference using collapsed shapes return reference::scatter_layer<T>(src, updates, indices, out_shape, info); } @@ -219,10 +240,10 @@ class ScatterValidationFixture : public ScatterGenericValidationFixture<TensorTy { 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) + 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, + indices_shape, out_shape, data_type, ScatterInfo(func, zero_init), inplace, padding, QuantizationInfo(), QuantizationInfo()); } }; diff --git a/tests/validation/reference/ScatterLayer.cpp b/tests/validation/reference/ScatterLayer.cpp index 283022e8e2..55c48a9002 100644 --- a/tests/validation/reference/ScatterLayer.cpp +++ b/tests/validation/reference/ScatterLayer.cpp @@ -63,6 +63,7 @@ T reduce_op(const T ¤t,const T &update,const ScatterFunction func) } 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. @@ -138,7 +139,13 @@ SimpleTensor<T> scatter_layer(const SimpleTensor<T> &src, const SimpleTensor<T> } 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 |