diff options
Diffstat (limited to 'arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h')
| -rw-r--r-- | arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h | 107 |
1 files changed, 27 insertions, 80 deletions
diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h index dc9783f9eb..6d07675d3d 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2021 Arm Limited. + * Copyright (c) 2017-2021, 2023-2024 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,53 +21,34 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#ifndef ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H -#define ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H +#ifndef ACL_ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEGEMMLOWPMATRIXMULTIPLYCORE_H +#define ACL_ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEGEMMLOWPMATRIXMULTIPLYCORE_H -#include "NEActivationLayer.h" -#include "arm_compute/core/ITensorPack.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/function_info/GEMMInfo.h" #include "arm_compute/runtime/IFunction.h" #include "arm_compute/runtime/IMemoryManager.h" #include "arm_compute/runtime/IWeightsManager.h" -#include "arm_compute/runtime/MemoryGroup.h" -#include "arm_compute/runtime/Tensor.h" #include <memory> namespace arm_compute { class ITensor; -class NEConvertQuantizedSignednessKernel; -class NEGEMMInterleave4x4Kernel; -class NEGEMMLowpMatrixMultiplyKernel; -class NEGEMMLowpOffsetContributionKernel; -class NEGEMMLowpOffsetContributionOutputStageKernel; -class NEGEMMLowpMatrixAReductionKernel; -class NEGEMMLowpMatrixBReductionKernel; -class NEGEMMTranspose1xWKernel; -namespace cpu -{ -class CpuGemmAssemblyDispatch; -} +class ITensorInfo; -/** Basic function to execute GEMMLowpMatrixMultiplyCore. This function calls the following kernels if the DOT product instruction is not available: - * - * -# @ref NEGEMMInterleave4x4Kernel - * -# @ref NEGEMMTranspose1xWKernel - * -# @ref NEGEMMLowpMatrixMultiplyKernel - * -# @ref NEGEMMLowpOffsetContributionKernel - * -# @ref NEActivationLayer +/** Function to run Gemm on quantized types. * - * otherwise if the DOT product instruction is available: + * This function calls the following: * - * -# @ref NEGEMMLowpOffsetContributionKernel - * -*/ + * -# @ref cpu::CpuGemmLowpMatrixMultiplyCore + */ class NEGEMMLowpMatrixMultiplyCore : public IFunction { public: /** Constructor */ - NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager = nullptr, IWeightsManager *weights_manager = nullptr); + NEGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager = nullptr, + IWeightsManager *weights_manager = nullptr); /** Prevent instances of this class from being copied (As this class contains pointers) */ NEGEMMLowpMatrixMultiplyCore(const NEGEMMLowpMatrixMultiplyCore &) = delete; /** Default move constructor */ @@ -99,6 +80,7 @@ public: * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |S32 | * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |S32 | * |QASYMM8_SIGNED |QSYMM8 |S32 |S32 | + * |QASYMM8_SIGNED |QASYMM8_SIGNED |F32 |F32 | * * @note GEMM_LOWP: low precision GEMM kernel * This kernel performs the following computations: @@ -107,71 +89,36 @@ 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/QASYMM8_SIGNED otherwise + * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8/QASYMM8_SIGNED/F32 otherwise * * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8/QASYMM8_SIGNED. * @param[in] b Second input tensor (Matrix B). Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL. - * @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/QASYMM8/QASYMM8_SIGNED + * @param[in] c Third input tensor (Matrix C). It can be a nullptr. Data type supported: S32/F32 + * @param[out] output Output tensor. Data type supported: Data type supported: S32/QASYMM8/QASYMM8_SIGNED/F32 * @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()); + 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 * - * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8/QASYMM8_SIGNED otherwise - * - * @param[in] a First input tensor info (Matrix A). Data type supported: QASYMM8/QASYMM8_SIGNED. - * @param[in] b Second input tensor info (Matrix B). Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL. - * @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/QASYMM8_SIGNED - * @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 + * Similar to @ref NEGEMMLowpMatrixMultiplyCore::configure() * * @return a status */ - static Status validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info = GEMMInfo()); + static Status validate(const ITensorInfo *a, + const ITensorInfo *b, + const ITensorInfo *c, + const ITensorInfo *output, + const GEMMInfo &gemm_info = GEMMInfo()); // Inherited methods overridden void run() override; void prepare() override; private: - MemoryGroup _memory_group; - IWeightsManager *_weights_manager; - std::unique_ptr<cpu::CpuGemmAssemblyDispatch> _asm_glue; - std::unique_ptr<NEGEMMLowpMatrixMultiplyKernel> _mm_kernel; - std::unique_ptr<NEGEMMInterleave4x4Kernel> _mtx_a_reshape_kernel; - std::unique_ptr<NEGEMMTranspose1xWKernel> _mtx_b_reshape_kernel; - std::unique_ptr<NEGEMMLowpMatrixAReductionKernel> _mtx_a_reduction_kernel; - std::unique_ptr<NEGEMMLowpMatrixBReductionKernel> _mtx_b_reduction_kernel; - std::unique_ptr<NEGEMMLowpOffsetContributionKernel> _offset_contribution_kernel; - std::unique_ptr<NEGEMMLowpOffsetContributionOutputStageKernel> _offset_contribution_output_stage_kernel; - NEActivationLayer _activation_func; - std::unique_ptr<NEConvertQuantizedSignednessKernel> _convert_to_signed_asymm; - std::unique_ptr<NEConvertQuantizedSignednessKernel> _convert_from_signed_asymm; - - Tensor _vector_sum_col; - Tensor _vector_sum_row; - Tensor _tmp_a; - Tensor _tmp_b; - Tensor _mm_result_s32; - Tensor _signed_a; - Tensor _signed_output; - const ITensor *_original_b; - int32_t _a_offset; - int32_t _b_offset; - - bool _run_vector_matrix_multiplication; - bool _assembly_path; - bool _fused_assembly_path; - bool _reshape_b_only_on_first_run; - bool _is_prepared; - bool _fuse_output_stage; - bool _run_activation; - bool _flip_signedness; - - ITensorPack _asm_glue_tensors{}; + struct Impl; + std::unique_ptr<Impl> _impl; }; } // namespace arm_compute -#endif /*ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H */ +#endif // ACL_ARM_COMPUTE_RUNTIME_NEON_FUNCTIONS_NEGEMMLOWPMATRIXMULTIPLYCORE_H |