Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 1 | /* |
Matthew Bentham | f1aeab9 | 2023-05-30 13:35:34 +0000 | [diff] [blame] | 2 | * Copyright (c) 2021, 2023 Arm Limited. |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
SiCong Li | c5ab4df | 2023-10-17 17:38:57 +0100 | [diff] [blame] | 24 | #ifndef ACL_SRC_CPU_OPERATORS_CPUGEMMLOWPMATRIXMULTIPLYCORE_H |
| 25 | #define ACL_SRC_CPU_OPERATORS_CPUGEMMLOWPMATRIXMULTIPLYCORE_H |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 26 | |
| 27 | #include "arm_compute/core/TensorInfo.h" |
SiCong Li | 9129549 | 2023-07-21 18:16:13 +0100 | [diff] [blame] | 28 | #include "arm_compute/function_info/GEMMInfo.h" |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 29 | |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 30 | #include "src/core/common/Macros.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 31 | #include "src/cpu/ICpuOperator.h" |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 32 | |
| 33 | #include <memory> |
| 34 | |
| 35 | namespace arm_compute |
| 36 | { |
| 37 | namespace cpu |
| 38 | { |
| 39 | namespace kernels |
| 40 | { |
| 41 | class CpuGemmInterleave4x4Kernel; |
| 42 | class CpuGemmLowpMatrixMultiplyKernel; |
| 43 | class CpuGemmLowpOffsetContributionKernel; |
| 44 | class CpuGemmLowpOffsetContributionOutputStageKernel; |
| 45 | class CpuGemmLowpMatrixAReductionKernel; |
| 46 | class CpuGemmLowpMatrixBReductionKernel; |
| 47 | class CpuGemmTranspose1xWKernel; |
| 48 | class CpuConvertQuantizedSignednessKernel; |
| 49 | } // namespace kernels |
| 50 | class CpuGemmAssemblyDispatch; |
| 51 | class CpuActivation; |
| 52 | |
| 53 | /** Basic function to execute GEMMLowpMatrixMultiplyCore. This function calls the following kernels if the DOT product instruction is not available: |
| 54 | * |
| 55 | * -# @ref kernels::CpuGemmInterleave4x4Kernel |
| 56 | * -# @ref kernels::CpuGemmTranspose1xWKernel |
| 57 | * -# @ref kernels::CpuGemmLowpMatrixMultiplyKernel |
| 58 | * -# @ref kernels::CpuGemmLowpOffsetContributionKernel |
| 59 | * -# @ref CpuActivation |
| 60 | * |
| 61 | * otherwise if the DOT product instruction is available: |
| 62 | * |
| 63 | * -# @ref kernels::CpuGemmLowpOffsetContributionKernel |
| 64 | * |
| 65 | */ |
| 66 | class CpuGemmLowpMatrixMultiplyCore : public ICpuOperator |
| 67 | { |
| 68 | public: |
| 69 | /** Constructor */ |
| 70 | CpuGemmLowpMatrixMultiplyCore(); |
| 71 | ARM_COMPUTE_DISALLOW_COPY_ALLOW_MOVE(CpuGemmLowpMatrixMultiplyCore); |
| 72 | /** Destructor */ |
| 73 | ~CpuGemmLowpMatrixMultiplyCore(); |
| 74 | /** Initialise the kernel's inputs, output |
| 75 | * |
| 76 | * Valid data layouts: |
| 77 | * - NHWC |
| 78 | * - NCHW |
| 79 | * |
| 80 | * Valid data type configurations: |
| 81 | * |src0 |src1 |src2 |dst | |
| 82 | * |:--------------|:------------------|:--------|:--------------| |
| 83 | * |QASYMM8 |QASYMM8 |S32 |QASYMM8 | |
| 84 | * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |QASYMM8 | |
| 85 | * |QASYMM8 |QSYMM8 |S32 |QASYMM8 | |
| 86 | * |QASYMM8 |QASYMM8 |S32 |S32 | |
| 87 | * |QASYMM8 |QSYMM8_PER_CHANNEL |S32 |S32 | |
| 88 | * |QASYMM8 |QSYMM8 |S32 |S32 | |
| 89 | * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |QASYMM8_SIGNED | |
| 90 | * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |QASYMM8_SIGNED | |
| 91 | * |QASYMM8_SIGNED |QSYMM8 |S32 |QASYMM8_SIGNED | |
| 92 | * |QASYMM8_SIGNED |QASYMM8_SIGNED |S32 |S32 | |
| 93 | * |QASYMM8_SIGNED |QSYMM8_PER_CHANNEL |S32 |S32 | |
| 94 | * |QASYMM8_SIGNED |QSYMM8 |S32 |S32 | |
| 95 | * |
| 96 | * @note GEMM_LOWP: low precision GEMM kernel |
| 97 | * This kernel performs the following computations: |
| 98 | * |
| 99 | * -# Convert a values from QASYMM8 to int32 and add a_offset to each of them. |
| 100 | * -# Convert b values from QASYMM8 to int32 add b_offset to each of them. |
| 101 | * -# Compute the matrix product of the resulting a * b in int32. |
| 102 | * |
| 103 | * @note The @p output type is S32 if @p gemm_info.type == GEMMLowpOutputStageType::NONE. It is QASYMM8/QASYMM8_SIGNED otherwise |
| 104 | * |
| 105 | * @param[in] a First input tensor info (Matrix A). Data type supported: QASYMM8/QASYMM8_SIGNED. |
| 106 | * @param[in] b Second input tensor info (Matrix B). Data type supported: QASYMM8/QASYMM8_SIGNED/QSYMM8/QSYMM8_PER_CHANNEL. |
| 107 | * @param[in] c Third input tensor info (Matrix C). It can be a nullptr. Data type supported: S32 |
| 108 | * @param[out] dst Output tensor info. Data type supported: Data type supported: S32/QASYMM8/QASYMM8_SIGNED |
| 109 | * @param[in] gemm_info (Optional) Specifies if the matrix A and/or matrix B have been reshaped and |
| 110 | * if the reshape of matrix B should be executed only for the first run |
| 111 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 112 | void configure(const ITensorInfo *a, |
| 113 | const ITensorInfo *b, |
| 114 | const ITensorInfo *c, |
| 115 | ITensorInfo *dst, |
| 116 | const GEMMInfo &gemm_info = GEMMInfo()); |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 117 | /** Static function to check if given info will lead to a valid configuration |
| 118 | * |
| 119 | * Similar to CpuGemmLowpMatrixMultiplyCore::configure() |
| 120 | * |
| 121 | * @return a status |
| 122 | */ |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 123 | static Status validate(const ITensorInfo *a, |
| 124 | const ITensorInfo *b, |
| 125 | const ITensorInfo *c, |
| 126 | const ITensorInfo *dst, |
| 127 | const GEMMInfo &gemm_info = GEMMInfo()); |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 128 | |
| 129 | // Inherited methods overridden: |
Felix Thomasmathibalan | afd38f0 | 2023-09-27 17:46:17 +0100 | [diff] [blame] | 130 | void run(ITensorPack &tensors) override; |
| 131 | void prepare(ITensorPack &tensors) override; |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 132 | experimental::MemoryRequirements workspace() const override; |
| 133 | |
| 134 | private: |
| 135 | enum AuxTensorIdx |
| 136 | { |
SiCong Li | c5ab4df | 2023-10-17 17:38:57 +0100 | [diff] [blame] | 137 | /* Slots 0 - 2 reserved for CpuGemmAssemblyDispatch */ |
| 138 | VectorSumCol = 3, |
Manuel Bottini | cfac51c | 2021-06-18 15:47:28 +0100 | [diff] [blame] | 139 | VectorSumRow, |
| 140 | TmpA, |
| 141 | TmpB, |
| 142 | MMResultS32, |
| 143 | SignedA, |
| 144 | SignedOutput, |
| 145 | Count |
| 146 | }; |
| 147 | |
| 148 | std::unique_ptr<CpuGemmAssemblyDispatch> _asm_glue; |
| 149 | std::unique_ptr<kernels::CpuGemmLowpMatrixMultiplyKernel> _mm_kernel; |
| 150 | std::unique_ptr<kernels::CpuGemmInterleave4x4Kernel> _mtx_a_reshape_kernel; |
| 151 | std::unique_ptr<kernels::CpuGemmTranspose1xWKernel> _mtx_b_reshape_kernel; |
| 152 | std::unique_ptr<kernels::CpuGemmLowpMatrixAReductionKernel> _mtx_a_reduction_kernel; |
| 153 | std::unique_ptr<kernels::CpuGemmLowpMatrixBReductionKernel> _mtx_b_reduction_kernel; |
| 154 | std::unique_ptr<kernels::CpuGemmLowpOffsetContributionKernel> _offset_contribution_kernel; |
| 155 | std::unique_ptr<kernels::CpuGemmLowpOffsetContributionOutputStageKernel> _offset_contribution_output_stage_kernel; |
| 156 | std::unique_ptr<CpuActivation> _activation_func; |
| 157 | std::unique_ptr<kernels::CpuConvertQuantizedSignednessKernel> _convert_to_signed_asymm; |
| 158 | std::unique_ptr<kernels::CpuConvertQuantizedSignednessKernel> _convert_from_signed_asymm; |
| 159 | |
| 160 | TensorInfo _vector_sum_col; |
| 161 | TensorInfo _vector_sum_row; |
| 162 | TensorInfo _tmp_a; |
| 163 | TensorInfo _tmp_b; |
| 164 | TensorInfo _mm_result_s32; |
| 165 | TensorInfo _signed_a; |
| 166 | TensorInfo _signed_output; |
| 167 | int32_t _a_offset; |
| 168 | int32_t _b_offset; |
| 169 | |
| 170 | bool _run_vector_matrix_multiplication; |
| 171 | bool _assembly_path; |
| 172 | bool _fused_assembly_path; |
| 173 | bool _reshape_b_only_on_first_run; |
| 174 | bool _is_prepared; |
| 175 | bool _fuse_output_stage; |
| 176 | bool _run_activation; |
| 177 | bool _flip_signedness; |
| 178 | GEMMInfo _gemm_info; |
| 179 | experimental::MemoryRequirements _aux_mem{}; |
| 180 | }; |
| 181 | } // namespace cpu |
| 182 | } // namespace arm_compute |
SiCong Li | c5ab4df | 2023-10-17 17:38:57 +0100 | [diff] [blame] | 183 | #endif // ACL_SRC_CPU_OPERATORS_CPUGEMMLOWPMATRIXMULTIPLYCORE_H |