diff options
Diffstat (limited to 'tests')
-rw-r--r-- | tests/CMakeLists.txt | 1 | ||||
-rw-r--r-- | tests/validation/CL/ConvolutionLayer.cpp | 291 | ||||
-rw-r--r-- | tests/validation/CL/GEMMMatrixMultiplyNative.cpp | 244 | ||||
-rw-r--r-- | tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp | 421 | ||||
-rw-r--r-- | tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp | 294 | ||||
-rw-r--r-- | tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp | 10 | ||||
-rw-r--r-- | tests/validation/fixtures/GEMMFixture.h | 740 | ||||
-rw-r--r-- | tests/validation/reference/PostOps.cpp | 93 | ||||
-rw-r--r-- | tests/validation/reference/PostOps.h | 47 |
9 files changed, 40 insertions, 2101 deletions
diff --git a/tests/CMakeLists.txt b/tests/CMakeLists.txt index c4b12e770a..3f2223596f 100644 --- a/tests/CMakeLists.txt +++ b/tests/CMakeLists.txt @@ -67,7 +67,6 @@ target_sources( validation/reference/L2NormalizeLayer.cpp validation/reference/ActivationLayer.cpp validation/reference/SpaceToBatch.cpp - validation/reference/PostOps.cpp validation/reference/Im2Col.cpp validation/reference/DequantizationLayer.cpp validation/reference/DeconvolutionLayer.cpp diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp index bced540d2a..986d76708d 100644 --- a/tests/validation/CL/ConvolutionLayer.cpp +++ b/tests/validation/CL/ConvolutionLayer.cpp @@ -22,7 +22,6 @@ * SOFTWARE. */ #include "arm_compute/core/Types.h" -#include "arm_compute/core/experimental/PostOps.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" @@ -69,53 +68,30 @@ constexpr float tolerance_num = 0.07f; /**< T const auto CNNDataTypes = framework::dataset::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", { DataType::F16, - DataType::F32 + DataType::F32 }); -const auto ActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +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, 0.5f) + 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", { ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 0.5f) }); - -bool is_post_op_list_valid_in_gemmconv(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &output_shape, DataType data_type, DataLayout data_layout, - const PadStrideInfo &conv_info, const experimental::PostOpList<ITensorInfo *> &post_ops) -{ - 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); - const int idx_kernels = get_data_layout_dimension_index(data_layout, DataLayoutDimension::BATCHES); - - const auto dilation = Size2D(1U, 1U); - const unsigned int num_groups = 1U; - - TensorInfo input_info(input_shape, 1, data_type, data_layout); - TensorInfo weights_info(weights_shape, 1, data_type, data_layout); - - TensorInfo output_info(output_shape, 1, data_type, data_layout); - - WeightsInfo w_info(false, weights_info.dimension(idx_width), weights_info.dimension(idx_height), weights_info.dimension(idx_kernels)); - - const auto status = CLGEMMConvolutionLayer::validate(&input_info.clone()->set_is_resizable(true), - &weights_info.clone()->set_is_resizable(true), nullptr, &output_info.clone()->set_is_resizable(true), - conv_info, w_info, dilation, ActivationLayerInfo(), num_groups, post_ops); - return bool(status); -} } // namespace TEST_SUITE(CL) @@ -207,72 +183,6 @@ DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(z enable_fast_math); ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); } - -DATA_TEST_CASE(ValidatePostOpSupportInConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( - framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 17U, 31U), 1, DataType::F32, DataLayout::NHWC), // Select GEMM - TensorInfo(TensorShape(17U, 31U, 32U), 1, DataType::F32, DataLayout::NCHW), // Select WINOGRAD - TensorInfo(TensorShape(27U, 27U, 48U), 1, DataType::F32, DataLayout::NCHW), // Select Direct - TensorInfo(TensorShape(27U, 27U, 48U), 1, DataType::F32, DataLayout::NCHW), // Select FFT - }), - framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(2U, 1U, 1U, 19U), 1, DataType::F32, DataLayout::NHWC), - TensorInfo(TensorShape(5U, 5U, 32U, 19U), 1, DataType::F32, DataLayout::NCHW), - TensorInfo(TensorShape(5U, 5U, 48U, 128U), 1, DataType::F32, DataLayout::NCHW), - TensorInfo(TensorShape(11U, 11U, 48U, 24), 1, DataType::F32, DataLayout::NCHW), - })), - framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(19U, 17U, 31U), 1, DataType::F32, DataLayout::NHWC), - TensorInfo(TensorShape(17U, 31U, 19U), 1, DataType::F32, DataLayout::NCHW), - TensorInfo(TensorShape(27U, 27U, 128U), 1, DataType::F32, DataLayout::NCHW), - TensorInfo(TensorShape(27U, 27U, 24U), 1, DataType::F32, DataLayout::NCHW), - })), - framework::dataset::make("ConvInfo", { PadStrideInfo(1U, 1U, 0U, 0U), - PadStrideInfo(1U, 1U, 2U, 2U), - PadStrideInfo(1U, 1U, 2U, 2U), - PadStrideInfo(1U, 1U, 5U, 5U), - })), - framework::dataset::make("EnableFastMath", { false, true, false, false})), - framework::dataset::make("ExpectedMethod",{ ConvolutionMethod::GEMM, - ConvolutionMethod::WINOGRAD, - ConvolutionMethod::DIRECT, - ConvolutionMethod::FFT, - })), - framework::dataset::make("PostOpSupported",{ true, false, false, false - })), - input_info, weights_info, output_info, conv_info, enable_fast_math, expected_method, post_op_supported) -{ - const int idx_width = get_data_layout_dimension_index(input_info.data_layout(), DataLayoutDimension::WIDTH); - const int idx_height = get_data_layout_dimension_index(input_info.data_layout(), DataLayoutDimension::HEIGHT); - const int idx_kernels = get_data_layout_dimension_index(input_info.data_layout(), DataLayoutDimension::BATCHES); - - const auto dilation = Size2D(1U, 1U); - const unsigned int num_groups = 1U; - - WeightsInfo w_info(false, weights_info.dimension(idx_width), weights_info.dimension(idx_height), weights_info.dimension(idx_kernels)); - - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<ITensorInfo*>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - - ConvolutionMethod actual_method = CLConvolutionLayer::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(), - ActivationLayerInfo(), - GPUTarget::BIFROST, - dilation, - enable_fast_math); - ARM_COMPUTE_EXPECT(actual_method == expected_method, framework::LogLevel::ERRORS); - const auto is_valid = CLConvolutionLayer::validate(&input_info.clone()->set_is_resizable(true), - &weights_info.clone()->set_is_resizable(true), - nullptr, - &output_info.clone()->set_is_resizable(true), - conv_info, - w_info, - dilation, - ActivationLayerInfo(), - enable_fast_math, - num_groups, - post_ops); - ARM_COMPUTE_EXPECT( bool(is_valid) == post_op_supported, framework::LogLevel::ERRORS); -} // clang-format on // *INDENT-ON* TEST_SUITE_END() // ConvolutionLayer @@ -285,167 +195,11 @@ using CLGEMMConvolutionLayerMixedDataLayoutFixture = ConvolutionValidationFixtur template <typename T> using CLConvolutionValidationWithPaddingFixture = ConvolutionValidationWithPaddingFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; -TEST_SUITE(ValidateFusedPostOpsConfigs) -TEST_SUITE(Invalid) -TEST_CASE(UnsupportedPostOpSequence, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const auto data_layout = DataLayout::NHWC; - const auto conv_info = PadStrideInfo(1, 1, 0, 0); - const auto input_shape = TensorShape(16U, 14U, 12U, 2U); - const auto weights_shape = TensorShape(16U, 1U, 1U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - const TensorShape post_op_arg0_shape(output_shape); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - auto post_op_arg1_info = post_op_arg_info.clone(); - - // Unsupported sequence of post ops - experimental::PostOpList<ITensorInfo *> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - post_op_arg1_info.get(), - 0, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OnlyNHWCIsSupported, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const auto data_layout = DataLayout::NCHW; - const auto conv_info = PadStrideInfo(1, 1, 0, 0); - const auto input_shape = TensorShape(14U, 12U, 16U, 2U); - const auto weights_shape = TensorShape(1U, 1U, 16U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - const TensorShape post_op_arg0_shape(output_shape); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - - experimental::PostOpList<ITensorInfo *> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OnlyFloatingTypeIsSupported, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::QASYMM8; - const auto data_layout = DataLayout::NHWC; - const auto conv_info = PadStrideInfo(1, 1, 0, 0); - const auto input_shape = TensorShape(16U, 14U, 12U, 2U); - const auto weights_shape = TensorShape(16U, 1U, 1U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - const TensorShape post_op_arg0_shape(output_shape); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - - experimental::PostOpList<ITensorInfo *> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OnlyConv1x1Stride1IsSupported_UnsupportedKernelSize, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const auto data_layout = DataLayout::NHWC; - const auto conv_info = PadStrideInfo(1, 1, 0, 0); - const auto input_shape = TensorShape(16U, 14U, 12U, 2U); - const auto weights_shape = TensorShape(16U, 3U, 3U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - const TensorShape post_op_arg0_shape(output_shape); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - - experimental::PostOpList<ITensorInfo *> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OnlyConv1x1Stride1IsSupported_UnsupportedStride, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const auto data_layout = DataLayout::NHWC; - const auto conv_info = PadStrideInfo(3, 3, 0, 0); - const auto input_shape = TensorShape(16U, 14U, 12U, 2U); - const auto weights_shape = TensorShape(16U, 1U, 1U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - const TensorShape post_op_arg0_shape(output_shape); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - - experimental::PostOpList<ITensorInfo *> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Invalid -TEST_SUITE(Valid) -TEST_CASE(EmptyPostOpList, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const auto data_layout = DataLayout::NHWC; - const auto conv_info = PadStrideInfo(1, 1, 0, 0); - const auto input_shape = TensorShape(16U, 14U, 12U, 2U); - const auto weights_shape = TensorShape(16U, 1U, 1U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - experimental::PostOpList<ITensorInfo *> post_ops{}; - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(SupportedPostOps, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const auto data_layout = DataLayout::NHWC; - const auto conv_info = PadStrideInfo(1, 1, 0, 0); - const auto input_shape = TensorShape(16U, 14U, 12U, 2U); - const auto weights_shape = TensorShape(16U, 1U, 1U, 24U); - - const auto output_shape = misc::shape_calculator::compute_deep_convolution_shape(input_shape, data_layout, weights_shape, conv_info); - - TensorShape post_op_arg0_shape(output_shape); - post_op_arg0_shape[1] = 1; // Broadcast in "Y" (second) dimension - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - - experimental::PostOpList<ITensorInfo *> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo *>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid_in_gemmconv(input_shape, weights_shape, output_shape, data_type, data_layout, conv_info, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Valid -TEST_SUITE_END() // ValidateFusedPostOps 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 })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsSmallDataset)) { // Validate output @@ -456,10 +210,7 @@ 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 })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), ActivationFunctionsSmallDataset)) { // Validate output @@ -503,16 +254,16 @@ using CLGEMMConvolutionLayerQuantizedMixedDataLayoutFixture = ConvolutionValidat template <typename T> using CLGEMMConvolutionLayerQuantizedPerChannelFixture = ConvolutionValidationQuantizedPerChannelFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T, int8_t>; -const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationInfo", +const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationInfo", { ActivationLayerInfo(), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) + 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) + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f) }); TEST_SUITE(Quantized) @@ -520,8 +271,8 @@ TEST_SUITE(Quantized) const auto QuantizationData = framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10), - QuantizationInfo(0.3f, 3), - QuantizationInfo(1.1f, 10), + QuantizationInfo(0.3f, 3), + QuantizationInfo(1.1f, 10), }); TEST_SUITE(QASYMM8) @@ -637,9 +388,7 @@ 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 })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsSmallDataset)) { // Validate output @@ -661,9 +410,7 @@ 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 })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsSmallDataset)) { // Validate output diff --git a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp index 7f63a03371..0ddf43766f 100644 --- a/tests/validation/CL/GEMMMatrixMultiplyNative.cpp +++ b/tests/validation/CL/GEMMMatrixMultiplyNative.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2022 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -53,11 +53,6 @@ using CLGEMMMatrixMultiplyNative = CLSynthetizeOperator<ClGemmMatrixMultiplyNati template <typename T> using CLGEMMMatrixMultiplyNativeFixture = GEMMMatrixMultiplyNativeValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; -// Fixture for CLGEMMMatrixMultiplyNative with post ops -template <typename T> -using CLGEMMMatrixMultiplyNativeWithPostOpsFixture = - GEMMMatrixMultiplyNativeWithPostOpsValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; - // Fixture for CLGEMMMatrixMultiplyNative3D template <typename T> using CLGEMMMatrixMultiplyNative3DFixture = GEMMMatrixMultiplyNative3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMMatrixMultiplyNative>; @@ -146,105 +141,6 @@ const auto boundary_handling_cases = combine(combine(combine(combine(combine(com broadcast_bias_values), framework::dataset::make("Activation", ActivationLayerInfo())); -/** Post Ops */ -using PostOpArgBroadcast = CLGEMMMatrixMultiplyNativeWithPostOpsFixture<float>::PostOpArgBroadcast; -experimental::PostOpList<PostOpArgBroadcast> post_ops_1() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(true, true, false), // If broadcast in dims 0, 1 and 2 - 0, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -experimental::PostOpList<PostOpArgBroadcast> post_ops_2() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(false, true, true), // If broadcast in dims 0, 1 and 2 - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -experimental::PostOpList<PostOpArgBroadcast> post_ops_3() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - // post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(false, false, false), // If broadcast in dims 0, 1 and 2 - 1, - ConvertPolicy::SATURATE); - return post_ops; -} -// To test that the output of the main op is the first parameter in prelu post op -experimental::PostOpList<PostOpArgBroadcast> post_ops_4() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>( - std::make_tuple(false, false, true), // If true, broadcast in corresponding dim: 0, 1 or 2 - 0, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -// To test that the output of the main op is the second parameter in prelu post op i.e. it is the alpha_param -experimental::PostOpList<PostOpArgBroadcast> post_ops_5() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>( - std::make_tuple(false, false, false), // If true, broadcast in corresponding dim: 0, 1 or 2 - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -/** Different Post Op Lists */ -const auto post_op_lists = framework::dataset::make("post_op_lists", { - post_ops_1(), - post_ops_2(), - post_ops_3(), - post_ops_4(), - post_ops_5() - } ); - -bool is_post_op_list_valid(unsigned int m, unsigned int n, unsigned int k, unsigned int batch, DataType data_type, const experimental::PostOpList<ITensorInfo*>& post_ops) -{ - const auto lhs_info = GEMMLHSMatrixInfo(4,4,1,false,true); - const auto rhs_info = GEMMRHSMatrixInfo(4,4,1,true,true,false); - - // Create TensorInfo for post op arguments - TensorInfo input0_info(TensorShape(k, m, batch), 1, data_type); - TensorInfo input1_info(TensorShape(n, k, batch), 1, data_type); - TensorInfo input2_info(TensorShape(n), 1, data_type); - TensorInfo output_info(TensorShape(n, m, batch), 1, data_type); - - GEMMKernelInfo gemm_info(m, n, k, 0 /**< Depth of the output tensor in case is reinterpreted as 3D */, - false /**< reinterpret the input as 3D */, - true /**< Flag used to broadcast the bias addition */, - false /**< wider accumm */, - false /**< has pad y */, - ActivationLayerInfo::ActivationFunction::IDENTITY, - 1 /**< Multiplication factor for the width of the 1xW transposed block */, - 1 /**< Multiplication factor for the height of the 4x4 interleaved block */, - lhs_info, - rhs_info, - 0 /**< Offset to be added to each element of the matrix A */, - 0 /**< Offset to be added to each element of the matrix B */, - post_ops); - return bool(ClGemmMatrixMultiplyNativeKernel::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, - lhs_info, - rhs_info, - gemm_info)); -} - /** Configuration test */ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, bool broadcast_bias, DataType data_type, const ActivationLayerInfo &act_info) { @@ -295,119 +191,6 @@ void validate_configuration(unsigned int m_value, unsigned int n_value, unsigned TEST_SUITE(CL) TEST_SUITE(GEMMMatrixMultiplyNative) -TEST_SUITE(ValidateFusedPostOpsConfigs) -TEST_SUITE(Invalid) -TEST_CASE(UnsupportedPostOpSequence, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 17; - const unsigned int n = 1; - const unsigned int k = 13; - const unsigned int batch = 2; - TensorShape post_op_arg0_shape(n, m, batch); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - auto post_op_arg1_info = post_op_arg_info.clone(); - - // Unsupported sequence of post ops - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( - post_op_arg1_info.get(), - 0, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OutputWidened, framework::DatasetMode::ALL) -{ - // Invalid broadcast: post op tensors "widen" the output tensor - const auto data_type = DataType::F32; - const unsigned int m = 1; - const unsigned int n = 18; - const unsigned int k = 13; - const unsigned int batch = 2; - TensorShape post_op_arg_shape(n, m + 1, batch); // output's Y dimension (m) is "widened", which is not allowed - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInXDimOnly, framework::DatasetMode::ALL) -{ - // Invalid broadcast: post op tensors broadcast in the first dimension (X) only - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, m, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Invalid -TEST_SUITE(Valid) -TEST_CASE(EmptyPostOpList, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - experimental::PostOpList<ITensorInfo*> post_ops{}; - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInYDimOnly, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(n, 1, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInBothXandYDims, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, 1, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInAllDims, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, 1, 1); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Valid -TEST_SUITE_END() // ValidateFusedPostOps TEST_SUITE(Float) TEST_SUITE(FP32) DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(combine(combine(combine( @@ -541,31 +324,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyNative3DFixture<float>, f validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyNativeWithPostOpsFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - m_values, - n_values), - k_values), - b_values), - framework::dataset::make("M0", { 4 })), - n0_values_precommit), - k0_values_precommit), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("alpha", {1.0f} )), - framework::dataset::make("beta", {1.0f} )), - framework::dataset::make("broadcast_bias", { false, true } )), - framework::dataset::make("Activation", { ActivationLayerInfo() })), - post_op_lists) - ) -{ - // Validate output - validate(CLAccessor(_target), _reference, rel_tolerance_f32, 0.f, abs_tolerance_f32); -} - -TEST_SUITE_END() // FusedPostOps - 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 cdd89670fa..b06e4bf213 100644 --- a/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp +++ b/tests/validation/CL/GEMMMatrixMultiplyReshaped.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2022 Arm Limited. + * Copyright (c) 2018-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,7 +23,6 @@ */ #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" -#include "arm_compute/core/experimental/PostOps.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" @@ -62,21 +61,11 @@ using CLGEMMMatrixMultiplyReshaped = CLSynthetizeOperator<ClGemmMatrixMultiplyRe template <typename T> using CLGEMMMatrixMultiplyReshapedFixture = GEMMMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>; -// Fixture for CLGEMMMatrixMultiplyReshaped with post ops -template <typename T> -using CLGEMMMatrixMultiplyReshapedWithPostOpsFixture = - GEMMMatrixMultiplyReshapedWithPostOpsValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>; - // Fixture for CLGEMMMatrixMultiplyReshaped mixed precision template <typename T> using CLGEMMMatrixMultiplyReshapedMixedPrecisionFixture = GEMMMatrixMultiplyReshapedValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped, true>; -// Fixture for CLGEMMMatrixMultiplyReshaped mixed precision with post ops -template <typename T> -using CLGEMMMatrixMultiplyReshapedMixedPrecisionWithPostOpsFixture = - GEMMMatrixMultiplyReshapedWithPostOpsValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped, true>; - // Fixture for CLGEMMMatrixMultiplyReshaped3D template <typename T> using CLGEMMMatrixMultiplyReshaped3DFixture = GEMMMatrixMultiplyReshaped3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeLHSMatrix, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshaped>; @@ -184,108 +173,6 @@ const auto broadcast_bias_values = framework::dataset::make("broadcast_bias", { /** LHS transposed values */ const auto lhs_transpose_values = framework::dataset::make("lhs_transpose", { false, true } ); -/** Post Ops */ -using PostOpArgBroadcast = CLGEMMMatrixMultiplyReshapedWithPostOpsFixture<float>::PostOpArgBroadcast; -experimental::PostOpList<PostOpArgBroadcast> post_ops_1() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(true, true, false), // If broadcast in dims 0, 1 and 2 - 0, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -experimental::PostOpList<PostOpArgBroadcast> post_ops_2() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(false, true, true), // If broadcast in dims 0, 1 and 2 - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -experimental::PostOpList<PostOpArgBroadcast> post_ops_3() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(false, false, true), // If broadcast in dims 0, 1 and 2 - 1, - ConvertPolicy::SATURATE); - return post_ops; -} -// To test that the output of the main op is the first parameter in prelu post op -experimental::PostOpList<PostOpArgBroadcast> post_ops_4() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>( - std::make_tuple(false, false, true), // If true, broadcast in corresponding dim: 0, 1 or 2 - 0, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -// To test that the output of the main op is the second parameter in prelu post op i.e. it is the alpha_param -experimental::PostOpList<PostOpArgBroadcast> post_ops_5() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>( - std::make_tuple(false, false, false), // If true, broadcast in corresponding dim: 0, 1 or 2 - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -/** Different Post Op Lists */ -const auto post_op_lists = framework::dataset::make("post_op_lists", { - post_ops_1(), - post_ops_2(), - post_ops_3(), - post_ops_4(), - post_ops_5() - } ); - -bool is_post_op_list_valid(unsigned int m, unsigned int n, unsigned int k, unsigned int batch, DataType data_type, const experimental::PostOpList<ITensorInfo*>& post_ops) -{ - const auto lhs_info = GEMMLHSMatrixInfo(4,4,1,false,true); - const auto rhs_info = GEMMRHSMatrixInfo(4,4,1,true,true,false); - - // Create TensorInfo for post op arguments - TensorInfo input0_info(TensorShape(k, m, batch), 1, data_type); - TensorInfo input1_info(TensorShape(n, k, batch), 1, data_type); - TensorInfo input2_info(TensorShape(n), 1, data_type); - TensorInfo output_info(TensorShape(n, m, batch), 1, data_type); - - const TensorInfo reshaped_input0_info = input0_info.clone()->set_tensor_shape(misc::shape_calculator::compute_lhs_reshaped_shape(input0_info, lhs_info)); - const TensorInfo reshaped_input1_info = input1_info.clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(input1_info, rhs_info)); - - GEMMKernelInfo gemm_info(m, n, k, 0 /**< Depth of the output tensor in case is reinterpreted as 3D */, - false /**< reinterpret the input as 3D */, - true /**< Flag used to broadcast the bias addition */, - false /**< wider accumm */, - false /**< has pad y */, - ActivationLayerInfo::ActivationFunction::IDENTITY, - 1 /**< Multiplication factor for the width of the 1xW transposed block */, - 1 /**< Multiplication factor for the height of the 4x4 interleaved block */, - lhs_info, - rhs_info, - 0 /**< Offset to be added to each element of the matrix A */, - 0 /**< Offset to be added to each element of the matrix B */, - post_ops); - return bool(ClGemmMatrixMultiplyReshapedKernel::validate(&reshaped_input0_info.clone()->set_is_resizable(true), - &reshaped_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, - lhs_info, - rhs_info, - gemm_info)); -} - } // namespace TEST_SUITE(CL) @@ -450,119 +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(ValidateFusedPostOpsConfigs) -TEST_SUITE(Invalid) -TEST_CASE(UnsupportedPostOpSequence, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 17; - const unsigned int n = 1; - const unsigned int k = 13; - const unsigned int batch = 2; - TensorShape post_op_arg0_shape(n, m, batch); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - auto post_op_arg1_info = post_op_arg_info.clone(); - - // Unsupported sequence of post ops - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( - post_op_arg1_info.get(), - 0, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OutputWidened, framework::DatasetMode::ALL) -{ - // Invalid broadcast: post op tensors "widen" the output tensor - const auto data_type = DataType::F32; - const unsigned int m = 17; - const unsigned int n = 1; - const unsigned int k = 13; - const unsigned int batch = 2; - TensorShape post_op_arg_shape(n + 4, m, batch); // output's X dimension (n) is "widened", which is not allowed - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInXDimOnly, framework::DatasetMode::ALL) -{ - // Invalid broadcast: post op tensors broadcast in the first dimension (X) only - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, m, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Invalid -TEST_SUITE(Valid) -TEST_CASE(EmptyPostOpList, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - experimental::PostOpList<ITensorInfo*> post_ops{}; - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInYDimOnly, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(n, 1, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInBothXandYDims, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, 1, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInAllDims, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, 1, 1); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Valid -TEST_SUITE_END() // ValidateFusedPostOps + TEST_SUITE(Float) TEST_SUITE(FP32) @@ -697,44 +472,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::ARM_COMPUTE_PRINT_INFO(); } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedWithPostOpsFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(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), - v0_values_precommit), - h0_values_precommit), - framework::dataset::make("interleave_lhs", { false })), - framework::dataset::make("interleave_rhs", { false })), - framework::dataset::make("export_to_cl_image_rhs", false)), - framework::dataset::make("DataType", DataType::F32)), - a_values_precommit), - beta_values_precommit), - framework::dataset::make("broadcast_bias", { true } )), - lhs_transpose_values), - act_values), - post_op_lists) - ) -{ - // Validate output - 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_END() // FusedPostOps TEST_SUITE(ExportToCLImage) DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zip( @@ -1002,44 +739,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<float>, framework::ARM_COMPUTE_PRINT_INFO(); } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedWithPostOpsFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(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), - v0_values_precommit), - h0_values_precommit), - framework::dataset::make("interleave_lhs", { false })), - framework::dataset::make("interleave_rhs", { false })), - framework::dataset::make("export_to_cl_image_rhs", true)), - framework::dataset::make("DataType", DataType::F32)), - a_values_precommit), - beta_values_precommit), - framework::dataset::make("broadcast_bias", { true } )), - lhs_transpose_values), - act_values), - post_op_lists) - ) -{ - // Validate output only if validate() is successful - 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_END() // FusedPostOps TEST_SUITE_END() // ExportToCLImage TEST_SUITE_END() // FP32 @@ -1178,45 +877,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedWithPostOpsFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(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), - v0_values_precommit), - h0_values_precommit), - framework::dataset::make("interleave_lhs", { false })), - framework::dataset::make("interleave_rhs", { false })), - framework::dataset::make("export_to_cl_image_rhs", false)), - framework::dataset::make("DataType", DataType::F16)), - a_values_precommit), - beta_values_precommit), - framework::dataset::make("broadcast_bias", { true } )), - lhs_transpose_values), - act_values), - post_op_lists) - ) -{ - // Validate output - 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_END() // FusedPostOps - 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::F16), // OK or incorrect if cl_khr_image2d_from_buffer not supported @@ -1483,44 +1143,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DFixture<half>, framework::ARM_COMPUTE_PRINT_INFO(); } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedWithPostOpsFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(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), - v0_values_precommit), - h0_values_precommit), - framework::dataset::make("interleave_lhs", { false })), - framework::dataset::make("interleave_rhs", { false })), - framework::dataset::make("export_to_cl_image_rhs", true)), - framework::dataset::make("DataType", DataType::F16)), - a_values_precommit), - beta_values_precommit), - framework::dataset::make("broadcast_bias", { true } )), - lhs_transpose_values), - act_values), - post_op_lists) - ) -{ - // Validate output only if validate() is successful - 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_END() // FusedPostOps TEST_SUITE_END() // ExportToCLImage TEST_SUITE_END() // FP16 @@ -1659,45 +1281,6 @@ FIXTURE_DATA_TEST_CASE(RunLarge3D, CLGEMMMatrixMultiplyReshaped3DMixedPrecisionF } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMMatrixMultiplyReshapedMixedPrecisionWithPostOpsFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(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), - v0_values_precommit), - h0_values_precommit), - framework::dataset::make("interleave_lhs", { false })), - framework::dataset::make("interleave_rhs", { false })), - framework::dataset::make("export_to_cl_image_rhs", { true, false })), - framework::dataset::make("DataType", DataType::F16)), - a_values_precommit), - beta_values_precommit), - framework::dataset::make("broadcast_bias", { true } )), - lhs_transpose_values), - act_values), - post_op_lists) - ) -{ - // Validate output - 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() // FusedPostOps - 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 53038c8177..dafc8dc5ec 100644 --- a/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp +++ b/tests/validation/CL/GEMMMatrixMultiplyReshapedOnlyRHS.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2019-2022 Arm Limited. + * Copyright (c) 2019-2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -23,7 +23,6 @@ */ #include "arm_compute/core/KernelDescriptors.h" #include "arm_compute/core/Types.h" -#include "arm_compute/core/experimental/PostOps.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" @@ -62,11 +61,6 @@ using CLGEMMMatrixMultiplyReshapedOnlyRHSFixture = GEMMMatrixMultiplyReshapedOnl template <typename T> using CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixture = GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRHS>; -// Fixture for CLGEMMMatrixMultiplyReshapedOnlyRHS with post ops -template <typename T> -using CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture = - GEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsValidationFixture<CLTensor, CLAccessor, T, CLGEMMReshapeRHSMatrix, CLGEMMMatrixMultiplyReshapedOnlyRHS>; - namespace { // *INDENT-OFF* @@ -164,106 +158,6 @@ const auto boundary_handling_cases = combine(combine(combine(combine(combine(com broadcast_bias_values), framework::dataset::make("Activation", ActivationLayerInfo())); -/** Post Ops */ -using PostOpArgBroadcast = CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture<float>::PostOpArgBroadcast; -experimental::PostOpList<PostOpArgBroadcast> post_ops_1() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(true, true, false), // If broadcast in dims 0, 1 and 2 - 0, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -experimental::PostOpList<PostOpArgBroadcast> post_ops_2() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(false, true, true), // If broadcast in dims 0, 1 and 2 - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -experimental::PostOpList<PostOpArgBroadcast> post_ops_3() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<PostOpArgBroadcast>>( - std::make_tuple(false, false, true), // If broadcast in dims 0, 1 and 2 - 1, - ConvertPolicy::SATURATE); - return post_ops; -} -// To test that the output of the main op is the first parameter in prelu post op -experimental::PostOpList<PostOpArgBroadcast> post_ops_4() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>( - std::make_tuple(false, false, true), // If true, broadcast in corresponding dim: 0, 1 or 2 - 0, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -// To test that the output of the main op is the second parameter in prelu post op i.e. it is the alpha_param -experimental::PostOpList<PostOpArgBroadcast> post_ops_5() -{ - experimental::PostOpList<PostOpArgBroadcast> post_ops{}; - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::LINEAR, 0.5F, 0.0F}); - post_ops.push_back_op<experimental::PostOpEltwisePRelu<PostOpArgBroadcast>>( - std::make_tuple(false, false, false), // If true, broadcast in corresponding dim: 0, 1 or 2 - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpAct<PostOpArgBroadcast>>(ActivationLayerInfo{ActivationLayerInfo::ActivationFunction::RELU, 2.1F, 1.3F}); - return post_ops; -} -/** Different Post Op Lists */ -const auto post_op_lists = framework::dataset::make("post_op_lists", { - post_ops_1(), - post_ops_2(), - post_ops_3(), - post_ops_4(), - post_ops_5() - } ); - - bool is_post_op_list_valid(unsigned int m, unsigned int n, unsigned int k, unsigned int batch, DataType data_type, const experimental::PostOpList<ITensorInfo*>& post_ops) -{ - const auto lhs_info = GEMMLHSMatrixInfo(4,4,1,false,true); - const auto rhs_info = GEMMRHSMatrixInfo(4,4,1,true,true,false); - - // Create TensorInfo for post op arguments - TensorInfo input0_info(TensorShape(k, m, batch), 1, data_type); - TensorInfo input1_info(TensorShape(n, k, batch), 1, data_type); - TensorInfo input2_info(TensorShape(n), 1, data_type); - TensorInfo output_info(TensorShape(n, m, batch), 1, data_type); - - const TensorInfo reshaped_input1_info = input1_info.clone()->set_tensor_shape(misc::shape_calculator::compute_rhs_reshaped_shape(input1_info, rhs_info)); - - GEMMKernelInfo gemm_info(m, n, k, 0 /**< Depth of the output tensor in case is reinterpreted as 3D */, - false /**< reinterpret the input as 3D */, - true /**< Flag used to broadcast the bias addition */, - false /**< wider accumm */, - false /**< has pad y */, - ActivationLayerInfo::ActivationFunction::IDENTITY, - 1 /**< Multiplication factor for the width of the 1xW transposed block */, - 1 /**< Multiplication factor for the height of the 4x4 interleaved block */, - lhs_info, - rhs_info, - 0 /**< Offset to be added to each element of the matrix A */, - 0 /**< Offset to be added to each element of the matrix B */, - post_ops); - return bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(&input0_info.clone()->set_is_resizable(true), - &reshaped_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, - lhs_info, - rhs_info, - gemm_info)); -} /** Configuration test */ bool validate_configuration(unsigned int m_value, unsigned int n_value, unsigned int k_value, unsigned int b_value, unsigned int m0_value, unsigned int n0_value, unsigned int k0_value, unsigned int h0_value, @@ -370,119 +264,6 @@ b_value, m0_value, n0_value, k0_value, broadcast_bias, input_as_3d, depth_output ARM_COMPUTE_EXPECT(status == expected_value, framework::LogLevel::ERRORS); } -TEST_SUITE(ValidateFusedPostOpsConfigs) -TEST_SUITE(Invalid) -TEST_CASE(UnsupportedPostOpSequence, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 17; - const unsigned int n = 1; - const unsigned int k = 13; - const unsigned int batch = 2; - TensorShape post_op_arg0_shape(n, m, batch); - TensorInfo post_op_arg_info(post_op_arg0_shape, 1, data_type); - auto post_op_arg1_info = post_op_arg_info.clone(); - - // Unsupported sequence of post ops - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( - &post_op_arg_info, - 1, - ConvertPolicy::SATURATE); - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( - post_op_arg1_info.get(), - 0, - ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(OutputWidened, framework::DatasetMode::ALL) -{ - // Invalid broadcast: post op tensors "widen" the output tensor - const auto data_type = DataType::F32; - const unsigned int m = 17; - const unsigned int n = 1; - const unsigned int k = 1; - const unsigned int batch = 1; - TensorShape post_op_arg_shape(n, m, batch + 4); // output's batch dimension is "widened", which is not allowed - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInXDimOnly, framework::DatasetMode::ALL) -{ - // Invalid broadcast: post op tensors broadcast in the first dimension (X) only - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, m, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == false, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Invalid -TEST_SUITE(Valid) -TEST_CASE(EmptyPostOpList, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - experimental::PostOpList<ITensorInfo*> post_ops{}; - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInYDimOnly, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(n, 1, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInBothXandYDims, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, 1, batch); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_CASE(BroadcastInAllDims, framework::DatasetMode::ALL) -{ - const auto data_type = DataType::F32; - const unsigned int m = 22; - const unsigned int n = 16; - const unsigned int k = 15; - const unsigned int batch = 3; - TensorShape post_op_arg_shape(1, 1, 1); - TensorInfo post_op_arg_info(post_op_arg_shape, 1, data_type); - experimental::PostOpList<ITensorInfo*> post_ops{}; - post_ops.push_back_op<experimental::PostOpEltwiseAdd<ITensorInfo*>>( &post_op_arg_info, 0, ConvertPolicy::SATURATE); - - ARM_COMPUTE_EXPECT(is_post_op_list_valid(m, n, k, batch, data_type, post_ops) == true, framework::LogLevel::ERRORS); -} -TEST_SUITE_END() // Valid -TEST_SUITE_END() // ValidateFusedPostOps TEST_SUITE(Float) TEST_SUITE(FP32) @@ -684,43 +465,6 @@ FIXTURE_DATA_TEST_CASE(RunNightly3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixtur } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunPrecommit, CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture<float>, framework::DatasetMode::ALL, - combine(combine(combine(combine(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("H0", {1})), - framework::dataset::make("interleave_rhs", { true })), - t_values_rhs), - framework::dataset::make("export_to_cl_image_rhs", {false, true})), - framework::dataset::make("DataType", DataType::F32)), - a_values), - beta_values), - framework::dataset::make("broadcast_bias", { false } )), - act_values), - post_op_lists) - ) -{ - // Validate output only if the target platform supports the OpenCL cl_khr_image2d_from_buffer extension - 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_END() // FusedPostOps - TEST_SUITE_END() // FP32 TEST_SUITE(FP16) @@ -849,42 +593,6 @@ FIXTURE_DATA_TEST_CASE(RunNightly3D, CLGEMMMatrixMultiplyReshapedOnlyRHS3DFixtur framework::ARM_COMPUTE_PRINT_INFO(); } } -TEST_SUITE(FusedPostOps) - -FIXTURE_DATA_TEST_CASE(RunPrecommit, CLGEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsFixture<half>, framework::DatasetMode::ALL, - combine(combine(combine(combine(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("H0", {1})), - framework::dataset::make("interleave_rhs", { true })), - t_values_rhs), - framework::dataset::make("export_to_cl_image_rhs", true)), - framework::dataset::make("DataType", DataType::F16)), - a_values), - beta_values), - framework::dataset::make("broadcast_bias", { false } )), - act_values), - post_op_lists) - ) -{ - // Validate output only if the target platform supports the OpenCL cl_khr_image2d_from_buffer extension - 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_END() // FusedPostOps TEST_SUITE_END() // FP16 diff --git a/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp index f27a1796c9..bae8cbf868 100644 --- a/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp +++ b/tests/validation/dynamic_fusion/gpu/cl/DirectConv2d.cpp @@ -58,7 +58,7 @@ TEST_SUITE(DYNAMIC_FUSION) * 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/post op tests | Not needed in fusion + * 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 @@ -70,9 +70,7 @@ 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()))) + 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); @@ -81,9 +79,7 @@ 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()))) + 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); diff --git a/tests/validation/fixtures/GEMMFixture.h b/tests/validation/fixtures/GEMMFixture.h index f1e0ee9150..afde3d8067 100644 --- a/tests/validation/fixtures/GEMMFixture.h +++ b/tests/validation/fixtures/GEMMFixture.h @@ -21,14 +21,12 @@ * 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" #include "arm_compute/core/Types.h" -#include "arm_compute/core/experimental/IPostOp.h" -#include "src/core/experimental/PostOpUtils.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/IAccessor.h" @@ -38,7 +36,6 @@ #include "tests/validation/reference/ActivationLayer.h" #include "tests/validation/reference/ElementwiseOperations.h" #include "tests/validation/reference/GEMM.h" -#include "tests/validation/reference/PostOps.h" #include <random> @@ -304,8 +301,7 @@ 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; @@ -426,8 +422,7 @@ 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; @@ -580,8 +575,7 @@ 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; @@ -734,8 +728,7 @@ 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; @@ -908,8 +901,7 @@ 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; @@ -960,262 +952,6 @@ protected: SimpleTensor<T> _reference{}; }; -/** (EXPERIMENTAL_POST_OPS)*/ -template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType, bool fp_mixed_precision = false> -class GEMMMatrixMultiplyReshapedWithPostOpsValidationFixture : public framework::Fixture -{ -public: - using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument -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 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, - const experimental::PostOpList<PostOpArgBroadcast> &post_ops) - { - GEMMLHSMatrixInfo lhs_info; - lhs_info.m0 = m0; - lhs_info.k0 = k0; - lhs_info.v0 = v0; - lhs_info.interleave = interleave_lhs; - lhs_info.transpose = lhs_transpose; - - GEMMRHSMatrixInfo rhs_info; - rhs_info.n0 = n0; - rhs_info.k0 = k0; - rhs_info.h0 = h0; - rhs_info.interleave = interleave_rhs; - rhs_info.transpose = !lhs_transpose; - 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); - auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops, - [ = ](auto broadcast) - { - return TensorShape - { - std::get<0>(broadcast) ? 1 : n, - std::get<1>(broadcast) ? 1 : m, - std::get<2>(broadcast) ? 1 : batch_size, - }; - }); - - _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); - if(validate_result) - { - _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); - } - } - -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, const experimental::PostOpList<TensorShape> &post_ops) - { - // 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); - - // Create post op tensors and populate post op with them - std::vector<TensorType> post_op_tensors_holder{}; - auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops, - [&post_op_tensors_holder, &data_type](auto shape) - { - auto t = create_tensor<TensorType>(shape, data_type, 1); - post_op_tensors_holder.push_back(std::move(t)); - return post_op_tensors_holder.back().info(); - }); - TensorType lhs_reshaped; - 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; - kernel_info.fp_mixed_precision = fp_mixed_precision; - kernel_info.post_ops = populated_post_ops; - - // The output tensor will be auto-initialized within the function - - // Create and configure function - ReshapeLHSOperatorType reshape_lhs; - ReshapeRHSOperatorType reshape_rhs; - GEMMOperatorType gemm; - - validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); - validate_result = validate_result || !rhs_info.export_to_cl_image; - if(!validate_result) - { - return nullptr; - } - - reshape_lhs.configure(lhs.info(), lhs_reshaped.info(), lhs_info); - reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); - gemm.configure(lhs_reshaped.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()); - for(const auto &tensor : post_op_tensors_holder) - { - ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); - } - - // We do not pad when using image as it needs to comply to strict pitch alignment restrictions - if(!rhs_info.export_to_cl_image) - { - add_padding_x({ &lhs, &rhs, &lhs_reshaped, &rhs_reshaped, &bias, &dst }); - for(auto &tensor : post_op_tensors_holder) - { - add_padding_x({ &tensor }); - } - } - - // Allocate tensors - lhs.allocator()->allocate(); - rhs.allocator()->allocate(); - lhs_reshaped.allocator()->allocate(); - rhs_reshaped.allocator()->allocate(); - bias.allocator()->allocate(); - dst.allocator()->allocate(); - for(auto &tensor : post_op_tensors_holder) - { - tensor.allocator()->allocate(); - } - - ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!lhs_reshaped.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!rhs_reshaped.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); - for(const auto &tensor : post_op_tensors_holder) - { - ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable()); - } - - // Fill tensors - fill(AccessorType(lhs), 0); - fill(AccessorType(rhs), 1); - fill(AccessorType(bias), 2); - for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) - { - fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i); - } - - // Compute GEMM - ITensorPack reshape_lhs_pack = { { ACL_SRC, &lhs }, { ACL_DST, &lhs_reshaped } }; - reshape_lhs.run(reshape_lhs_pack); - ITensorPack reshape_rhs_pack = { { ACL_SRC, &rhs }, { ACL_DST, &rhs_reshaped } }; - reshape_rhs.run(reshape_rhs_pack); - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs_reshaped }, - { ACL_SRC_1, &rhs_reshaped }, - { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); - for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) - { - gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i)); - } - 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, const experimental::PostOpList<TensorShape> &post_ops) - { - 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 }; - // Create post op tensors and populate post op with them - auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape) - { - return SimpleTensor<T> { 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 - int tensor_idx = 0; - fill(lhs, tensor_idx++); - fill(rhs, tensor_idx++); - fill(bias, tensor_idx++); - for(auto &op : populated_post_ops.get_list()) - { - for(auto tensor : op->arguments()) - { - fill(*tensor, tensor_idx++); - } - } - - 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)); - } - } - - SimpleTensor<T> out; - if(fp_mixed_precision) - { - out = reference::gemm_mixed_precision<T>(lhs, rhs, bias, alpha, beta); - } - else - { - out = reference::gemm<T>(lhs, rhs, bias, alpha, beta); - } - // Ignore activation info if post ops are used instead - if(populated_post_ops.size() > 0) - { - out = reference::post_ops<T>(out, populated_post_ops); - } - else - { - out = reference::activation_layer(out, act_info); - } - return out; - } - - bool validate_result = true; - TensorType _target{}; - SimpleTensor<T> _reference{}; -}; - template <typename TensorType, typename AccessorType, typename T, typename ReshapeLHSOperatorType, typename ReshapeRHSOperatorType, typename GEMMOperatorType, bool fp_mixed_precision = false> class GEMMMatrixMultiplyReshaped3DValidationFixture : public framework::Fixture { @@ -1344,8 +1080,7 @@ 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; @@ -1515,8 +1250,7 @@ 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; @@ -1560,242 +1294,6 @@ protected: SimpleTensor<T> _reference{}; }; -/** (EXPERIMENTAL_POST_OPS)*/ -template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> -class GEMMMatrixMultiplyReshapedOnlyRHSWithPostOpsValidationFixture : public framework::Fixture -{ -public: - using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument - 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, - const experimental::PostOpList<PostOpArgBroadcast> &post_ops) - { - 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; - 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); - auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops, - [ = ](auto broadcast) - { - return TensorShape - { - std::get<0>(broadcast) ? 1 : n, - std::get<1>(broadcast) ? 1 : m, - std::get<2>(broadcast) ? 1 : batch_size, - }; - }); - - _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); - if(validate_result) - { - _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); - } - } - -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, const experimental::PostOpList<TensorShape> &post_ops) - { - // 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; - // Create post op tensors and populate post op with them - std::vector<TensorType> post_op_tensors_holder{}; - auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops, - [&post_op_tensors_holder, &data_type](auto shape) - { - auto t = create_tensor<TensorType>(shape, data_type, 1); - post_op_tensors_holder.push_back(std::move(t)); - return post_op_tensors_holder.back().info(); - }); - - 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; - kernel_info.post_ops = populated_post_ops; - - // The output tensor will be auto-initialized within the function - - // Create and configure function - ReshapeRHSOperatorType reshape_rhs; - GEMMOperatorType gemm; - - validate_result = bool(reshape_rhs.validate(rhs.info(), rhs_reshaped.info(), rhs_info)); - validate_result = validate_result || !rhs_info.export_to_cl_image; - if(!validate_result) - { - return nullptr; - } - - reshape_rhs.configure(rhs.info(), rhs_reshaped.info(), rhs_info); - 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()); - for(const auto &tensor : post_op_tensors_holder) - { - ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); - } - - // We do not pad when using image as it needs to comply to strict pitch alignment restrictions - if(!rhs_info.export_to_cl_image) - { - add_padding_x({ &lhs, &rhs, &rhs_reshaped, &bias, &dst }); - for(auto &tensor : post_op_tensors_holder) - { - add_padding_x({ &tensor }); - } - } - - // Allocate tensors - lhs.allocator()->allocate(); - rhs.allocator()->allocate(); - rhs_reshaped.allocator()->allocate(); - bias.allocator()->allocate(); - dst.allocator()->allocate(); - for(auto &tensor : post_op_tensors_holder) - { - tensor.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()); - for(const auto &tensor : post_op_tensors_holder) - { - ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable()); - } - - // Fill tensors - fill(AccessorType(lhs), 0); - fill(AccessorType(rhs), 1); - fill(AccessorType(bias), 2); - for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) - { - fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i); - } - - // 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 } - }); - for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) - { - gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i)); - } - 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, const experimental::PostOpList<TensorShape> &post_ops) - { - 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 }; - // Create post op tensors and populate post op with them - auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape) - { - return SimpleTensor<T> { 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 - int tensor_idx = 0; - fill(lhs, tensor_idx++); - fill(rhs, tensor_idx++); - fill(bias, tensor_idx++); - for(auto &op : populated_post_ops.get_list()) - { - for(auto tensor : op->arguments()) - { - fill(*tensor, tensor_idx++); - } - } - - 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)); - } - } - - SimpleTensor<T> out; - out = reference::gemm<T>(lhs, rhs, bias, alpha, beta); - // Ignore activation info if post ops are used instead - if(populated_post_ops.size() > 0) - { - out = reference::post_ops<T>(out, populated_post_ops); - } - else - { - out = reference::activation_layer(out, act_info); - } - return out; - } - - bool validate_result = true; - TensorType _target{}; - SimpleTensor<T> _reference{}; -}; - template <typename TensorType, typename AccessorType, typename T, typename ReshapeRHSOperatorType, typename GEMMOperatorType> class GEMMMatrixMultiplyReshapedOnlyRHS3DValidationFixture : public framework::Fixture { @@ -1921,8 +1419,7 @@ 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; @@ -2057,8 +1554,7 @@ 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; @@ -2102,212 +1598,6 @@ protected: }; template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> -class GEMMMatrixMultiplyNativeWithPostOpsValidationFixture : public framework::Fixture -{ -public: - using PostOpArgBroadcast = std::tuple<bool, bool, bool>; // Instruct fixture if we need broadcasting in dimension 0, 1, 2 of each PostOp argument -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, DataType data_type, float alpha, float beta, bool broadcast_bias, - const ActivationLayerInfo &act_info, const experimental::PostOpList<PostOpArgBroadcast> &post_ops) - { - GEMMLHSMatrixInfo lhs_info; - lhs_info.m0 = m0; - lhs_info.k0 = k0; - - GEMMRHSMatrixInfo rhs_info; - rhs_info.n0 = n0; - rhs_info.k0 = k0; - - // 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); - const auto post_ops_with_shapes = experimental::transform_post_op_list_arguments<PostOpArgBroadcast, TensorShape>(post_ops, - [ = ](auto broadcast) - { - return TensorShape - { - std::get<0>(broadcast) ? 1 : n, - std::get<1>(broadcast) ? 1 : m, - std::get<2>(broadcast) ? 1 : batch_size, - }; - }); - - _target = compute_target(lhs_shape, rhs_shape, bias_shape, lhs_info, rhs_info, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); - _reference = compute_reference(lhs_shape, rhs_shape, data_type, alpha, beta, broadcast_bias, act_info, post_ops_with_shapes); - } - -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, const experimental::PostOpList<TensorShape> &post_ops) - { - // 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 dst; - // Create post op tensors and populate post op with them - std::vector<TensorType> post_op_tensors_holder{}; - auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, ITensorInfo *>(post_ops, - [&post_op_tensors_holder, &data_type](auto shape) - { - auto t = create_tensor<TensorType>(shape, data_type, 1); - post_op_tensors_holder.push_back(std::move(t)); - return post_op_tensors_holder.back().info(); - }); - - 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; - kernel_info.post_ops = populated_post_ops; - - // Create and configure function - GEMMOperatorType gemm; - gemm.configure(lhs.info(), rhs.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()); - for(const auto &tensor : post_op_tensors_holder) - { - ARM_COMPUTE_ASSERT(tensor.info()->is_resizable()); - } - - add_padding_x({ &lhs, &rhs, &bias, &dst }); - for(auto &tensor : post_op_tensors_holder) - { - add_padding_x({ &tensor }); - } - - // Allocate tensors - lhs.allocator()->allocate(); - rhs.allocator()->allocate(); - bias.allocator()->allocate(); - dst.allocator()->allocate(); - for(auto &tensor : post_op_tensors_holder) - { - tensor.allocator()->allocate(); - } - - ARM_COMPUTE_ASSERT(!lhs.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!rhs.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!bias.info()->is_resizable()); - ARM_COMPUTE_ASSERT(!dst.info()->is_resizable()); - for(const auto &tensor : post_op_tensors_holder) - { - ARM_COMPUTE_ASSERT(!tensor.info()->is_resizable()); - } - - // Fill tensors - fill(AccessorType(lhs), 0); - fill(AccessorType(rhs), 1); - fill(AccessorType(bias), 2); - for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) - { - fill(AccessorType(post_op_tensors_holder.at(i)), 3 + i); - } - - // Compute GEMM - ITensorPack gemm_pack({ { ACL_SRC_0, &lhs }, - { ACL_SRC_1, &rhs }, - { ACL_SRC_2, &bias }, - { ACL_DST, &dst } - }); - for(size_t i = 0; i < post_op_tensors_holder.size(); ++i) - { - gemm_pack.add_tensor(experimental::get_post_op_arg_type(i), &post_op_tensors_holder.at(i)); - } - 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, const experimental::PostOpList<TensorShape> &post_ops) - { - 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 }; - // Create post op tensors and populate post op with them - auto populated_post_ops = experimental::transform_post_op_list_arguments<TensorShape, SimpleTensor<T>>(post_ops, [&data_type](auto shape) - { - return SimpleTensor<T> { 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 - int tensor_idx = 0; - fill(lhs, tensor_idx++); - fill(rhs, tensor_idx++); - fill(bias, tensor_idx++); - for(auto &op : populated_post_ops.get_list()) - { - for(auto tensor : op->arguments()) - { - fill(*tensor, tensor_idx++); - } - } - - 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)); - } - } - - SimpleTensor<T> out; - out = reference::gemm<T>(lhs, rhs, bias, alpha, beta); - // Ignore activation info if post ops are used instead - if(populated_post_ops.size() > 0) - { - out = reference::post_ops<T>(out, populated_post_ops); - } - else - { - out = reference::activation_layer(out, act_info); - } - return out; - } - - TensorType _target{}; - SimpleTensor<T> _reference{}; -}; - -template <typename TensorType, typename AccessorType, typename T, typename GEMMOperatorType> class GEMMMatrixMultiplyNative3DValidationFixture : public framework::Fixture { public: @@ -2398,8 +1688,7 @@ 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; @@ -2557,8 +1846,7 @@ 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; @@ -2608,4 +1896,4 @@ protected: } // 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/reference/PostOps.cpp b/tests/validation/reference/PostOps.cpp deleted file mode 100644 index ecfed4c48f..0000000000 --- a/tests/validation/reference/PostOps.cpp +++ /dev/null @@ -1,93 +0,0 @@ -/* - * 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 "PostOps.h" - -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/experimental/PostOps.h" -#include "support/Cast.h" -#include "tests/validation/reference/ActivationLayer.h" -#include "tests/validation/reference/ElementwiseOperations.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -namespace reference -{ -template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> -SimpleTensor<T> post_ops(const SimpleTensor<T> &a, experimental::PostOpList<SimpleTensor<T>> post_ops) -{ - // Create reference - SimpleTensor<T> dst{ a }; - - for(auto &post_op : post_ops.get_list()) - { - switch(post_op->type()) - { - case experimental::PostOpType::Activation: - { - const auto _post_op = utils::cast::polymorphic_downcast<const experimental::PostOpAct<SimpleTensor<T>> *>(post_op.get()); - dst = reference::activation_layer(dst, _post_op->_act_info); - break; - } - case experimental::PostOpType::Eltwise_Add: - { - const auto _post_op = utils::cast::polymorphic_downcast<const experimental::PostOpEltwiseAdd<SimpleTensor<T>> *>(post_op.get()); - dst = reference::arithmetic_operation(ArithmeticOperation::ADD, dst, _post_op->_addend, dst, _post_op->_policy); - break; - } - case experimental::PostOpType::Eltwise_PRelu: - { - const auto _post_op = utils::cast::polymorphic_downcast<const experimental::PostOpEltwisePRelu<SimpleTensor<T>> *>(post_op.get()); - - // If previous main operation output is the the first pRelu argument, then pass it as src1 parameter of the arithmetic operation - if(_post_op->_prev_dst_pos == 0) - { - dst = reference::arithmetic_operation(ArithmeticOperation::PRELU, dst, _post_op->_alpha_param, dst, _post_op->_policy); - } - // If previous main operation output is the the second pRelu argument, then pass it as src2 parameter of the arithmetic operation - else if(_post_op->_prev_dst_pos == 1) - { - dst = reference::arithmetic_operation(ArithmeticOperation::PRELU, _post_op->_alpha_param, dst, dst, _post_op->_policy); - } - break; - } - default: - { - ARM_COMPUTE_ERROR("Unsupported PostOpType"); - } - } - } - return dst; -} - -template SimpleTensor<float> post_ops(const SimpleTensor<float> &a, experimental::PostOpList<SimpleTensor<float>> post_ops); -template SimpleTensor<half> post_ops(const SimpleTensor<half> &a, experimental::PostOpList<SimpleTensor<half>> post_ops); -} // namespace reference -} // namespace validation -} // namespace test -} // namespace arm_compute
\ No newline at end of file diff --git a/tests/validation/reference/PostOps.h b/tests/validation/reference/PostOps.h deleted file mode 100644 index 5fe0fe71f5..0000000000 --- a/tests/validation/reference/PostOps.h +++ /dev/null @@ -1,47 +0,0 @@ -/* - * 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_POSTOPS_H -#define ARM_COMPUTE_TEST_POSTOPS_H - -#include "arm_compute/core/experimental/IPostOp.h" -#include "tests/SimpleTensor.h" -#include "tests/validation/Helpers.h" - -namespace arm_compute -{ -namespace test -{ -namespace validation -{ -namespace reference -{ -/** (EXPERIMENTAL_POST_OPS) */ -template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0> -SimpleTensor<T> post_ops(const SimpleTensor<T> &a, experimental::PostOpList<SimpleTensor<T>> post_ops); - -} // namespace reference -} // namespace validation -} // namespace test -} // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_POSTOPS_H */ |