From b3be45759bdd0749ae3a16fe470820f0d9830ea9 Mon Sep 17 00:00:00 2001 From: Sang-Hoon Park Date: Tue, 18 May 2021 10:46:00 +0100 Subject: Implement memory injection in CpuDirectGemmConv2d The following operators are now stateless by implementing memory injection. - CpuDirectGemmConv2d - CpuGemmAssemblyDispatch A test case is added to test if CpuDirectGemmConv2d can run on different group of tensors with a single configure. Resolves: COMPMID-4506 Change-Id: I48f44ed41236ca7e18da2de07bdbacc9007a3c5e Signed-off-by: Sang-Hoon Park Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5718 Comments-Addressed: Arm Jenkins Tested-by: Arm Jenkins Reviewed-by: Pablo Marquez Tello --- tests/validation/NEON/ConvolutionLayer.cpp | 239 ++++++++++++++++++++--------- 1 file changed, 163 insertions(+), 76 deletions(-) (limited to 'tests/validation/NEON/ConvolutionLayer.cpp') diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index b435744cdc..f38f9034a4 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -28,6 +28,8 @@ #include "arm_compute/runtime/NEON/functions/NEWinogradConvolutionLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" +#include "src/core/helpers/MemoryHelpers.h" +#include "src/runtime/cpu/operators/CpuGemmDirectConv2d.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/LargeConvolutionLayerDataset.h" @@ -169,13 +171,13 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradConvolutionLayerFixture, frame validate(Accessor(_target), _reference, abs_tolerance_f32); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEWinogradConvolutionLayerMixedDataLayoutFixture, framework::DatasetMode::PRECOMMIT, - combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(8U, 8U, 32U)), - framework::dataset::make("Weight", TensorShape(1U, 3U, 32U, 1U))), - framework::dataset::make("Bias", TensorShape(1U))), - framework::dataset::make("Output", TensorShape(8U, 6U, 1U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1U, 1U))), + combine(combine(combine(combine(combine(combine(combine(combine( + framework::dataset::make("Input", TensorShape(8U, 8U, 32U)), + framework::dataset::make("Weight", TensorShape(1U, 3U, 32U, 1U))), + framework::dataset::make("Bias", TensorShape(1U))), + framework::dataset::make("Output", TensorShape(8U, 6U, 1U))), + framework::dataset::make("PadStrideInfo", PadStrideInfo(1, 1, 0, 0))), + framework::dataset::make("Dilation", Size2D(1U, 1U))), framework::dataset::make("DataType", { DataType::F32 })), ActivationFunctionsDataset), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC }))) @@ -408,9 +410,7 @@ TEST_SUITE(Float) #if defined(__ARM_FEATURE_BF16_VECTOR_ARITHMETIC) || defined(ARM_COMPUTE_FORCE_BF16) TEST_SUITE(BFLOAT16) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::BFLOAT16)), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::BFLOAT16)), framework::dataset::make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { // Validate output @@ -422,10 +422,7 @@ TEST_SUITE_END() // BFLOAT16 #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture, 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 })), - ActivationFunctionsDataset)) + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F16)), framework::dataset::make("DataLayout", { DataLayout::NCHW })), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f16, tolerance_num, abs_tolerance_f16); @@ -435,26 +432,24 @@ TEST_SUITE_END() // FP16 TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture, 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 })), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEGEMMConvolutionLayerMixedDataLayoutFixture, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), - framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), - framework::dataset::make("Bias", TensorShape(2U))), - framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1, 1))), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::F32)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - ActivationFunctionsDataset)) + combine(combine(combine(combine(combine(combine(combine(combine(combine( + framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), + framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), + framework::dataset::make("Bias", TensorShape(2U))), + framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), + framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), + framework::dataset::make("Dilation", Size2D(1, 1))), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::F32)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); @@ -479,28 +474,25 @@ const auto QuantizedActivationFunctionsDataset = framework::dataset::make("Activ TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerQuantizedFixture, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - QuantizedActivationFunctionsDataset)) + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), QuantizedActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEGEMMConvolutionLayerQuantizedFixture, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), - framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), - framework::dataset::make("Bias", TensorShape(2U))), - framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1, 1))), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - QuantizedActivationFunctionsDataset)) + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), + framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), + framework::dataset::make("Bias", TensorShape(2U))), + framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), + framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), + framework::dataset::make("Dilation", Size2D(1, 1))), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + QuantizedActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -509,28 +501,25 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerQuantizedFixture, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.01f, -10) })), - QuantizedActivationFunctionsDataset)) + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.01f, -10) })), QuantizedActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); } FIXTURE_DATA_TEST_CASE(RunMixedDataLayout, NEGEMMConvolutionLayerQuantizedFixture, framework::DatasetMode::ALL, - combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( - framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), - framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), - framework::dataset::make("Bias", TensorShape(2U))), - framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), - framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), - framework::dataset::make("Dilation", Size2D(1, 1))), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - QuantizedActivationFunctionsDataset)) + combine(combine(combine(combine(combine(combine(combine(combine(combine(combine( + framework::dataset::make("Input", TensorShape(23U, 27U, 5U)), + framework::dataset::make("Weights", TensorShape(3U, 3U, 5U, 2U))), + framework::dataset::make("Bias", TensorShape(2U))), + framework::dataset::make("Output", TensorShape(11U, 25U, 2U))), + framework::dataset::make("PadStrideInfo", PadStrideInfo(2, 1, 0, 0))), + framework::dataset::make("Dilation", Size2D(1, 1))), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), + framework::dataset::make("DataLayout", { DataLayout::NCHW, DataLayout::NHWC })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), + QuantizedActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -571,13 +560,117 @@ TEST_SUITE(DirectGEMMConv2d) template using NEDirectGEMMConv2dLayerFixture = ConvolutionValidationFixture; +/** Test case to test if an operator configured once can be run for different group of tensors */ +TEST_CASE(MemoryInjection, framework::DatasetMode::ALL) +{ + auto conv = std::make_unique(); + const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(1U, 7U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto conv_info = Conv2dInfo{}; + + conv->configure(&src_info, &weight_info, &bias_info, &dst_info, conv_info); + + auto run_conv = [&]() -> Tensor + { + auto pack = ITensorPack{}; + auto mg = MemoryGroup{}; + auto ws = manage_workspace(conv->workspace(), mg, pack); + + // tensors are newly created every call of this lambda function + auto src = create_tensor(src_info); + auto weight = create_tensor(weight_info); + auto bias = create_tensor(bias_info); + auto dst = create_tensor(dst_info); + + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + pack.add_const_tensor(TensorType::ACL_SRC_0, &src); + pack.add_const_tensor(TensorType::ACL_SRC_1, &weight); + pack.add_const_tensor(TensorType::ACL_SRC_2, &bias); + pack.add_tensor(TensorType::ACL_DST, &dst); + + // value aren't too important if the same values are used in each execution. + std::vector static_number_to_be_filled{}; + for(size_t i = 0; i < weight_info.tensor_shape().total_size(); i++) + { + static_number_to_be_filled.emplace_back(i); + } + + library->fill_static_values(Accessor(src), static_number_to_be_filled); + library->fill_static_values(Accessor(weight), static_number_to_be_filled); + library->fill_static_values(Accessor(bias), static_number_to_be_filled); + + // This operator is configured once and captured by this lambda. + conv->run(pack); + return dst; + }; + + auto result_0 = run_conv(); + auto result_1 = run_conv(); + + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); i++) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + +TEST_CASE(MultipleExecutionWithConfigure, framework::DatasetMode::ALL) +{ + auto conv = std::make_unique(); + const auto src_info = TensorInfo(TensorShape(1U, 5U, 2U), 1, DataType::F32, DataLayout::NHWC); + const auto weight_info = TensorInfo(TensorShape(1U, 3U, 2U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto bias_info = TensorInfo(TensorShape(3U), 1, DataType::F32, DataLayout::NHWC); + auto dst_info = TensorInfo(TensorShape(1U, 7U, 3U), 1, DataType::F32, DataLayout::NHWC); + const auto conv_info = Conv2dInfo{}; + + auto run_conv = [&]() + { + auto src = create_tensor(src_info); + auto weight = create_tensor(weight_info); + auto bias = create_tensor(bias_info); + auto dst = create_tensor(dst_info); + + conv->configure(&src, &weight, &bias, &dst, conv_info); + + src.allocator()->allocate(); + weight.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + + // value aren't too important if the same values are used in each execution. + std::vector static_number_to_be_filled{}; + for(size_t i = 0; i < weight_info.tensor_shape().total_size(); i++) + { + static_number_to_be_filled.emplace_back(i); + } + + library->fill_static_values(Accessor(src), static_number_to_be_filled); + library->fill_static_values(Accessor(weight), static_number_to_be_filled); + library->fill_static_values(Accessor(bias), static_number_to_be_filled); + + conv->run(); + + return dst; + }; + + auto result_0 = run_conv(); + auto result_1 = run_conv(); + + for(size_t i = 0; i < result_0.info()->tensor_shape().total_size(); i++) + { + ARM_COMPUTE_EXPECT(((float *)result_0.buffer())[i] == ((float *)result_1.buffer())[i], framework::LogLevel::ERRORS); + } +} + TEST_SUITE(Float) TEST_SUITE(FP32) FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectGEMMConv2dLayerFixture, 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::NHWC })), - ActivationFunctionsDataset)) + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::F32)), framework::dataset::make("DataLayout", { DataLayout::NHWC })), ActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, rel_tolerance_f32, 0.f, float(abs_tolerance_f32)); @@ -601,11 +694,8 @@ const auto QuantizedActivationFunctionsDataset = framework::dataset::make("Activ TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectGEMMConv2dLayerQuantizedFixture, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), - QuantizedActivationFunctionsDataset)) + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) })), QuantizedActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); @@ -614,11 +704,8 @@ TEST_SUITE_END() // QASYMM8 TEST_SUITE(QASYMM8_SIGNED) FIXTURE_DATA_TEST_CASE(RunSmall, NEDirectGEMMConv2dLayerQuantizedFixture, framework::DatasetMode::ALL, combine(combine(combine(combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), - framework::dataset::make("DataLayout", { DataLayout::NHWC })), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.01f, -10) })), - QuantizedActivationFunctionsDataset)) + framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8_SIGNED)), framework::dataset::make("DataLayout", { DataLayout::NHWC })), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.01f, -10) })), QuantizedActivationFunctionsDataset)) { // Validate output validate(Accessor(_target), _reference, tolerance_qasymm8); -- cgit v1.2.1