Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017-2021 Arm Limited. |
| 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 | */ |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 24 | #include "src/gpu/cl/operators/ClGemm.h" |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 27 | #include "arm_compute/core/CL/ICLTensor.h" |
| 28 | #include "arm_compute/core/Error.h" |
| 29 | #include "arm_compute/core/GPUTarget.h" |
| 30 | #include "arm_compute/core/Helpers.h" |
| 31 | #include "arm_compute/core/KernelDescriptors.h" |
| 32 | #include "arm_compute/core/Log.h" |
| 33 | #include "arm_compute/core/TensorInfo.h" |
| 34 | #include "arm_compute/core/Types.h" |
| 35 | #include "arm_compute/core/Utils.h" |
| 36 | #include "arm_compute/core/Validate.h" |
| 37 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 38 | #include "arm_compute/runtime/CL/CLScheduler.h" |
| 39 | #include "arm_compute/runtime/ITensorAllocator.h" |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 40 | |
SiCongLi | 579ca84 | 2021-10-18 09:38:33 +0100 | [diff] [blame] | 41 | #include "arm_compute/core/experimental/IPostOp.h" |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 42 | #include "src/core/helpers/AutoConfiguration.h" |
| 43 | #include "src/core/helpers/MemoryHelpers.h" |
| 44 | #include "src/core/utils/helpers/float_ops.h" |
Georgios Pinitas | 7891a73 | 2021-08-20 21:39:25 +0100 | [diff] [blame] | 45 | #include "src/gpu/cl/IClKernel.h" |
| 46 | #include "src/gpu/cl/utils/ClAuxTensorHandler.h" |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 47 | #include "src/runtime/CL/gemm/CLGEMMKernelSelection.h" |
| 48 | #include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h" |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 49 | |
ramelg01 | 2e53f17 | 2021-09-22 10:48:25 +0100 | [diff] [blame] | 50 | #include "src/common/utils/Log.h" |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 51 | #include "support/Cast.h" |
| 52 | #include "utils/TypePrinter.h" |
| 53 | |
| 54 | namespace arm_compute |
| 55 | { |
| 56 | namespace opencl |
| 57 | { |
| 58 | using namespace arm_compute::misc::shape_calculator; |
| 59 | using namespace arm_compute::cl_gemm; |
| 60 | using namespace arm_compute::experimental; |
| 61 | using namespace arm_compute::utils::cast; |
| 62 | using namespace arm_compute::opencl::kernels; |
| 63 | |
| 64 | namespace |
| 65 | { |
| 66 | inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type) |
| 67 | { |
SiCongLi | 579ca84 | 2021-10-18 09:38:33 +0100 | [diff] [blame] | 68 | return kernel_type == CLGEMMKernelType::NATIVE ? false : true; |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 69 | } |
| 70 | //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 71 | inline CLGEMMKernelType auto_select_gemm_kernel(auto_heuristics::CommonQuery query, bool reshape_b_only_on_first_run, bool constant_weights) |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 72 | { |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 73 | if(!constant_weights) |
| 74 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 75 | return CLGEMMKernelType::NATIVE; |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 76 | } |
| 77 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 78 | auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run); |
| 79 | if(bool(gemm_kernel)) |
| 80 | { |
| 81 | if(validate_gemm_kernel(gemm_kernel.gemm_type)) |
| 82 | { |
| 83 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); |
| 84 | return gemm_kernel.gemm_type; |
| 85 | } |
| 86 | } |
| 87 | gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run); |
| 88 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); |
| 89 | return gemm_kernel.gemm_type; |
| 90 | } |
| 91 | // Validate lhs_info and rhs_info for reshaped only rhs kernel |
| 92 | 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 *c, |
| 93 | const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info) |
| 94 | { |
| 95 | // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel |
| 96 | TensorInfo tmp_b_info{}; |
| 97 | // Validate reshape RHS kernel |
| 98 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 99 | if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info))) |
| 100 | { |
| 101 | return false; |
| 102 | } |
| 103 | // Validate mm kernel |
| 104 | gemm_kernel_info.lhs_info = lhs_info; |
| 105 | gemm_kernel_info.rhs_info = rhs_info; |
| 106 | gemm_kernel_info.has_pad_y = false; |
| 107 | if(!bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info))) |
| 108 | { |
| 109 | return false; |
| 110 | } |
| 111 | gemm_kernel_info.has_pad_y = true; |
| 112 | if(!bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info))) |
| 113 | { |
| 114 | return false; |
| 115 | } |
| 116 | return true; |
| 117 | } |
| 118 | |
| 119 | //Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs |
| 120 | inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a, |
| 121 | const ITensorInfo *b, |
| 122 | const ITensorInfo *c, const ITensorInfo *output) |
| 123 | { |
| 124 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query); |
| 125 | if(config) |
| 126 | { |
| 127 | if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info)) |
| 128 | { |
| 129 | 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()); |
| 130 | return { config.lhs_info, config.rhs_info }; |
| 131 | } |
| 132 | } |
| 133 | config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query); |
| 134 | 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()); |
| 135 | return { config.lhs_info, config.rhs_info }; |
| 136 | } |
| 137 | |
| 138 | // Validate lhs_info and rhs_info for reshaped kernel |
| 139 | inline bool validate_lhs_rhs_info_reshaped(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, |
| 140 | const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info, bool reinterpret_input_as_3d) |
| 141 | { |
| 142 | // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped kernel |
| 143 | TensorInfo tmp_a_info{}; |
| 144 | TensorInfo tmp_b_info{}; |
| 145 | |
| 146 | // Validate reshape LHS kernel |
| 147 | auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, reinterpret_input_as_3d))); |
| 148 | if(!bool(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, reinterpret_input_as_3d))) |
| 149 | { |
| 150 | return false; |
| 151 | } |
| 152 | |
| 153 | // Validate reshape RHS kernel |
| 154 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 155 | if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info))) |
| 156 | { |
| 157 | return false; |
| 158 | } |
| 159 | // Validate mm kernel |
| 160 | gemm_kernel_info.lhs_info = lhs_info; |
| 161 | gemm_kernel_info.rhs_info = rhs_info; |
| 162 | if(!bool(ClGemmMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info))) |
| 163 | { |
| 164 | return false; |
| 165 | } |
| 166 | return true; |
| 167 | } |
| 168 | |
| 169 | //Automatically select between mlgo (prioritized) and default heuristics for reshaped kernel configs |
| 170 | inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a, const ITensorInfo *b, |
| 171 | const ITensorInfo *c, const ITensorInfo *output, bool reinterpret_input_as_3d) |
| 172 | { |
| 173 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped(query); |
| 174 | if(config) |
| 175 | { |
| 176 | if(validate_lhs_rhs_info_reshaped(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info, reinterpret_input_as_3d)) |
| 177 | { |
| 178 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped config from mlgo heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); |
| 179 | return { config.lhs_info, config.rhs_info }; |
| 180 | } |
| 181 | } |
| 182 | config = auto_heuristics::select_default_gemm_config_reshaped(query); |
| 183 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use reshaped config from default heuristics: LHS info: %s ; RHS info: %s ", to_string(config.lhs_info).c_str(), to_string(config.rhs_info).c_str()); |
| 184 | return { config.lhs_info, config.rhs_info }; |
| 185 | } |
| 186 | } // namespace |
| 187 | |
| 188 | ClGemm::ClGemm() |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 189 | : _reshape_lhs_kernel(std::make_unique<ClGemmReshapeLhsMatrixKernel>()), |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 190 | _reshape_rhs_kernel(std::make_unique<ClGemmReshapeRhsMatrixKernel>()), |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 191 | _mm_native_kernel(std::make_unique<ClGemmMatrixMultiplyNativeKernel>()), |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 192 | _mm_reshaped_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedKernel>()), |
| 193 | _mm_reshaped_only_rhs_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedOnlyRhsKernel>()), |
| 194 | _mm_reshaped_only_rhs_fallback_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedOnlyRhsKernel>()), |
| 195 | _tmp_a(), |
| 196 | _tmp_b(), |
| 197 | _reshape_b_only_on_first_run(false), |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 198 | _gemm_kernel_type(CLGEMMKernelType::NATIVE), |
Manuel Bottini | d87aded | 2021-07-16 10:23:31 +0100 | [diff] [blame] | 199 | _is_prepared(false), |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 200 | _aux_mem(AuxTensorIdx::Count) |
| 201 | { |
| 202 | } |
| 203 | |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 204 | void ClGemm::configure_native(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 205 | const GEMMInfo &gemm_info) |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 206 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 207 | DataType data_type = a->data_type(); |
| 208 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 209 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 210 | const unsigned int n = b->dimension(0); |
| 211 | const unsigned int k = a->dimension(0); |
| 212 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 213 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 214 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 215 | bool broadcast_bias = gemm_info.broadcast_bias(); |
| 216 | |
| 217 | GEMMKernelInfo kernel_info; |
| 218 | kernel_info.m = m; |
| 219 | kernel_info.n = n; |
| 220 | kernel_info.k = k; |
| 221 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 222 | kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 223 | kernel_info.broadcast_bias = broadcast_bias; |
| 224 | kernel_info.activation_info = gemm_info.activation_info(); |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 225 | kernel_info.post_ops = gemm_info.post_ops(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 226 | |
| 227 | // Set the target for the kernels |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 228 | _mm_native_kernel->set_target(gpu_target); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 229 | |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 230 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 231 | |
| 232 | // Configure and tune matrix multiply kernel |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 233 | _mm_native_kernel->configure(compile_context, a, b, c, output, alpha, beta, config.lhs_info, config.rhs_info, kernel_info); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 234 | } |
| 235 | |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 236 | void ClGemm::configure_reshaped(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 237 | const GEMMInfo &gemm_info) |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 238 | { |
| 239 | DataType data_type = a->data_type(); |
| 240 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 241 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 242 | const unsigned int n = b->dimension(0); |
| 243 | const unsigned int k = a->dimension(0); |
| 244 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 245 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 246 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 247 | bool broadcast_bias = gemm_info.broadcast_bias(); |
| 248 | |
| 249 | GEMMKernelInfo kernel_info; |
| 250 | kernel_info.m = m; |
| 251 | kernel_info.n = n; |
| 252 | kernel_info.k = k; |
| 253 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 254 | kernel_info.reinterpret_input_as_3d = false; |
| 255 | kernel_info.broadcast_bias = broadcast_bias; |
| 256 | kernel_info.activation_info = gemm_info.activation_info(); |
SiCongLi | 579ca84 | 2021-10-18 09:38:33 +0100 | [diff] [blame] | 257 | kernel_info.post_ops = gemm_info.post_ops(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 258 | |
| 259 | // Set the target for the kernels |
| 260 | _reshape_lhs_kernel->set_target(gpu_target); |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 261 | _mm_reshaped_kernel->set_target(gpu_target); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 262 | |
| 263 | GEMMLHSMatrixInfo lhs_info{}; |
| 264 | GEMMRHSMatrixInfo rhs_info{}; |
| 265 | |
| 266 | // Pick up the GEMM configuration |
| 267 | std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }, kernel_info, a, b, |
| 268 | c, output, gemm_info.reinterpret_input_as_3d()); |
| 269 | |
| 270 | _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d()); |
| 271 | _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info); |
| 272 | |
| 273 | // Configure and tune matrix multiply kernel |
| 274 | _mm_reshaped_kernel->configure(compile_context, &_tmp_a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); |
| 275 | |
| 276 | // Request memory for LHS and RHS reshape matrix |
| 277 | _aux_mem[LhsReshape] = MemoryInfo(offset_int_vec(LhsReshape), MemoryLifetime::Temporary, _tmp_a.total_size()); |
| 278 | _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); |
| 279 | } |
| 280 | |
| 281 | void ClGemm::configure_reshaped_only_rhs(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 282 | const GEMMInfo &gemm_info) |
| 283 | { |
| 284 | DataType data_type = a->data_type(); |
| 285 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 286 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 287 | const unsigned int n = b->dimension(0); |
| 288 | const unsigned int k = a->dimension(0); |
| 289 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 290 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 291 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 292 | bool broadcast_bias = gemm_info.broadcast_bias(); |
| 293 | |
| 294 | GEMMKernelInfo kernel_info; |
| 295 | kernel_info.m = m; |
| 296 | kernel_info.n = n; |
| 297 | kernel_info.k = k; |
| 298 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 299 | kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 300 | kernel_info.broadcast_bias = broadcast_bias; |
| 301 | kernel_info.activation_info = gemm_info.activation_info(); |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 302 | kernel_info.post_ops = gemm_info.post_ops(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 303 | |
| 304 | // Set the target for the kernels |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 305 | _mm_reshaped_only_rhs_kernel->set_target(gpu_target); |
| 306 | _mm_reshaped_only_rhs_fallback_kernel->set_target(gpu_target); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 307 | |
| 308 | GEMMLHSMatrixInfo lhs_info{}; |
| 309 | GEMMRHSMatrixInfo rhs_info{}; |
| 310 | |
| 311 | // Pick up the GEMM configuration |
| 312 | std::tie(lhs_info, rhs_info) = auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }, kernel_info, a, b, c, output); |
| 313 | |
| 314 | // Transpose matrix |
| 315 | _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info); |
| 316 | |
| 317 | // Configure two variants of CLGEMMMatrixMultiplyReshapedOnlyRHSKernel (has_pad_y = false/true) |
| 318 | // During the prepare stage we check the padding requirement for the lhs and dst tensors. If they do not have |
| 319 | // pad y, we dispatch CLGEMMMatrixMultiplyReshapedOnlyRHSKernel with has_pad_y = false |
| 320 | |
| 321 | // Configure matrix multiply kernel with no y padding support |
| 322 | kernel_info.has_pad_y = false; |
| 323 | _mm_reshaped_only_rhs_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); |
| 324 | |
| 325 | // Configure matrix multiply kernel with y padding support |
| 326 | kernel_info.has_pad_y = true; |
| 327 | _mm_reshaped_only_rhs_fallback_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); |
| 328 | |
| 329 | // Request memory for RHS reshape matrix |
| 330 | _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); |
| 331 | } |
| 332 | |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 333 | Status ClGemm::validate_native(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 334 | { |
| 335 | ARM_COMPUTE_UNUSED(alpha); |
| 336 | ARM_COMPUTE_UNUSED(output); |
| 337 | |
| 338 | // Get the GPU target |
| 339 | const GPUTarget gpu_target = CLScheduler::get().target(); |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 340 | DataType data_type = a->data_type(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 341 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 342 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 343 | const unsigned int n = b->dimension(0); |
| 344 | const unsigned int k = a->dimension(0); |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 345 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 346 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 347 | const bool broadcast_bias = gemm_info.broadcast_bias(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 348 | |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 349 | GEMMKernelInfo kernel_info; |
| 350 | kernel_info.m = m; |
| 351 | kernel_info.n = n; |
| 352 | kernel_info.k = k; |
| 353 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 354 | kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 355 | kernel_info.broadcast_bias = broadcast_bias; |
| 356 | kernel_info.activation_info = gemm_info.activation_info(); |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 357 | kernel_info.post_ops = gemm_info.post_ops(); |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 358 | |
| 359 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 360 | |
| 361 | // Validate matrix multiply |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 362 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyNativeKernel::validate(a, b, c, output, alpha, beta, config.lhs_info, config.rhs_info, kernel_info)); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 363 | |
| 364 | return Status{}; |
| 365 | } |
| 366 | |
| 367 | Status ClGemm::validate_reshaped(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 368 | { |
| 369 | ARM_COMPUTE_UNUSED(alpha); |
| 370 | ARM_COMPUTE_UNUSED(output); |
| 371 | |
| 372 | TensorInfo tmp_a_info{}; |
| 373 | TensorInfo tmp_b_info{}; |
| 374 | |
| 375 | // Get the GPU target |
| 376 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 377 | DataType data_type = a->data_type(); |
| 378 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 379 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 380 | const unsigned int n = b->dimension(0); |
| 381 | const unsigned int k = a->dimension(0); |
| 382 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 383 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 384 | const bool broadcast_bias = gemm_info.broadcast_bias(); |
| 385 | |
| 386 | GEMMKernelInfo kernel_info; |
| 387 | kernel_info.m = m; |
| 388 | kernel_info.n = n; |
| 389 | kernel_info.k = k; |
| 390 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 391 | kernel_info.reinterpret_input_as_3d = false; |
| 392 | kernel_info.broadcast_bias = broadcast_bias; |
| 393 | kernel_info.activation_info = gemm_info.activation_info(); |
SiCongLi | 579ca84 | 2021-10-18 09:38:33 +0100 | [diff] [blame] | 394 | kernel_info.post_ops = gemm_info.post_ops(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 395 | |
| 396 | GEMMLHSMatrixInfo lhs_info; |
| 397 | GEMMRHSMatrixInfo rhs_info; |
| 398 | |
| 399 | // Pick up the GEMM configuration |
| 400 | // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails |
| 401 | const auto gemm_config = select_default_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); |
| 402 | lhs_info = gemm_config.lhs_info; |
| 403 | rhs_info = gemm_config.rhs_info; |
| 404 | |
| 405 | auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, gemm_info.reinterpret_input_as_3d()))); |
| 406 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d())); |
| 407 | |
| 408 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 409 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)); |
| 410 | |
| 411 | // Validate matrix multiply |
| 412 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 413 | |
| 414 | return Status{}; |
| 415 | } |
| 416 | |
| 417 | Status ClGemm::validate_reshaped_only_rhs(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 418 | { |
| 419 | ARM_COMPUTE_UNUSED(alpha); |
| 420 | ARM_COMPUTE_UNUSED(output); |
| 421 | |
| 422 | TensorInfo tmp_b_info{}; |
| 423 | |
| 424 | // Get the GPU target |
| 425 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 426 | const DataType data_type = a->data_type(); |
| 427 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 428 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 429 | const unsigned int n = b->dimension(0); |
| 430 | const unsigned int k = a->dimension(0); |
| 431 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 432 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 433 | const bool broadcast_bias = gemm_info.broadcast_bias(); |
| 434 | |
| 435 | GEMMKernelInfo kernel_info; |
| 436 | kernel_info.m = m; |
| 437 | kernel_info.n = n; |
| 438 | kernel_info.k = k; |
| 439 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 440 | kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 441 | kernel_info.broadcast_bias = broadcast_bias; |
| 442 | kernel_info.activation_info = gemm_info.activation_info(); |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 443 | kernel_info.post_ops = gemm_info.post_ops(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 444 | |
| 445 | GEMMLHSMatrixInfo lhs_info; |
| 446 | GEMMRHSMatrixInfo rhs_info; |
| 447 | |
| 448 | // Pick up the GEMM configuration |
| 449 | // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails |
| 450 | const auto gemm_config = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); |
| 451 | lhs_info = gemm_config.lhs_info; |
| 452 | rhs_info = gemm_config.rhs_info; |
| 453 | |
| 454 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 455 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)); |
| 456 | |
| 457 | // Validate matrix multiply |
| 458 | kernel_info.has_pad_y = false; |
| 459 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 460 | |
| 461 | kernel_info.has_pad_y = true; |
| 462 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 463 | |
| 464 | return Status{}; |
| 465 | } |
| 466 | |
| 467 | void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 468 | { |
| 469 | ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); |
| 470 | |
| 471 | // Perform validation step |
| 472 | ARM_COMPUTE_ERROR_THROW_ON(validate(a, b, c, output, alpha, beta, gemm_info)); |
ramelg01 | 2e53f17 | 2021-09-22 10:48:25 +0100 | [diff] [blame] | 473 | ARM_COMPUTE_LOG_PARAMS(a, b, c, output, alpha, beta, gemm_info); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 474 | |
| 475 | // Check if we need to reshape the matrix B only on the first run |
| 476 | _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
Georgios Pinitas | f5d51f3 | 2021-08-17 16:09:10 +0100 | [diff] [blame] | 477 | _is_prepared = gemm_info.retain_internal_weights(); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 478 | |
| 479 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 480 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 481 | const unsigned int n = b->dimension(0); |
| 482 | const unsigned int k = a->dimension(0); |
| 483 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 484 | |
| 485 | // Select GEMMType |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 486 | _gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery{ CLScheduler::get().target(), a->data_type(), m, n, k, batch_size }, _reshape_b_only_on_first_run, |
Giorgio Arena | 63e0beb | 2021-09-24 14:04:27 +0100 | [diff] [blame] | 487 | b->are_values_constant()); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 488 | |
| 489 | const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); |
| 490 | |
| 491 | ITensorInfo *c_to_use = fuse_add_c ? c : nullptr; |
| 492 | |
| 493 | switch(_gemm_kernel_type) |
| 494 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 495 | case CLGEMMKernelType::NATIVE: |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 496 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 497 | configure_native(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 498 | break; |
| 499 | } |
| 500 | case CLGEMMKernelType::RESHAPED: |
| 501 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 502 | configure_reshaped(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 503 | break; |
| 504 | } |
| 505 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 506 | { |
| 507 | configure_reshaped_only_rhs(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
| 508 | break; |
| 509 | } |
| 510 | default: |
| 511 | { |
| 512 | ARM_COMPUTE_ERROR("GEMMType not supported"); |
| 513 | } |
| 514 | } |
| 515 | } |
| 516 | |
| 517 | Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 518 | { |
| 519 | // Get the GPU target |
| 520 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 521 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 522 | const unsigned int n = b->dimension(0); |
| 523 | const unsigned int k = a->dimension(0); |
| 524 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 525 | |
| 526 | // Select GEMMType |
| 527 | CLGEMMKernelType gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery |
| 528 | { |
| 529 | CLScheduler::get().target(), a->data_type(), m, n, k, batch_size, |
| 530 | }, |
Giorgio Arena | 63e0beb | 2021-09-24 14:04:27 +0100 | [diff] [blame] | 531 | gemm_info.reshape_b_only_on_first_run(), b->are_values_constant()); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 532 | |
| 533 | const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); |
| 534 | |
| 535 | const ITensorInfo *c_to_use = fuse_add_c ? c : nullptr; |
| 536 | |
| 537 | switch(gemm_kernel_type) |
| 538 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 539 | case CLGEMMKernelType::NATIVE: |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 540 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 541 | ARM_COMPUTE_RETURN_ON_ERROR(validate_native(a, b, c_to_use, output, alpha, beta, gemm_info)); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 542 | break; |
| 543 | } |
| 544 | case CLGEMMKernelType::RESHAPED: |
| 545 | { |
| 546 | ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped(a, b, c_to_use, output, alpha, beta, gemm_info)); |
| 547 | break; |
| 548 | } |
| 549 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 550 | { |
| 551 | ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_only_rhs(a, b, c_to_use, output, alpha, beta, gemm_info)); |
| 552 | break; |
| 553 | } |
| 554 | default: |
| 555 | { |
| 556 | ARM_COMPUTE_RETURN_ERROR_MSG("GEMMType not supported"); |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | return Status{}; |
| 561 | } |
| 562 | |
| 563 | void ClGemm::run(ITensorPack &tensors) |
| 564 | { |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 565 | const ITensor *lhs = tensors.get_const_tensor(ACL_SRC_0); |
| 566 | const ITensor *rhs = tensors.get_const_tensor(ACL_SRC_1); |
| 567 | ITensor *dst = tensors.get_tensor(ACL_DST); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 568 | |
| 569 | ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, dst); |
| 570 | |
| 571 | CLAuxTensorHandler lhs_reshaped(offset_int_vec(LhsReshape), _tmp_a, tensors, true); |
| 572 | CLAuxTensorHandler rhs_reshaped(offset_int_vec(RhsReshape), _tmp_b, tensors, true); |
| 573 | |
| 574 | // Prepare the consts if needed |
| 575 | prepare(tensors); |
| 576 | |
| 577 | // Run matrix multiply kernel |
| 578 | switch(_gemm_kernel_type) |
| 579 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 580 | case CLGEMMKernelType::NATIVE: |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 581 | { |
Gian Marco Iodice | c9cecc0 | 2021-10-15 10:23:24 +0100 | [diff] [blame] | 582 | CLScheduler::get().enqueue_op(*_mm_native_kernel, tensors, true); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 583 | break; |
| 584 | } |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 585 | case CLGEMMKernelType::RESHAPED: |
| 586 | { |
| 587 | // Run interleave kernel |
| 588 | ITensorPack reshape_lhs_pack{ { ACL_SRC, lhs }, { ACL_DST, lhs_reshaped.get() } }; |
| 589 | CLScheduler::get().enqueue_op(*_reshape_lhs_kernel, reshape_lhs_pack, false); |
| 590 | |
| 591 | if(!_reshape_b_only_on_first_run) |
| 592 | { |
| 593 | // Run transpose kernel |
| 594 | ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } }; |
| 595 | CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false); |
| 596 | } |
SiCongLi | 579ca84 | 2021-10-18 09:38:33 +0100 | [diff] [blame] | 597 | // Copy original tensor pack and overwrite lhs and rhs with reshaped counterparts |
| 598 | ITensorPack gemm_reshaped_pack(tensors); |
| 599 | gemm_reshaped_pack.add_const_tensor(ACL_SRC_0, lhs_reshaped.get()); |
| 600 | gemm_reshaped_pack.add_const_tensor(ACL_SRC_1, rhs_reshaped.get()); |
Manuel Bottini | d87aded | 2021-07-16 10:23:31 +0100 | [diff] [blame] | 601 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 602 | if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED) |
| 603 | { |
| 604 | CLScheduler::get().enqueue_op(*_mm_reshaped_kernel, gemm_reshaped_pack, true); |
| 605 | } |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 606 | break; |
| 607 | } |
| 608 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 609 | { |
| 610 | if(!_reshape_b_only_on_first_run) |
| 611 | { |
| 612 | // Run transpose kernel |
| 613 | ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } }; |
| 614 | CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false); |
| 615 | } |
| 616 | // In case of RESHAPED_ONLY_RHS, we need to check the padding requirement |
| 617 | // Check if the lhs or dst tensors have padding |
| 618 | const unsigned int cross_plane_pad_lhs = lhs->info()->padding().top + lhs->info()->padding().bottom; |
| 619 | const unsigned int cross_plane_pad_dst = dst->info()->padding().top + dst->info()->padding().bottom; |
| 620 | bool has_pad_y = (cross_plane_pad_lhs != 0) || (cross_plane_pad_dst != 0); |
| 621 | |
SiCongLi | afa1972 | 2021-10-24 19:12:33 +0100 | [diff] [blame] | 622 | // Copy original tensor pack and overwrite rhs with reshaped counterpart |
| 623 | ITensorPack gemm_reshaped_onlyrhs_pack(tensors); |
| 624 | gemm_reshaped_onlyrhs_pack.add_const_tensor(ACL_SRC_1, rhs_reshaped.get()); |
| 625 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 626 | if(has_pad_y) |
| 627 | { |
| 628 | CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_fallback_kernel, gemm_reshaped_onlyrhs_pack, true); |
| 629 | } |
| 630 | else |
| 631 | { |
| 632 | CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_kernel, gemm_reshaped_onlyrhs_pack, true); |
| 633 | } |
| 634 | break; |
| 635 | } |
| 636 | default: |
| 637 | { |
| 638 | ARM_COMPUTE_ERROR("GEMMType not supported"); |
| 639 | } |
| 640 | } |
| 641 | } |
| 642 | |
| 643 | void ClGemm::prepare(ITensorPack &constants) |
| 644 | { |
Manuel Bottini | d87aded | 2021-07-16 10:23:31 +0100 | [diff] [blame] | 645 | if(!_is_prepared) |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 646 | { |
Manuel Bottini | d87aded | 2021-07-16 10:23:31 +0100 | [diff] [blame] | 647 | const ITensor *src1 = constants.get_const_tensor(ACL_SRC_1); |
| 648 | ICLTensor *rhs_aux = utils::cast::polymorphic_downcast<ICLTensor *>(constants.get_tensor(offset_int_vec(RhsReshape))); |
Georgios Pinitas | 2b147ee | 2021-07-08 18:14:45 +0100 | [diff] [blame] | 649 | |
Manuel Bottini | d87aded | 2021-07-16 10:23:31 +0100 | [diff] [blame] | 650 | // If memory for RHS is persistent and src1 is provided re-transform else assume that RHS is transformed |
| 651 | if((_aux_mem[AuxTensorIdx::RhsReshape].lifetime == MemoryLifetime::Persistent) && (src1 != nullptr && rhs_aux != nullptr) && rhs_aux) |
| 652 | { |
| 653 | ARM_COMPUTE_LOG_INFO_WITH_FUNCNAME_ACL("Transforming RHS Matrix!"); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 654 | |
Manuel Bottini | d87aded | 2021-07-16 10:23:31 +0100 | [diff] [blame] | 655 | CLAuxTensorHandler rhs_reshaped(_tmp_b, *rhs_aux); |
| 656 | ARM_COMPUTE_ERROR_ON(rhs_reshaped.get()->cl_buffer().get() == nullptr); |
| 657 | |
| 658 | ITensorPack reshape_rhs_pack{ { ACL_SRC, src1 }, { ACL_DST, rhs_reshaped.get() } }; |
| 659 | CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, true); |
| 660 | } |
| 661 | _is_prepared = true; |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 662 | } |
| 663 | } |
| 664 | |
| 665 | experimental::MemoryRequirements ClGemm::workspace() const |
| 666 | { |
| 667 | return _aux_mem; |
| 668 | } |
| 669 | } // namespace opencl |
| 670 | } // namespace arm_compute |