From 2d7e683e79c8ad328d4930c1f82a46827313faf4 Mon Sep 17 00:00:00 2001 From: George Wort Date: Fri, 22 Feb 2019 16:37:41 +0000 Subject: COMPMID-1694: Fuse offset contribution with the output stage when we use NEGEMMLowpMatrixMultiplyCore Change-Id: Ic1a681e4cc03e1eba3bf8485d9cdb17b3e926047 Signed-off-by: giuros01 Reviewed-on: https://review.mlplatform.org/c/561 Reviewed-by: Gian Marco Iodice Tested-by: Arm Jenkins --- tests/validation/CL/GEMMLowp.cpp | 16 +++ tests/validation/NEON/GEMMLowp.cpp | 15 +++ tests/validation/fixtures/GEMMLowpFixture.h | 194 +++++++++++++++++++--------- 3 files changed, 167 insertions(+), 58 deletions(-) (limited to 'tests/validation') diff --git a/tests/validation/CL/GEMMLowp.cpp b/tests/validation/CL/GEMMLowp.cpp index 08641dbaa3..efefbd645b 100644 --- a/tests/validation/CL/GEMMLowp.cpp +++ b/tests/validation/CL/GEMMLowp.cpp @@ -28,6 +28,7 @@ #include "arm_compute/runtime/CL/functions/CLGEMMLowpOutputStage.h" #include "tests/CL/CLAccessor.h" #include "tests/PaddingCalculator.h" +#include "tests/datasets/GEMMLowpFusedOffsetOutputDataset.h" #include "tests/datasets/LargeGEMMLowpDataset.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/datasets/SmallGEMMLowpDataset.h" @@ -83,6 +84,21 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFixture, framework: validate(CLAccessor(_target), _reference); } +using CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture; +TEST_SUITE(FusedOffsetOutput) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpFusedOffsetOutputDataset()) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpFusedOffsetOutputDataset()) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() // FusedOffsetOutput + TEST_SUITE(Output3D) using CLGEMMLowpMatrixMultiplyCoreOutput3DFixture = GEMMLowpMatrixMultiplyCoreValidationFixture; FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMLowpMatrixMultiplyCoreOutput3DFixture, framework::DatasetMode::PRECOMMIT, datasets::SmallGEMMLowpOutput3DDataset()) diff --git a/tests/validation/NEON/GEMMLowp.cpp b/tests/validation/NEON/GEMMLowp.cpp index 57067f140f..f0460b4a23 100644 --- a/tests/validation/NEON/GEMMLowp.cpp +++ b/tests/validation/NEON/GEMMLowp.cpp @@ -30,6 +30,7 @@ #include "tests/NEON/Accessor.h" #include "tests/NEON/Helper.h" #include "tests/PaddingCalculator.h" +#include "tests/datasets/GEMMLowpFusedOffsetOutputDataset.h" #include "tests/datasets/LargeGEMMLowpDataset.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/datasets/SmallGEMMLowpDataset.h" @@ -144,6 +145,20 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFixture, framework: validate(Accessor(_target), _reference); } +using NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture = GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture; +TEST_SUITE(FusedOffsetOutput) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::ALL, datasets::SmallGEMMLowpFusedOffsetOutputDataset()) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMLowpMatrixMultiplyCoreFusedOffsetOutputFixture, framework::DatasetMode::NIGHTLY, datasets::LargeGEMMLowpFusedOffsetOutputDataset()) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() // FusedOffsetOutput TEST_SUITE_END() // MatrixMultiplyCore TEST_SUITE(OutputStage) diff --git a/tests/validation/fixtures/GEMMLowpFixture.h b/tests/validation/fixtures/GEMMLowpFixture.h index 836f8eddfe..90a4b5cf40 100644 --- a/tests/validation/fixtures/GEMMLowpFixture.h +++ b/tests/validation/fixtures/GEMMLowpFixture.h @@ -42,86 +42,164 @@ namespace test { namespace validation { -template -class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture +namespace { -public: - template - void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, int32_t a_offset, int32_t b_offset) +template +void fill(U &&tensor, int i) +{ + // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path + std::uniform_int_distribution<> distribution(1, 254); + library->fill(tensor, distribution, i); +} + +template +TensorType compute_gemmlowp_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, + GEMMLowpOutputStageInfo output_stage = GEMMLowpOutputStageInfo()) +{ + // Create tensors + TensorType a = create_tensor(shape_a, DataType::QASYMM8, 1); + TensorType b = create_tensor(shape_b, DataType::QASYMM8, 1); + TensorType output = create_tensor(shape_output, output_stage.type == GEMMLowpOutputStageType::NONE ? DataType::S32 : DataType::QASYMM8, 1); + + a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset)); + b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset)); + + TensorType bias; + if(is_fused) { - _target = compute_target(shape_a, shape_b, shape_c, a_offset, b_offset); - _reference = compute_reference(shape_a, shape_b, shape_c, a_offset, b_offset); + TensorShape bias_shape(shape_b[0]); + bias = create_tensor(bias_shape, DataType::S32, 1); } -protected: - template - void fill(U &&tensor, int i) + // Create and configure function + // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output + FunctionType gemmlowp; + // TODO (COMPMID-1672) - Extending the test to validate add bias in offset contribution + gemmlowp.configure(&a, &b, is_fused ? &bias : nullptr, &output, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_output[2] : 0), reinterpret_input_as_3d, false, output_stage)); + + ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(output.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + a.allocator()->allocate(); + b.allocator()->allocate(); + output.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!output.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(a), 0); + fill(AccessorType(b), 1); + + if(is_fused) { - // Between 1 and 254 in order to avoid having -128 and 128 for the DOT product path - std::uniform_int_distribution<> distribution(1, 254); - library->fill(tensor, distribution, i); + ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); + bias.allocator()->allocate(); + ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); + fill(AccessorType(bias), 2); } - TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, int32_t a_offset, int32_t b_offset) + // Compute GEMM function + gemmlowp.run(); + return output; +} + +template +SimpleTensor compute_gemmlowp_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) +{ + TensorShape shape_a_to_use = shape_a; + if(reinterpret_input_as_3d) { - // Create tensors - TensorType a = create_tensor(shape_a, DataType::QASYMM8, 1); - TensorType b = create_tensor(shape_b, DataType::QASYMM8, 1); - TensorType c = create_tensor(shape_c, DataType::S32, 1); + // Collapse the second and third dimension if the input is 3D + shape_a_to_use.collapse(2U, 1U); + } - a.info()->set_quantization_info(QuantizationInfo(1.0f / 255, a_offset)); - b.info()->set_quantization_info(QuantizationInfo(1.0f / 255, b_offset)); + // Create reference + SimpleTensor a{ shape_a_to_use, DataType::QASYMM8, 1 }; + SimpleTensor b{ shape_b, DataType::QASYMM8, 1 }; - // Create and configure function - // The GEMMinfo includes the values of the depth in case of reinterpreted 3d input/output - FunctionType gemmlowp; - // TODO (COMPMID-1672) - Extending the test to validate add bias in offset contribution - gemmlowp.configure(&a, &b, nullptr, &c, GEMMInfo(false, false, false, (reinterpret_output_as_3d ? shape_c[2] : 0), reinterpret_input_as_3d)); + // Fill reference + fill(a, 0); + fill(b, 1); - ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(c.info()->is_resizable(), framework::LogLevel::ERRORS); + return reference::gemmlowp_matrix_multiply_core(a, b, shape_output, a_offset, b_offset); +} +} - // Allocate tensors - a.allocator()->allocate(); - b.allocator()->allocate(); - c.allocator()->allocate(); +template +class GEMMLowpMatrixMultiplyCoreValidationFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset) + { + _target = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset); + _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset); + } - ARM_COMPUTE_EXPECT(!a.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(!b.info()->is_resizable(), framework::LogLevel::ERRORS); - ARM_COMPUTE_EXPECT(!c.info()->is_resizable(), framework::LogLevel::ERRORS); +protected: + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) + { + return compute_gemmlowp_target(shape_a, shape_b, shape_output, a_offset, b_offset); + } - // Fill tensors - fill(AccessorType(a), 0); - fill(AccessorType(b), 1); + SimpleTensor compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset) + { + return compute_gemmlowp_reference(shape_a, shape_b, shape_output, a_offset, b_offset); + } - // Compute GEMM function - gemmlowp.run(); - return c; + TensorType _target{}; + SimpleTensor _reference{}; +}; + +template +class GEMMLowpMatrixMultiplyCoreFusedOffsetOutputValidationFixture : public framework::Fixture +{ +public: + template + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) + { + ARM_COMPUTE_EXPECT(output_stage.type != GEMMLowpOutputStageType::NONE, framework::LogLevel::ERRORS); + _target = compute_target(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage); + _reference = compute_reference(shape_a, shape_b, shape_output, a_offset, b_offset, output_stage); } - SimpleTensor compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_c, int32_t a_offset, int32_t b_offset) +protected: + TensorType compute_target(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, GEMMLowpOutputStageInfo output_stage) { - TensorShape shape_a_to_use = shape_a; - if(reinterpret_input_as_3d) - { - // Collapse the second and third dimension if the input is 3D - shape_a_to_use.collapse(2U, 1U); - } + return compute_gemmlowp_target(shape_a, shape_b, shape_output, a_offset, b_offset, + output_stage); + } - // Create reference - SimpleTensor a{ shape_a_to_use, DataType::QASYMM8, 1 }; - SimpleTensor b{ shape_b, DataType::QASYMM8, 1 }; + SimpleTensor compute_reference(const TensorShape &shape_a, const TensorShape &shape_b, const TensorShape &shape_output, int32_t a_offset, int32_t b_offset, + GEMMLowpOutputStageInfo output_stage) + { + SimpleTensor output = compute_gemmlowp_reference(shape_a, shape_b, shape_output, a_offset, b_offset); - // Fill reference - fill(a, 0); - fill(b, 1); + TensorShape bias_shape(shape_b[0]); + SimpleTensor bias{ bias_shape, DataType::S32, 1 }; + fill(bias, 2); - return reference::gemmlowp_matrix_multiply_core(a, b, shape_c, a_offset, b_offset); + switch(output_stage.type) + { + case GEMMLowpOutputStageType::QUANTIZE_DOWN: + return reference::gemmlowp_quantize_down_int32_to_uint8_scale(output, bias, + output_stage.gemmlowp_offset, output_stage.gemmlowp_multiplier, output_stage.gemmlowp_shift, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); + break; + case GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT: + return reference::gemmlowp_quantize_down_int32_to_uint8_scale_by_fixedpoint(output, bias, + output_stage.gemmlowp_multiplier, output_stage.gemmlowp_shift, output_stage.gemmlowp_offset, output_stage.gemmlowp_min_bound, output_stage.gemmlowp_max_bound); + break; + default: + ARM_COMPUTE_ERROR("Not Supported!"); + } } - TensorType _target{}; - SimpleTensor _reference{}; + TensorType _target{}; + SimpleTensor _reference{}; }; template @@ -536,4 +614,4 @@ protected: } // namespace validation } // namespace test } // namespace arm_compute -#endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */ \ No newline at end of file +#endif /* ARM_COMPUTE_TEST_GEMMLOWP_FIXTURE */ -- cgit v1.2.1