From 7359a87040c7a901619de21701f540dd5a9a960c Mon Sep 17 00:00:00 2001 From: Jakub Sujak Date: Thu, 5 Jan 2023 14:24:13 +0000 Subject: Add Multiplication operator (FP only) to Dynamic Fusion Interface Note: we use a separate test fixture for Multiplication op instead of reusing ElementwiseBinaryFixture to avoid exposing the internal enum ElementwiseOp to the public utils/TypePrinters.h as required by the data test case macros to print the test data. We also do not consider modifying the enum ArithmeticOp in the standard interface to include MUL without an implementation. Future work should consider refactoring this test fixture into the ElementwiseBinaryFixture to reduce the total number of fixtures/code duplication. Resolves: COMPMID-5779 Change-Id: I84207658ce0407095b028fca0ab7bfa2950255ec Signed-off-by: Jakub Sujak Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/9013 Reviewed-by: Viet-Hoa Do Reviewed-by: SiCong Li Reviewed-by: Gunes Bayir Tested-by: Arm Jenkins Comments-Addressed: Arm Jenkins Benchmark: Arm Jenkins --- .../gpu/cl/ElementwiseBinaryFixture.h | 41 ++-- .../fixtures/dynamic_fusion/operators/MulFixture.h | 235 +++++++++++++++++++++ 2 files changed, 255 insertions(+), 21 deletions(-) create mode 100644 tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h (limited to 'tests/validation/fixtures') diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h index faed610874..b0680c0e4a 100644 --- a/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h +++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/ElementwiseBinaryFixture.h @@ -31,12 +31,9 @@ #include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" #include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" -#include "tests/CL/CLAccessor.h" #include "tests/framework/Fixture.h" #include "tests/framework/Macros.h" -#include "tests/validation/Validation.h" #include "tests/validation/reference/ElementwiseOperations.h" -#include "tests/validation/reference/Permute.h" using namespace arm_compute::experimental::dynamic_fusion; @@ -51,12 +48,13 @@ class DynamicFusionGpuElementwiseBinaryValidationGenericFixture : public framewo { public: template - void setup(ArithmeticOperation op, TensorShape shape0, TensorShape shape1, TensorShape shape2, const DataType data_type, const bool is_inplace) + void setup(ArithmeticOperation ref_op, const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2, DataType data_type, bool is_inplace, bool fuse_two_ops = false) { - _op = op; + _ref_op = ref_op; _is_inplace = is_inplace; _data_type = data_type; - _fuse = shape2.total_size() != 0; + _fuse = fuse_two_ops; + ARM_COMPUTE_ERROR_ON_MSG(_fuse && shape2.total_size() == 0, "No shape2 provided for fusion of two ops."); ARM_COMPUTE_ERROR_ON_MSG(_fuse && _is_inplace, "In place for fusing case not supported yet."); _target = compute_target(shape0, shape1, shape2); _reference = compute_reference(shape0, shape1, shape2); @@ -68,7 +66,7 @@ protected: { if(is_data_type_float(tensor.data_type())) { - switch(_op) + switch(_ref_op) { case ArithmeticOperation::DIV: library->fill_tensor_uniform_ranged(tensor, i, { std::pair(-0.001f, 0.001f) }); @@ -82,7 +80,7 @@ protected: } else if(tensor.data_type() == DataType::S32) { - switch(_op) + switch(_ref_op) { case ArithmeticOperation::DIV: library->fill_tensor_uniform_ranged(tensor, i, { std::pair(-1U, 1U) }); @@ -97,7 +95,7 @@ protected: } } - TensorType compute_target(TensorShape shape0, TensorShape shape1, TensorShape shape2) + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) { // Create a new workload sketch auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); @@ -105,7 +103,7 @@ protected: GpuWorkloadSketch sketch{ &gpu_ctx }; // Fuse first element wise binary Op - TensorInfo lhs_info = sketch.create_tensor_info(shape0, 1, _data_type); + TensorInfo lhs_info = sketch.create_tensor_info(TensorInfo(shape0, 1, _data_type)); TensorInfo rhs_info = sketch.create_tensor_info(TensorInfo(shape1, 1, _data_type)); TensorInfo dst_info = sketch.create_tensor_info(); @@ -115,7 +113,7 @@ protected: if(_fuse) { - rhs_info_fuse = sketch.create_tensor_info(shape2, 1, _data_type); + rhs_info_fuse = sketch.create_tensor_info(TensorInfo(shape2, 1, _data_type)); ITensorInfo *ans2_info = FunctionType::create_op(sketch, ans_info, &rhs_info_fuse); GpuOutput::create_op(sketch, ans2_info, &dst_info); } @@ -183,7 +181,7 @@ protected: return t_dst; } - SimpleTensor compute_reference(TensorShape shape0, TensorShape shape1, TensorShape shape2) + SimpleTensor compute_reference(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) { const TensorShape out_shape = TensorShape::broadcast_shape(shape0, shape1); const TensorShape out_shape_fuse = TensorShape::broadcast_shape(out_shape, shape1); @@ -194,21 +192,22 @@ protected: SimpleTensor ref_rhs_fuse{ shape2, _data_type, 1, QuantizationInfo() }; SimpleTensor ref_dst{ out_shape, _data_type, 1, QuantizationInfo() }; SimpleTensor ref_dst_fuse{ out_shape_fuse, _data_type, 1, QuantizationInfo() }; + // Fill reference fill(ref_lhs, 0); fill(ref_rhs, 1); - reference::arithmetic_operation(_op, ref_lhs, ref_rhs, ref_dst, ConvertPolicy::WRAP); + reference::arithmetic_operation(_ref_op, ref_lhs, ref_rhs, ref_dst, ConvertPolicy::WRAP); if(_fuse) { fill(ref_rhs_fuse, 2); - reference::arithmetic_operation(_op, ref_dst, ref_rhs_fuse, ref_dst_fuse, ConvertPolicy::WRAP); + reference::arithmetic_operation(_ref_op, ref_dst, ref_rhs_fuse, ref_dst_fuse, ConvertPolicy::WRAP); } SimpleTensor *ret = _fuse ? &ref_dst_fuse : &ref_dst; return *ret; } - ArithmeticOperation _op{ ArithmeticOperation::ADD }; + ArithmeticOperation _ref_op{ ArithmeticOperation::ADD }; TensorType _target{}; SimpleTensor _reference{}; DataType _data_type{}; @@ -222,9 +221,9 @@ class DynamicFusionGpuElementwiseBinaryOneOpValidationFixture : public DynamicFu { public: template - void setup(ArithmeticOperation op, TensorShape shape, const DataType data_type, const bool is_inplace) + void setup(ArithmeticOperation ref_op, const TensorShape &shape0, DataType data_type, bool is_inplace) { - DynamicFusionGpuElementwiseBinaryValidationGenericFixture::setup(op, shape, shape, TensorShape(), data_type, is_inplace); + DynamicFusionGpuElementwiseBinaryValidationGenericFixture::setup(ref_op, shape0, shape0, TensorShape(), data_type, is_inplace); } }; @@ -233,9 +232,9 @@ class DynamicFusionGpuElementwiseBinaryBroadcastOneOpValidationFixture : public { public: template - void setup(ArithmeticOperation op, TensorShape shape0, TensorShape shape1, const DataType data_type, const bool is_inplace) + void setup(ArithmeticOperation ref_op, const TensorShape &shape0, const TensorShape &shape1, DataType data_type, bool is_inplace) { - DynamicFusionGpuElementwiseBinaryValidationGenericFixture::setup(op, shape0, shape1, TensorShape(), data_type, is_inplace); + DynamicFusionGpuElementwiseBinaryValidationGenericFixture::setup(ref_op, shape0, shape1, TensorShape(), data_type, is_inplace); } }; @@ -244,9 +243,9 @@ class DynamicFusionGpuElementwiseBinaryTwoOpsValidationFixture : public DynamicF { public: template - void setup(ArithmeticOperation op, TensorShape shape0, TensorShape shape1, TensorShape shape2, const DataType data_type, const bool is_inplace) + void setup(ArithmeticOperation ref_op, const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2, DataType data_type, bool is_inplace, bool fuse_two_ops) { - DynamicFusionGpuElementwiseBinaryValidationGenericFixture::setup(op, shape0, shape1, shape2, data_type, is_inplace); + DynamicFusionGpuElementwiseBinaryValidationGenericFixture::setup(ref_op, shape0, shape1, shape2, data_type, is_inplace, fuse_two_ops); } }; diff --git a/tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h b/tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h new file mode 100644 index 0000000000..0530707c38 --- /dev/null +++ b/tests/validation/fixtures/dynamic_fusion/operators/MulFixture.h @@ -0,0 +1,235 @@ +/* +* Copyright (c) 2023 Arm Limited. +* +* SPDX-License-Identifier: MIT +* +* Permission is hereby granted, free of charge, to any person obtaining a copy +* of this software and associated documentation files (the "Software"), to +* deal in the Software without restriction, including without limitation the +* rights to use, copy, modify, merge, publish, distribute, sublicense, and/or +* sell copies of the Software, and to permit persons to whom the Software is +* furnished to do so, subject to the following conditions: +* +* The above copyright notice and this permission notice shall be included in all +* copies or substantial portions of the Software. +* +* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR +* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, +* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE +* AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER +* LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, +* OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE +* SOFTWARE. +*/ +#ifndef TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_MULFIXTURE +#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_MULFIXTURE + +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h" +#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuOutput.h" + +#include "tests/Globals.h" +#include "tests/framework/Fixture.h" +#include "tests/framework/Macros.h" +#include "tests/validation/reference/PixelWiseMultiplication.h" + +using namespace arm_compute::experimental::dynamic_fusion; + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +/* We use a separate test fixture for Multiplication op instead of reusing ElementwiseBinaryFixture to avoid exposing + * the internal enum ElementwiseOp to the public utils/TypePrinters.h as required by the data test case macros + * to print the test data. + */ +template +class DynamicFusionMulValidationFixture : public framework::Fixture +{ +public: + template + void setup(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2, DataType data_type, bool is_inplace, bool fuse_two_ops = false) + { + _data_type = data_type; + _is_inplace = is_inplace; + _fuse = fuse_two_ops; + ARM_COMPUTE_ERROR_ON_MSG(_fuse && shape2.total_size() == 0, "No shape2 provided for fusion of two ops."); + ARM_COMPUTE_ERROR_ON_MSG(_fuse && _is_inplace, "In place for fusing case not supported yet."); + _target = compute_target(shape0, shape1, shape2); + _reference = compute_reference(shape0, shape1, shape2); + } + +protected: + template + void fill(U &&tensor, int i) + { + library->fill_tensor_uniform(tensor, i); + } + + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) + { + // Create a new workload sketch + auto cl_compile_ctx = CLKernelLibrary::get().get_compile_context(); + auto gpu_ctx = GpuWorkloadContext{ &cl_compile_ctx }; + GpuWorkloadSketch sketch{ &gpu_ctx }; + + // Fuse first multiplication op + TensorInfo lhs_info = sketch.create_tensor_info(TensorInfo(shape0, 1, _data_type)); + TensorInfo rhs_info = sketch.create_tensor_info(TensorInfo(shape1, 1, _data_type)); + TensorInfo dst_info = sketch.create_tensor_info(); + + TensorInfo rhs_info_fuse; + + ITensorInfo *ans_info = FunctionType::create_op(sketch, &lhs_info, &rhs_info); + + if(_fuse) + { + rhs_info_fuse = sketch.create_tensor_info(TensorInfo(shape2, 1, _data_type)); + ITensorInfo *ans2_info = FunctionType::create_op(sketch, ans_info, &rhs_info_fuse); + GpuOutput::create_op(sketch, ans2_info, &dst_info); + } + else + { + GpuOutput::create_op(sketch, ans_info, &dst_info); + } + + // Configure runtime + ClWorkloadRuntime runtime; + runtime.configure(sketch); + + // (Important) Allocate auxiliary tensor memory if there are any + for(auto &data : runtime.get_auxiliary_tensors()) + { + CLTensor *tensor = std::get<0>(data); + TensorInfo info = std::get<1>(data); + AuxMemoryInfo aux_mem_req = std::get<2>(data); + tensor->allocator()->init(info, aux_mem_req.alignment); + tensor->allocator()->allocate(); // Use ACL allocated memory + } + + // Construct user tensors + TensorType t_lhs{}; + TensorType t_rhs{}; + TensorType t_rhs_fuse{}; + TensorType t_dst{}; + + // Initialize user tensors + t_lhs.allocator()->init(lhs_info); + t_rhs.allocator()->init(rhs_info); + t_dst.allocator()->init(dst_info); + if(_fuse) + { + t_rhs_fuse.allocator()->init(rhs_info_fuse); + } + + // Allocate and fill user tensors + // Instead of using ACL allocator, the user can choose to import memory into the tensors + t_lhs.allocator()->allocate(); + t_rhs.allocator()->allocate(); + t_dst.allocator()->allocate(); + if(_fuse) + { + t_rhs_fuse.allocator()->allocate(); + } + + fill(AccessorType(t_lhs), 0); + fill(AccessorType(t_rhs), 1); + if(_fuse) + { + fill(AccessorType(t_rhs_fuse), 2); + } + + // Run runtime + if(_fuse) + { + runtime.run({ &t_lhs, &t_rhs, &t_rhs_fuse, &t_dst }); + } + else + { + runtime.run({ &t_lhs, &t_rhs, &t_dst }); + } + + return t_dst; + } + + SimpleTensor compute_reference(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2) + { + // Create reference + SimpleTensor ref_lhs{ shape0, _data_type, 1, QuantizationInfo() }; + SimpleTensor ref_rhs{ shape1, _data_type, 1, QuantizationInfo() }; + SimpleTensor ref_rhs_fuse{ shape2, _data_type, 1, QuantizationInfo() }; + + // Fill reference + fill(ref_lhs, 0); + fill(ref_rhs, 1); + SimpleTensor ref_dst = reference::pixel_wise_multiplication(ref_lhs, + ref_rhs, + 1.f, + ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_UP, + _data_type, + QuantizationInfo()); + if(_fuse) + { + fill(ref_rhs_fuse, 2); + SimpleTensor ref_dst_fuse = reference::pixel_wise_multiplication(ref_dst, + ref_rhs_fuse, + 1.f, + ConvertPolicy::SATURATE, + RoundingPolicy::TO_NEAREST_UP, + _data_type, + QuantizationInfo()); + return ref_dst_fuse; + } + return ref_dst; + } + + TensorType _target{}; + SimpleTensor _reference{}; + DataType _data_type{}; + bool _is_inplace{ false }; + bool _fuse{ false }; +}; + +template +class DynamicFusionMulOneOpValidationFixture : public DynamicFusionMulValidationFixture +{ +public: + template + void setup(const TensorShape &shape0, DataType data_type, bool is_inplace) + { + DynamicFusionMulValidationFixture::setup(shape0, shape0, TensorShape(), data_type, is_inplace); + } +}; + +template +class DynamicFusionMulBroadcastValidationFixture : public DynamicFusionMulValidationFixture +{ +public: + template + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type, bool is_inplace) + { + DynamicFusionMulValidationFixture::setup(shape0, shape1, TensorShape(), data_type, is_inplace); + } +}; + +template +class DynamicFusionMulTwoOpsValidationFixture : public DynamicFusionMulValidationFixture +{ +public: + template + void setup(const TensorShape &shape0, const TensorShape &shape1, const TensorShape &shape2, DataType data_type, bool is_inplace, bool fuse_two_ops) + { + DynamicFusionMulValidationFixture::setup(shape0, shape1, shape2, data_type, is_inplace, fuse_two_ops); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_OPERATORS_MULFIXTURE */ -- cgit v1.2.1