Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 1 | /* |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2021 Arm Limited. |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [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 | */ |
| 24 | #include "arm_compute/runtime/CL/functions/CLGEMMLowpMatrixMultiplyCore.h" |
| 25 | |
| 26 | #include "arm_compute/core/CL/ICLTensor.h" |
| 27 | #include "arm_compute/core/Error.h" |
| 28 | #include "arm_compute/core/Helpers.h" |
Michele Di Giorgio | f64d336 | 2020-04-03 12:40:10 +0100 | [diff] [blame] | 29 | #include "arm_compute/core/KernelDescriptors.h" |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 30 | #include "arm_compute/core/Log.h" |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 31 | #include "arm_compute/core/TensorInfo.h" |
| 32 | #include "arm_compute/core/Types.h" |
| 33 | #include "arm_compute/core/Validate.h" |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 34 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 35 | #include "arm_compute/core/utils/quantization/AsymmHelpers.h" |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 36 | #include "arm_compute/runtime/CL/CLScheduler.h" |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 37 | #include "src/core/CL/kernels/CLDepthConvertLayerKernel.h" |
| 38 | #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyNativeKernel.h" |
| 39 | #include "src/core/CL/kernels/CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.h" |
| 40 | #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionKernel.h" |
| 41 | #include "src/core/CL/kernels/CLGEMMLowpOffsetContributionOutputStageKernel.h" |
| 42 | #include "src/core/CL/kernels/CLGEMMLowpReductionKernel.h" |
| 43 | #include "src/core/CL/kernels/CLGEMMReshapeRHSMatrixKernel.h" |
Sang-Hoon Park | 68dd25f | 2020-10-19 16:00:11 +0100 | [diff] [blame] | 44 | #include "src/core/helpers/AutoConfiguration.h" |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 45 | #include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h" |
| 46 | #include "utils/TypePrinter.h" |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 47 | |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 48 | namespace arm_compute |
| 49 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 50 | using namespace arm_compute::misc::shape_calculator; |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 51 | using namespace arm_compute::cl_gemm; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 52 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 53 | namespace |
| 54 | { |
SiCong Li | 1a28e73 | 2021-02-10 16:57:33 +0000 | [diff] [blame] | 55 | inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type) |
| 56 | { |
| 57 | switch(kernel_type) |
| 58 | { |
| 59 | case CLGEMMKernelType::NATIVE: |
| 60 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 61 | { |
| 62 | return true; |
| 63 | } |
| 64 | default: |
| 65 | { |
| 66 | return false; |
| 67 | } |
| 68 | } |
| 69 | } |
| 70 | //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type |
| 71 | inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run) |
| 72 | { |
| 73 | auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run); |
| 74 | if(bool(gemm_kernel)) |
| 75 | { |
| 76 | if(validate_gemm_kernel(gemm_kernel.gemm_type)) |
| 77 | { |
| 78 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); |
| 79 | return gemm_kernel.gemm_type; |
| 80 | } |
| 81 | } |
| 82 | gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run); |
| 83 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); |
| 84 | return gemm_kernel.gemm_type; |
| 85 | } |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 86 | // Validate lhs_info and rhs_info for native kernel |
| 87 | inline bool validate_lhs_rhs_info_native(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const GEMMReshapeInfo &reshape_info) |
| 88 | { |
| 89 | // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel |
| 90 | TensorInfo mm_result_s32_info{}; |
| 91 | // Output tensor auto initialization if not yet initialized |
| 92 | auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*a, *b, false, reshape_info)).set_data_type(DataType::S32)); |
| 93 | // Validate mm kernel |
| 94 | // NOTE: Ignore all other parameters (eg. output stage etc.) and only validate lhs and rhs info |
| 95 | // NOTE: This assumes: |
| 96 | // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in CLGEMMLowpMatrixMultiplyNativeKernel.cpp validate_arguments). |
| 97 | // 2. lhs and rhs info does not cause window and padding issues through side effects (in CLGEMMLowpMatrixMultiplyNativeKernel.cpp validate_and_configure_window). |
| 98 | if(!bool(CLGEMMLowpMatrixMultiplyNativeKernel::validate(a, b, &mm_result_s32_info, lhs_info, rhs_info, reshape_info))) |
| 99 | { |
| 100 | return false; |
| 101 | } |
| 102 | return true; |
| 103 | } |
| 104 | |
| 105 | // Automatically select between mlgo (prioritized) and default heuristics for native kernel configs |
| 106 | std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_native(auto_heuristics::CommonQuery query, const ITensorInfo *a, const ITensorInfo *b, const GEMMReshapeInfo &reshape_info) |
| 107 | { |
| 108 | auto config = auto_heuristics::select_mlgo_gemm_config_native(query); |
| 109 | if(config) |
| 110 | { |
| 111 | if(validate_lhs_rhs_info_native(config.lhs_info, config.rhs_info, a, b, reshape_info)) |
| 112 | { |
| 113 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use native config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); |
| 114 | return { config.lhs_info, config.rhs_info }; |
| 115 | } |
| 116 | } |
| 117 | config = auto_heuristics::select_default_gemm_config_native(query); |
| 118 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use native config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); |
| 119 | return { config.lhs_info, config.rhs_info }; |
| 120 | } |
| 121 | |
| 122 | // Validate lhs_info and rhs_info for reshaped only rhs kernel |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 123 | inline bool validate_lhs_rhs_info_reshaped_only_rhs(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *output, |
| 124 | unsigned int m, unsigned int n, unsigned int k, bool reinterpret_input_as_3d, int depth_output_gemm3d) |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 125 | { |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 126 | // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel |
| 127 | TensorInfo tmp_b_info{}; |
| 128 | // Validate reshape RHS kernel |
| 129 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 130 | if(!bool(CLGEMMReshapeRHSMatrixKernel::validate(b, &tmp_b_info, rhs_info))) |
| 131 | { |
| 132 | return false; |
| 133 | } |
| 134 | // Validate mm kernel |
| 135 | // NOTE: Ignore all other parameters (eg. depth_output_gemm3d, output stage etc.) and only validate lhs and rhs info |
| 136 | // NOTE: This assumes: |
| 137 | // 1. lhs and rhs info's validity does not depend on these other parameters and vice versa(in CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_arguments). |
| 138 | // 2. lhs and rhs info does not cause window and padding issues through side effects (in CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel.cpp validate_and_configure_window). |
| 139 | GEMMKernelInfo gemm_kernel_info; |
| 140 | gemm_kernel_info.m = m; |
| 141 | gemm_kernel_info.n = n; |
| 142 | gemm_kernel_info.k = k; |
| 143 | gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 144 | gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 145 | gemm_kernel_info.lhs_info = lhs_info; |
| 146 | gemm_kernel_info.rhs_info = rhs_info; |
| 147 | // Since we ignore the output stage, output data type has to be S32 to pass the validation |
| 148 | TensorInfo output_info_copy(*output); |
| 149 | output_info_copy.set_data_type(DataType::S32); |
| 150 | if(!bool(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(a, &tmp_b_info, &output_info_copy, gemm_kernel_info))) |
| 151 | { |
| 152 | return false; |
| 153 | } |
| 154 | return true; |
| 155 | } |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 156 | |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 157 | // Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 158 | std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, bool reinterpret_input_as_3d, int depth_output_gemm3d, |
| 159 | const ITensorInfo *a, |
| 160 | const ITensorInfo *b, const ITensorInfo *output) |
| 161 | { |
| 162 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query); |
| 163 | if(config) |
| 164 | { |
| 165 | if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, output, query.m, query.n, query.k, reinterpret_input_as_3d, depth_output_gemm3d)) |
| 166 | { |
| 167 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); |
| 168 | return { config.lhs_info, config.rhs_info }; |
| 169 | } |
| 170 | } |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 171 | config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query); |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 172 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped_only_rhs config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); |
| 173 | return { config.lhs_info, config.rhs_info }; |
| 174 | } |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 175 | |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 176 | inline bool is_gemm_reshaped(CLGEMMKernelType kernel_type) |
| 177 | { |
| 178 | switch(kernel_type) |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 179 | { |
| 180 | case CLGEMMKernelType::NATIVE: |
| 181 | return false; |
| 182 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 183 | return true; |
| 184 | default: |
| 185 | ARM_COMPUTE_ERROR("Not supported gemmlowp kernel!"); |
| 186 | } |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 187 | } |
| 188 | } // namespace |
| 189 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 190 | CLGEMMLowpMatrixMultiplyCore::CLGEMMLowpMatrixMultiplyCore(std::shared_ptr<IMemoryManager> memory_manager) |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 191 | : _memory_group(std::move(memory_manager)), |
Georgios Pinitas | 40f51a6 | 2020-11-21 03:04:18 +0000 | [diff] [blame] | 192 | _weights_to_qasymm8(std::make_unique<CLDepthConvertLayerKernel>()), |
| 193 | _mm_native_kernel(std::make_unique<CLGEMMLowpMatrixMultiplyNativeKernel>()), |
| 194 | _mm_reshaped_only_rhs_kernel(std::make_unique<CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel>()), |
| 195 | _mtx_b_reshape_kernel(std::make_unique<CLGEMMReshapeRHSMatrixKernel>()), |
| 196 | _mtx_a_reduction_kernel(std::make_unique<CLGEMMLowpMatrixAReductionKernel>()), |
| 197 | _mtx_b_reduction_kernel(std::make_unique<CLGEMMLowpMatrixBReductionKernel>()), |
| 198 | _offset_contribution_kernel(std::make_unique<CLGEMMLowpOffsetContributionKernel>()), |
| 199 | _offset_contribution_output_stage_kernel(std::make_unique<CLGEMMLowpOffsetContributionOutputStageKernel>()), |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 200 | _qasymm8_weights(), |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 201 | _vector_sum_col(), |
| 202 | _vector_sum_row(), |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 203 | _tmp_b(), |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 204 | _mm_result_s32(), |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 205 | _gemm_output_stage_multipliers(), |
| 206 | _gemm_output_stage_shifts(), |
| 207 | _matrix_a(nullptr), |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 208 | _original_b(nullptr), |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 209 | _output(nullptr), |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 210 | _a_offset(0), |
| 211 | _b_offset(0), |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 212 | _is_gemm_reshaped(true), |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 213 | _reshape_b_only_on_first_run(false), |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 214 | _is_prepared(false), |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 215 | _run_output_stage(false), |
| 216 | _convert_to_qasymm8(false), |
| 217 | _run_offset_contribution(false) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 218 | { |
| 219 | } |
| 220 | |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 221 | CLGEMMLowpMatrixMultiplyCore::~CLGEMMLowpMatrixMultiplyCore() = default; |
| 222 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 223 | void CLGEMMLowpMatrixMultiplyCore::configure(const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, const GEMMInfo &gemm_info) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 224 | { |
Manuel Bottini | 2b84be5 | 2020-04-08 10:15:51 +0100 | [diff] [blame] | 225 | configure(CLKernelLibrary::get().get_compile_context(), a, b, c, output, gemm_info); |
| 226 | } |
| 227 | |
| 228 | void CLGEMMLowpMatrixMultiplyCore::configure(const CLCompileContext &compile_context, const ICLTensor *a, const ICLTensor *b, const ICLTensor *c, ICLTensor *output, const GEMMInfo &gemm_info) |
| 229 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 230 | ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 231 | ARM_COMPUTE_ERROR_THROW_ON(CLGEMMLowpMatrixMultiplyCore::validate(a->info(), b->info(), c != nullptr ? c->info() : nullptr, output->info(), gemm_info)); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 232 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 233 | _is_prepared = false; |
| 234 | _original_b = b; |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 235 | _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 236 | _a_offset = a->info()->quantization_info().uniform().offset; |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 237 | _matrix_a = a; |
| 238 | _output = output; |
| 239 | |
| 240 | _convert_to_qasymm8 = is_data_type_quantized_per_channel(b->info()->data_type()) && is_data_type_quantized_symmetric(b->info()->data_type()) |
morgolock | d13931d | 2020-06-23 15:49:35 +0100 | [diff] [blame] | 241 | && a->info()->data_type() == DataType::QASYMM8; |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 242 | _b_offset = _convert_to_qasymm8 ? -128 : b->info()->quantization_info().uniform().offset; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 243 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 244 | // Get the GPU target |
| 245 | const GPUTarget gpu_target = CLScheduler::get().target(); |
Gian Marco | 7b4d547 | 2018-01-10 15:56:30 +0000 | [diff] [blame] | 246 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 247 | // Set the target for the kernels |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 248 | _mm_native_kernel->set_target(gpu_target); |
| 249 | _mm_reshaped_only_rhs_kernel->set_target(gpu_target); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 250 | |
giuros01 | 8b6b4a9 | 2018-12-18 19:01:33 +0000 | [diff] [blame] | 251 | GEMMRHSMatrixInfo rhs_info; |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 252 | GEMMLHSMatrixInfo lhs_info; |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 253 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 254 | // Arguments used by GEMMReshapeInfo |
| 255 | // If we pass the matrix A and matrix B reshaped to CLGEMMMatrixMultiplyKernel, we need to pass m, n, k, mult_transpose1xW_width and mult_interleave4x4_height to CLGEMMReshapeInfo |
| 256 | // in order to know how the matrices have been reshaped |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 257 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 258 | const unsigned int m = reinterpret_input_as_3d ? (a->info()->dimension(1) * a->info()->dimension(2)) : a->info()->dimension(1); |
| 259 | const unsigned int n = b->info()->dimension(0); |
| 260 | const unsigned int k = a->info()->dimension(0); |
| 261 | const unsigned int batch_size = reinterpret_input_as_3d ? a->info()->dimension(3) : a->info()->dimension(2); |
| 262 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 263 | |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 264 | const auto reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d); |
| 265 | |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 266 | // Check if we need to reshape the matrix A and matrix B |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 267 | _is_gemm_reshaped = is_gemm_reshaped(auto_select_gemm_kernel(auto_heuristics::CommonQuery{ gpu_target, a->info()->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run)); |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 268 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 269 | if(_convert_to_qasymm8) |
| 270 | { |
| 271 | // Set data type for converted weights |
| 272 | TensorInfo weights_info(*b->info()); |
| 273 | weights_info.set_data_type(DataType::QASYMM8); |
| 274 | _qasymm8_weights.allocator()->init(weights_info); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 275 | _weights_to_qasymm8->configure(compile_context, b, &_qasymm8_weights, ConvertPolicy::WRAP, 0); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 276 | } |
| 277 | |
| 278 | const ICLTensor *matrix_b = _convert_to_qasymm8 ? &_qasymm8_weights : b; |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 279 | if(_is_gemm_reshaped) |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 280 | { |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 281 | matrix_b = &_tmp_b; |
| 282 | |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 283 | if(!_reshape_b_only_on_first_run) |
| 284 | { |
| 285 | _memory_group.manage(&_tmp_b); |
| 286 | } |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 287 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 288 | // Pick up the GEMM configuration |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 289 | // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration |
| 290 | std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, reinterpret_input_as_3d, |
| 291 | depth_output_gemm3d, |
| 292 | a->info(), _convert_to_qasymm8 ? _qasymm8_weights.info() : b->info(), output->info()); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 293 | |
Gian Marco Iodice | 926afe1 | 2019-03-19 11:44:13 +0000 | [diff] [blame] | 294 | // Configure reshape RHS kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 295 | _mtx_b_reshape_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_tmp_b, rhs_info); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 296 | } |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 297 | |
Michele Di Giorgio | f64d336 | 2020-04-03 12:40:10 +0100 | [diff] [blame] | 298 | // Using default reduction info |
| 299 | const GEMMLowpReductionKernelInfo reduction_info {}; |
| 300 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 301 | // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 |
| 302 | if(_a_offset != 0) |
| 303 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 304 | TensorInfo info_vector_sum_col(compute_reductionA_shape(*b->info()), 1, DataType::S32); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 305 | _vector_sum_col.allocator()->init(info_vector_sum_col); |
Giorgio Arena | bb54e4e | 2018-04-05 17:20:34 +0100 | [diff] [blame] | 306 | if(!_reshape_b_only_on_first_run) |
| 307 | { |
| 308 | _memory_group.manage(&_vector_sum_col); |
| 309 | } |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 310 | |
| 311 | // Configure Matrix B reduction kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 312 | _mtx_b_reduction_kernel->configure(compile_context, _convert_to_qasymm8 ? &_qasymm8_weights : b, &_vector_sum_col, reduction_info); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 313 | } |
| 314 | |
| 315 | // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 |
| 316 | if(_b_offset != 0) |
| 317 | { |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 318 | TensorInfo info_vector_sum_row(compute_reductionB_shape(*a->info()), 1, DataType::S32); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 319 | _vector_sum_row.allocator()->init(info_vector_sum_row); |
| 320 | _memory_group.manage(&_vector_sum_row); |
| 321 | |
| 322 | // Configure matrix A reduction kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 323 | _mtx_a_reduction_kernel->configure(compile_context, a, &_vector_sum_row, reduction_info); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 324 | } |
| 325 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 326 | GEMMKernelInfo gemm_kernel_info; |
| 327 | gemm_kernel_info.m = m; |
| 328 | gemm_kernel_info.n = n; |
| 329 | gemm_kernel_info.k = k; |
| 330 | gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 331 | gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 332 | gemm_kernel_info.lhs_info = lhs_info; |
| 333 | gemm_kernel_info.rhs_info = rhs_info; |
| 334 | gemm_kernel_info.a_offset = _a_offset; |
| 335 | gemm_kernel_info.b_offset = _b_offset; |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 336 | // If GEMMLowpOutputStage != NONE, fuse the offset contribution with the output stage |
| 337 | if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE) |
| 338 | { |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 339 | // Configure offset contribution kernel |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 340 | const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1; |
| 341 | |
| 342 | _gemm_output_stage_multipliers.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32)); |
| 343 | _gemm_output_stage_shifts.allocator()->init(TensorInfo(TensorShape(num_filters), 1, DataType::S32)); |
| 344 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 345 | GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage(); |
| 346 | gemmlowp_output_stage.output_data_type = _matrix_a->info()->data_type(); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 347 | |
| 348 | gemm_kernel_info.output_stage = gemmlowp_output_stage; |
| 349 | |
| 350 | if(_is_gemm_reshaped && gemmlowp_output_stage.type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) |
| 351 | { |
| 352 | // Configure and tune matrix multiply kernel with fused output stage |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 353 | _mm_reshaped_only_rhs_kernel->configure(compile_context, _matrix_a, matrix_b, output, gemm_kernel_info, _a_offset == 0 ? nullptr : &_vector_sum_col, |
| 354 | _b_offset == 0 ? nullptr : &_vector_sum_row, c, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 355 | } |
| 356 | else |
| 357 | { |
| 358 | _run_output_stage = true; |
| 359 | |
| 360 | _memory_group.manage(&_mm_result_s32); |
| 361 | |
| 362 | if(_is_gemm_reshaped) |
| 363 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 364 | _mm_reshaped_only_rhs_kernel->configure(compile_context, _matrix_a, matrix_b, &_mm_result_s32, gemm_kernel_info); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 365 | } |
| 366 | else |
| 367 | { |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 368 | // Pick up the GEMM configuration |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 369 | // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration |
| 370 | std::tie(lhs_info, rhs_info) = auto_select_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, |
| 371 | _matrix_a->info(), _convert_to_qasymm8 ? _qasymm8_weights.info() : matrix_b->info(), reshape_info); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 372 | |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 373 | // Configure matrix multiply kernel |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 374 | _mm_native_kernel->configure(compile_context, _matrix_a, matrix_b, &_mm_result_s32, lhs_info, rhs_info, reshape_info); |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 375 | |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 376 | _offset_contribution_output_stage_kernel->configure(compile_context, &_mm_result_s32, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, output, |
| 377 | a->info()->dimension(0), |
| 378 | _a_offset, _b_offset, gemmlowp_output_stage, &_gemm_output_stage_multipliers, &_gemm_output_stage_shifts); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 379 | _mm_result_s32.allocator()->allocate(); |
| 380 | } |
| 381 | } |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 382 | |
| 383 | _gemm_output_stage_multipliers.allocator()->allocate(); |
| 384 | _gemm_output_stage_shifts.allocator()->allocate(); |
| 385 | // Compute GEMM output multipliers and shifts for output stage |
| 386 | _gemm_output_stage_multipliers.map(); |
| 387 | _gemm_output_stage_shifts.map(); |
| 388 | std::memcpy(_gemm_output_stage_multipliers.ptr_to_element(Coordinates(0)), gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.data(), num_filters * sizeof(int32_t)); |
| 389 | std::memcpy(_gemm_output_stage_shifts.ptr_to_element(Coordinates(0)), gemm_info.gemmlowp_output_stage().gemmlowp_shifts.data(), num_filters * sizeof(int32_t)); |
| 390 | _gemm_output_stage_multipliers.unmap(); |
| 391 | _gemm_output_stage_shifts.unmap(); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 392 | } |
| 393 | else |
| 394 | { |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 395 | _run_offset_contribution = true; |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 396 | if(_is_gemm_reshaped) |
| 397 | { |
| 398 | // Configure and tune matrix multiply kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 399 | _mm_reshaped_only_rhs_kernel->configure(compile_context, _matrix_a, matrix_b, output, gemm_kernel_info); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 400 | } |
| 401 | else |
| 402 | { |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 403 | // Pick up the GEMM configuration |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 404 | // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration |
| 405 | std::tie(lhs_info, rhs_info) = auto_select_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }, |
| 406 | a->info(), _convert_to_qasymm8 ? _qasymm8_weights.info() : b->info(), reshape_info); |
Gian Marco Iodice | 06be6f8 | 2019-06-24 17:47:51 +0100 | [diff] [blame] | 407 | |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 408 | // Configure matrix multiply kernel |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 409 | _mm_native_kernel->configure(compile_context, _matrix_a, matrix_b, output, lhs_info, rhs_info, reshape_info); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 410 | } |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 411 | |
| 412 | // Configure offset contribution kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 413 | _offset_contribution_kernel->configure(compile_context, output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, c, a->info()->dimension(0), _a_offset, |
| 414 | _b_offset); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 415 | } |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 416 | |
| 417 | // Allocate tensors |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 418 | if(_is_gemm_reshaped) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 419 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 420 | if(!_reshape_b_only_on_first_run) |
| 421 | { |
| 422 | _tmp_b.allocator()->allocate(); |
| 423 | } |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 424 | } |
| 425 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 426 | if(_a_offset != 0 && !_reshape_b_only_on_first_run) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 427 | { |
| 428 | _vector_sum_col.allocator()->allocate(); |
| 429 | } |
| 430 | |
| 431 | if(_b_offset != 0) |
| 432 | { |
| 433 | _vector_sum_row.allocator()->allocate(); |
| 434 | } |
| 435 | } |
| 436 | |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 437 | Status CLGEMMLowpMatrixMultiplyCore::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, const GEMMInfo &gemm_info) |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 438 | { |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 439 | ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); |
| 440 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED); |
Michele Di Giorgio | 1c1b3aa | 2020-04-02 17:35:42 +0100 | [diff] [blame] | 441 | ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(b, 1, DataType::QASYMM8, DataType::QASYMM8_SIGNED, DataType::QSYMM8, DataType::QSYMM8_PER_CHANNEL); |
| 442 | ARM_COMPUTE_RETURN_ERROR_ON(a->data_type() == DataType::QASYMM8 && b->data_type() == DataType::QASYMM8_SIGNED); |
| 443 | ARM_COMPUTE_RETURN_ERROR_ON(a->data_type() == DataType::QASYMM8_SIGNED && b->data_type() == DataType::QASYMM8); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 444 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_a_reshaped(), "Matrix A already reshaped is not supported"); |
| 445 | ARM_COMPUTE_RETURN_ERROR_ON_MSG(gemm_info.is_b_reshaped(), "Matrix B already reshaped is not supported"); |
| 446 | |
Georgios Pinitas | 4c5469b | 2019-05-21 13:32:43 +0100 | [diff] [blame] | 447 | int32_t a_offset = a->quantization_info().uniform().offset; |
| 448 | int32_t b_offset = b->quantization_info().uniform().offset; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 449 | |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 450 | const ITensorInfo *matrix_a_info = a; |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 451 | |
giuros01 | 8b6b4a9 | 2018-12-18 19:01:33 +0000 | [diff] [blame] | 452 | TensorInfo tmp_b_info{}; |
| 453 | GEMMRHSMatrixInfo rhs_info; |
giuros01 | 1c9efeb | 2019-01-11 14:04:43 +0000 | [diff] [blame] | 454 | GEMMLHSMatrixInfo lhs_info; |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 455 | |
Gian Marco Iodice | 926afe1 | 2019-03-19 11:44:13 +0000 | [diff] [blame] | 456 | // Get the GPU target |
| 457 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 458 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 459 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 460 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 461 | const unsigned int n = b->dimension(0); |
| 462 | const unsigned int k = a->dimension(0); |
| 463 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 464 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
Gian Marco | 19835e5 | 2018-01-30 13:35:54 +0000 | [diff] [blame] | 465 | |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 466 | bool reshape_matrix_b = is_gemm_reshaped(auto_select_gemm_kernel(auto_heuristics::CommonQuery{ gpu_target, a->data_type(), m, n, k, batch_size }, gemm_info.reshape_b_only_on_first_run())); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 467 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 468 | const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d); |
Georgios Pinitas | ebf6b8a | 2018-09-24 16:31:08 +0100 | [diff] [blame] | 469 | |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 470 | bool convert_to_qasymm8 = is_data_type_quantized_per_channel(b->data_type()) && is_data_type_quantized_symmetric(b->data_type()) |
| 471 | && is_data_type_quantized_asymmetric(a->data_type()); |
| 472 | TensorInfo weights_info(*b); |
| 473 | if(convert_to_qasymm8) |
| 474 | { |
| 475 | b_offset = -128; |
| 476 | weights_info.set_data_type(DataType::QASYMM8); |
| 477 | ARM_COMPUTE_RETURN_ON_ERROR(CLDepthConvertLayerKernel::validate(b, &weights_info, ConvertPolicy::WRAP, 0)); |
| 478 | } |
| 479 | const ITensorInfo *matrix_b_info = &weights_info; |
Gian Marco Iodice | 2ec6c1e | 2019-04-09 12:03:05 +0100 | [diff] [blame] | 480 | if(reshape_matrix_b) |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 481 | { |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 482 | matrix_b_info = &tmp_b_info; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 483 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 484 | // Pick up the GEMM configuration |
SiCong Li | bbd8fac | 2021-02-04 13:12:19 +0000 | [diff] [blame] | 485 | // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails |
| 486 | // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration |
| 487 | const auto res = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }); |
| 488 | lhs_info = res.lhs_info; |
| 489 | rhs_info = res.rhs_info; |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 490 | |
Gian Marco Iodice | 926afe1 | 2019-03-19 11:44:13 +0000 | [diff] [blame] | 491 | // Validate reshape RHS kernel |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 492 | auto_init_if_empty(tmp_b_info, weights_info.clone()->set_tensor_shape(compute_rhs_reshaped_shape(weights_info, rhs_info))); |
| 493 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMReshapeRHSMatrixKernel::validate(&weights_info, &tmp_b_info, rhs_info)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 494 | } |
Isabella Gottardi | c4f582e | 2018-10-11 19:14:55 +0100 | [diff] [blame] | 495 | |
Michalis Spyrou | a4f378d | 2019-04-26 14:54:54 +0100 | [diff] [blame] | 496 | TensorInfo info_vector_sum_col{}; |
| 497 | TensorInfo info_vector_sum_row{}; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 498 | |
Michele Di Giorgio | f64d336 | 2020-04-03 12:40:10 +0100 | [diff] [blame] | 499 | const GEMMLowpReductionKernelInfo reduction_info; |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 500 | // Validate matrix B reduction kernel only if _a_offset is not equal to 0 |
| 501 | if(a_offset != 0) |
| 502 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 503 | info_vector_sum_col = TensorInfo(compute_reductionA_shape(weights_info), 1, DataType::S32); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 504 | |
| 505 | // Configure Matrix B reduction kernel |
Michele Di Giorgio | f64d336 | 2020-04-03 12:40:10 +0100 | [diff] [blame] | 506 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixBReductionKernel::validate(&weights_info, &info_vector_sum_col, reduction_info)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 507 | } |
| 508 | |
| 509 | // Validate Matrix A reduction kernel only if _b_offset is not equal to 0 |
| 510 | if(b_offset != 0) |
| 511 | { |
| 512 | info_vector_sum_row = TensorInfo(compute_reductionB_shape(*a), 1, DataType::S32); |
| 513 | |
| 514 | // Configure matrix A reduction kernel |
Michele Di Giorgio | f64d336 | 2020-04-03 12:40:10 +0100 | [diff] [blame] | 515 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixAReductionKernel::validate(a, &info_vector_sum_row, reduction_info)); |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 516 | } |
| 517 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 518 | GEMMKernelInfo gemm_kernel_info; |
| 519 | gemm_kernel_info.m = m; |
| 520 | gemm_kernel_info.n = n; |
| 521 | gemm_kernel_info.k = k; |
| 522 | gemm_kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 523 | gemm_kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 524 | gemm_kernel_info.lhs_info = lhs_info; |
| 525 | gemm_kernel_info.rhs_info = rhs_info; |
| 526 | gemm_kernel_info.a_offset = a_offset; |
| 527 | gemm_kernel_info.b_offset = b_offset; |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 528 | if(gemm_info.gemmlowp_output_stage().type != GEMMLowpOutputStageType::NONE) |
| 529 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 530 | const size_t num_filters = (gemm_info.gemmlowp_output_stage().is_quantized_per_channel) ? gemm_info.gemmlowp_output_stage().gemmlowp_multipliers.size() : 1; |
| 531 | |
| 532 | const TensorInfo gemm_output_stage_multipliers_shifts_info(TensorInfo(TensorShape(num_filters), 1, DataType::S32)); |
| 533 | |
Manuel Bottini | 959c26d | 2019-12-02 16:22:35 +0000 | [diff] [blame] | 534 | GEMMLowpOutputStageInfo gemmlowp_output_stage = gemm_info.gemmlowp_output_stage(); |
| 535 | gemmlowp_output_stage.output_data_type = a->data_type(); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 536 | |
| 537 | gemm_kernel_info.output_stage = gemmlowp_output_stage; |
| 538 | if(reshape_matrix_b && gemm_info.gemmlowp_output_stage().type == GEMMLowpOutputStageType::QUANTIZE_DOWN_FIXEDPOINT) |
| 539 | { |
| 540 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info, |
| 541 | a_offset == 0 ? nullptr : &info_vector_sum_col, |
| 542 | b_offset == 0 ? nullptr : &info_vector_sum_row, |
| 543 | c, |
| 544 | &gemm_output_stage_multipliers_shifts_info, |
| 545 | &gemm_output_stage_multipliers_shifts_info)); |
| 546 | } |
| 547 | else |
| 548 | { |
| 549 | TensorInfo mm_result_s32_info{}; |
| 550 | |
| 551 | if(reshape_matrix_b) |
| 552 | { |
| 553 | // Output tensor auto inizialitation if not yet initialized |
| 554 | auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, reshape_info)).set_data_type(DataType::S32)); |
| 555 | |
| 556 | // Validate matrix multiply |
| 557 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, gemm_kernel_info)); |
| 558 | } |
| 559 | else |
| 560 | { |
| 561 | // Output tensor auto inizialitation if not yet initialized |
| 562 | auto_init_if_empty(mm_result_s32_info, a->clone()->set_tensor_shape(compute_mm_shape(*matrix_a_info, *matrix_b_info, false, reshape_info)).set_data_type(DataType::S32)); |
| 563 | |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 564 | // Pick up the GEMM configuration |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 565 | // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails |
| 566 | // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration |
| 567 | const auto res = select_default_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }); |
| 568 | lhs_info = res.lhs_info; |
| 569 | rhs_info = res.rhs_info; |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 570 | |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 571 | // Validate matrix multiply |
| 572 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, &mm_result_s32_info, lhs_info, rhs_info, reshape_info)); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 573 | } |
| 574 | |
| 575 | // Validate offset contribution kernel |
| 576 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionOutputStageKernel::validate(&mm_result_s32_info, |
| 577 | a_offset == 0 ? nullptr : &info_vector_sum_col, |
| 578 | b_offset == 0 ? nullptr : &info_vector_sum_row, |
| 579 | c, |
| 580 | output, |
| 581 | a_offset, b_offset, |
| 582 | gemmlowp_output_stage, |
| 583 | &gemm_output_stage_multipliers_shifts_info, |
| 584 | &gemm_output_stage_multipliers_shifts_info)); |
| 585 | } |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 586 | } |
| 587 | else |
| 588 | { |
Gian Marco Iodice | 2ec6c1e | 2019-04-09 12:03:05 +0100 | [diff] [blame] | 589 | if(reshape_matrix_b) |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 590 | { |
| 591 | // Validate matrix multiply |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 592 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyReshapedOnlyRHSKernel::validate(matrix_a_info, matrix_b_info, output, gemm_kernel_info)); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 593 | } |
| 594 | else |
| 595 | { |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 596 | // Pick up the GEMM configuration |
SiCong Li | db35345 | 2021-02-08 15:16:13 +0000 | [diff] [blame] | 597 | // It doesn't matter whether Datatype is DataType::QASYMM8 or DataType::QASYMM8_SIGNED, since it only affect the shape configuration |
| 598 | const auto res = select_default_gemm_config_native(auto_heuristics::CommonQuery{ gpu_target, DataType::QASYMM8, m, n, k, batch_size }); |
| 599 | lhs_info = res.lhs_info; |
| 600 | rhs_info = res.rhs_info; |
Gian Marco Iodice | 06be6f8 | 2019-06-24 17:47:51 +0100 | [diff] [blame] | 601 | |
Gian Marco Iodice | eb65f6d | 2020-04-15 11:42:15 +0100 | [diff] [blame] | 602 | // Validate matrix multiply |
| 603 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpMatrixMultiplyNativeKernel::validate(matrix_a_info, matrix_b_info, output, lhs_info, rhs_info, reshape_info)); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 604 | } |
Gian Marco Iodice | 06be6f8 | 2019-06-24 17:47:51 +0100 | [diff] [blame] | 605 | |
giuros01 | 2f7c149 | 2019-03-18 12:30:02 +0000 | [diff] [blame] | 606 | if(output->total_size() != 0) |
| 607 | { |
| 608 | // Validate offset contribution kernel |
| 609 | ARM_COMPUTE_RETURN_ON_ERROR(CLGEMMLowpOffsetContributionKernel::validate(output, |
| 610 | a_offset == 0 ? nullptr : &info_vector_sum_col, |
| 611 | b_offset == 0 ? nullptr : &info_vector_sum_row, |
| 612 | c, |
| 613 | a_offset, b_offset)); |
| 614 | } |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 615 | } |
Georgios Pinitas | 358ca20 | 2017-12-07 16:47:52 +0000 | [diff] [blame] | 616 | |
| 617 | return Status{}; |
| 618 | } |
| 619 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 620 | void CLGEMMLowpMatrixMultiplyCore::run() |
| 621 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 622 | prepare(); |
| 623 | |
Georgios Pinitas | da953f2 | 2019-04-02 17:27:03 +0100 | [diff] [blame] | 624 | MemoryGroupResourceScope scope_mg(_memory_group); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 625 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 626 | if(_is_gemm_reshaped) |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 627 | { |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 628 | if(!_reshape_b_only_on_first_run) |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 629 | { |
| 630 | // Run reshape matrix B |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 631 | CLScheduler::get().enqueue(*_mtx_b_reshape_kernel, false); |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 632 | } |
| 633 | } |
| 634 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 635 | // Run matrix B reduction kernel only if _a_offset is not equal to 0 |
| 636 | if(_a_offset != 0 && !_reshape_b_only_on_first_run) |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 637 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 638 | CLScheduler::get().enqueue(*_mtx_b_reduction_kernel, false); |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 639 | } |
| 640 | |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 641 | // Run matrix A reduction kernel only if _b_offset is not equal to 0 |
| 642 | if(_b_offset != 0) |
| 643 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 644 | CLScheduler::get().enqueue(*_mtx_a_reduction_kernel, false); |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 645 | } |
| 646 | |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 647 | // Run matrix multiply |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 648 | if(_is_gemm_reshaped) |
| 649 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 650 | CLScheduler::get().enqueue(*_mm_reshaped_only_rhs_kernel, false); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 651 | } |
| 652 | else |
| 653 | { |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 654 | CLScheduler::get().enqueue(*_mm_native_kernel, false); |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 655 | } |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 656 | if(_run_output_stage) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 657 | { |
| 658 | // Run offset contribution/output stage kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 659 | CLScheduler::get().enqueue(*_offset_contribution_output_stage_kernel, true); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 660 | } |
Michele Di Giorgio | b54ba28 | 2020-01-14 15:31:55 +0000 | [diff] [blame] | 661 | if(_run_offset_contribution) |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 662 | { |
| 663 | // Run offset contribution kernel |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 664 | CLScheduler::get().enqueue(*_offset_contribution_kernel, true); |
Gian Marco Iodice | 4b90865 | 2018-10-18 10:21:02 +0100 | [diff] [blame] | 665 | } |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 666 | } |
Chunosov | 5124be5 | 2017-11-22 20:42:13 +0700 | [diff] [blame] | 667 | |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 668 | void CLGEMMLowpMatrixMultiplyCore::prepare() |
| 669 | { |
| 670 | if(!_is_prepared) |
| 671 | { |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 672 | if(_convert_to_qasymm8) |
| 673 | { |
| 674 | _qasymm8_weights.allocator()->allocate(); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 675 | CLScheduler::get().enqueue(*_weights_to_qasymm8, false); |
Vidhya Sudhan Loganathan | 951b8a4 | 2019-11-04 14:42:08 +0000 | [diff] [blame] | 676 | } |
| 677 | |
Gian Marco Iodice | db63b9c | 2019-01-17 09:47:04 +0000 | [diff] [blame] | 678 | if(_is_gemm_reshaped && _reshape_b_only_on_first_run) |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 679 | { |
| 680 | ARM_COMPUTE_ERROR_ON(!_original_b->is_used()); |
| 681 | |
| 682 | // Run reshape kernel and mark original weights tensor as unused |
| 683 | _tmp_b.allocator()->allocate(); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 684 | CLScheduler::get().enqueue(*_mtx_b_reshape_kernel, false); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 685 | _original_b->mark_as_unused(); |
| 686 | } |
| 687 | |
| 688 | // Run matrix B reduction kernel only if _a_offset is not equal to 0 |
| 689 | if(_a_offset != 0 && _reshape_b_only_on_first_run) |
| 690 | { |
| 691 | _vector_sum_col.allocator()->allocate(); |
Sang-Hoon Park | bef7fa2 | 2020-10-21 15:58:54 +0100 | [diff] [blame] | 692 | CLScheduler::get().enqueue(*_mtx_b_reduction_kernel, false); |
Georgios Pinitas | 7221933 | 2018-06-05 14:56:06 +0100 | [diff] [blame] | 693 | } |
| 694 | |
| 695 | CLScheduler::get().queue().finish(); |
| 696 | _is_prepared = true; |
| 697 | } |
Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 698 | } |
Sheri Zhang | 0cdbda5 | 2020-02-25 15:57:21 +0000 | [diff] [blame] | 699 | } // namespace arm_compute |