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 --- .../NEON/functions/NEGEMMLowpMatrixMultiplyCore.h | 60 ++++++++++++---------- 1 file changed, 34 insertions(+), 26 deletions(-) (limited to 'arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h') diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h index 682475c824..d3b27e4faf 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2018 ARM Limited. + * Copyright (c) 2017-2019 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -26,6 +26,7 @@ #include "arm_compute/core/NEON/INEKernel.h" #include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionOutputStageKernel.h" #include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h" #include "arm_compute/runtime/IFunction.h" #include "arm_compute/runtime/IMemoryManager.h" @@ -73,20 +74,24 @@ public: * -# Convert b values from QASYMM8 to int32 add b_offset to each of them. * -# Compute the matrix product of the resulting a * b in int32. * + * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8 otherwise + * * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8. * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a * @param[in] c Third input tensor (Matrix C). It can be a nullptr. Data type supported: S32 - * @param[out] output Output tensor. Data type supported: Data type supported: S32 + * @param[out] output Output tensor. Data type supported: Data type supported: S32/QASYMM8 * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and * if the reshape of matrix B should be executed only for the first run */ void configure(const ITensor *a, const ITensor *b, const ITensor *c, ITensor *output, const GEMMInfo &gemm_info = GEMMInfo()); /** Static function to check if given info will lead to a valid configuration of @ref NEGEMMLowpMatrixMultiplyCore * - * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8. - * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a - * @param[in] c Third input tensor (Matrix C). It can be a nullptr. Data type supported: S32 - * @param[in] output Output tensor. Data type supported: Data type supported: S32 + * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8 otherwise + * + * @param[in] a First input tensor info (Matrix A). Data type supported: QASYMM8. + * @param[in] b Second input tensor info (Matrix B). Data type supported: same as @p a + * @param[in] c Third input tensor info (Matrix C). It can be a nullptr. Data type supported: S32 + * @param[in] output Output tensor info. Data type supported: Data type supported: S32/QASYMM8 * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and * if the reshape of matrix B should be executed only for the first run * @@ -99,25 +104,28 @@ public: void prepare() override; private: - MemoryGroup _memory_group; - NEGEMMAssemblyDispatch _asm_glue; - std::unique_ptr _mm_kernel; - std::unique_ptr _mtx_a_reshape_kernel; - std::unique_ptr _mtx_b_reshape_kernel; - NEGEMMLowpMatrixAReductionKernel _mtx_a_reduction_kernel; - NEGEMMLowpMatrixBReductionKernel _mtx_b_reduction_kernel; - NEGEMMLowpOffsetContributionKernel _offset_contribution_kernel; - Tensor _vector_sum_col; - Tensor _vector_sum_row; - Tensor _tmp_a; - Tensor _tmp_b; - const ITensor *_original_b; - int32_t _a_offset; - int32_t _b_offset; - bool _run_vector_matrix_multiplication; - bool _dot_product_path; - bool _reshape_b_only_on_first_run; - bool _is_prepared; + MemoryGroup _memory_group; + NEGEMMAssemblyDispatch _asm_glue; + std::unique_ptr _mm_kernel; + std::unique_ptr _mtx_a_reshape_kernel; + std::unique_ptr _mtx_b_reshape_kernel; + NEGEMMLowpMatrixAReductionKernel _mtx_a_reduction_kernel; + NEGEMMLowpMatrixBReductionKernel _mtx_b_reduction_kernel; + NEGEMMLowpOffsetContributionKernel _offset_contribution_kernel; + NEGEMMLowpOffsetContributionOutputStageKernel _offset_contribution_output_stage_kernel; + Tensor _vector_sum_col; + Tensor _vector_sum_row; + Tensor _tmp_a; + Tensor _tmp_b; + Tensor _mm_result_s32; + const ITensor *_original_b; + int32_t _a_offset; + int32_t _b_offset; + bool _run_vector_matrix_multiplication; + bool _dot_product_path; + bool _reshape_b_only_on_first_run; + bool _is_prepared; + bool _fuse_output_stage; }; -} +} // namespace arm_compute #endif /*__ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H__ */ -- cgit v1.2.1