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 | */ |
| 24 | #include "src/runtime/gpu/cl/operators/ClGemm.h" |
| 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" |
| 40 | #include "src/core/gpu/cl/IClKernel.h" |
| 41 | #include "src/core/helpers/AutoConfiguration.h" |
| 42 | #include "src/core/helpers/MemoryHelpers.h" |
| 43 | #include "src/core/utils/helpers/float_ops.h" |
| 44 | #include "src/runtime/CL/gemm/CLGEMMKernelSelection.h" |
| 45 | #include "src/runtime/CL/gemm_auto_heuristics/CLGEMMAutoHeuristics.h" |
| 46 | #include "src/runtime/gpu/cl/utils/ClAuxTensorHandler.h" |
| 47 | |
| 48 | #include "support/Cast.h" |
| 49 | #include "utils/TypePrinter.h" |
| 50 | |
| 51 | namespace arm_compute |
| 52 | { |
| 53 | namespace opencl |
| 54 | { |
| 55 | using namespace arm_compute::misc::shape_calculator; |
| 56 | using namespace arm_compute::cl_gemm; |
| 57 | using namespace arm_compute::experimental; |
| 58 | using namespace arm_compute::utils::cast; |
| 59 | using namespace arm_compute::opencl::kernels; |
| 60 | |
| 61 | namespace |
| 62 | { |
| 63 | inline bool validate_gemm_kernel(CLGEMMKernelType kernel_type) |
| 64 | { |
| 65 | switch(kernel_type) |
| 66 | { |
| 67 | case CLGEMMKernelType::NATIVE_V1: |
| 68 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 69 | case CLGEMMKernelType::RESHAPED_V1: |
| 70 | case CLGEMMKernelType::RESHAPED: |
| 71 | { |
| 72 | return true; |
| 73 | } |
| 74 | default: |
| 75 | { |
| 76 | return false; |
| 77 | } |
| 78 | } |
| 79 | } |
| 80 | //Automatically select between mlgo (prioritized) and default heuristics for gemm kernel type |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 81 | 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] | 82 | { |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 83 | if(!constant_weights) |
| 84 | { |
| 85 | return CLGEMMKernelType::NATIVE_V1; |
| 86 | } |
| 87 | |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 88 | auto gemm_kernel = auto_heuristics::select_mlgo_gemm_kernel(query, reshape_b_only_on_first_run); |
| 89 | if(bool(gemm_kernel)) |
| 90 | { |
| 91 | if(validate_gemm_kernel(gemm_kernel.gemm_type)) |
| 92 | { |
| 93 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from mlgo heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); |
| 94 | return gemm_kernel.gemm_type; |
| 95 | } |
| 96 | } |
| 97 | gemm_kernel = auto_heuristics::select_default_gemm_kernel(query, reshape_b_only_on_first_run); |
| 98 | ARM_COMPUTE_LOG_INFO_MSG_WITH_FORMAT_CORE("Use gemm kernel from default heuristics: %s.", to_string(gemm_kernel.gemm_type).c_str()); |
| 99 | return gemm_kernel.gemm_type; |
| 100 | } |
| 101 | // Validate lhs_info and rhs_info for reshaped only rhs kernel |
| 102 | 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, |
| 103 | const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info) |
| 104 | { |
| 105 | // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped only rhs kernel |
| 106 | TensorInfo tmp_b_info{}; |
| 107 | // Validate reshape RHS kernel |
| 108 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 109 | if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info))) |
| 110 | { |
| 111 | return false; |
| 112 | } |
| 113 | // Validate mm kernel |
| 114 | gemm_kernel_info.lhs_info = lhs_info; |
| 115 | gemm_kernel_info.rhs_info = rhs_info; |
| 116 | gemm_kernel_info.has_pad_y = false; |
| 117 | if(!bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info))) |
| 118 | { |
| 119 | return false; |
| 120 | } |
| 121 | gemm_kernel_info.has_pad_y = true; |
| 122 | if(!bool(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info))) |
| 123 | { |
| 124 | return false; |
| 125 | } |
| 126 | return true; |
| 127 | } |
| 128 | |
| 129 | //Automatically select between mlgo (prioritized) and default heuristics for reshaped only rhs kernel configs |
| 130 | inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a, |
| 131 | const ITensorInfo *b, |
| 132 | const ITensorInfo *c, const ITensorInfo *output) |
| 133 | { |
| 134 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped_only_rhs(query); |
| 135 | if(config) |
| 136 | { |
| 137 | if(validate_lhs_rhs_info_reshaped_only_rhs(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info)) |
| 138 | { |
| 139 | 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()); |
| 140 | return { config.lhs_info, config.rhs_info }; |
| 141 | } |
| 142 | } |
| 143 | config = auto_heuristics::select_default_gemm_config_reshaped_only_rhs(query); |
| 144 | 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()); |
| 145 | return { config.lhs_info, config.rhs_info }; |
| 146 | } |
| 147 | |
| 148 | // Validate lhs_info and rhs_info for reshaped kernel |
| 149 | inline bool validate_lhs_rhs_info_reshaped(const GEMMLHSMatrixInfo &lhs_info, const GEMMRHSMatrixInfo &rhs_info, const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, |
| 150 | const ITensorInfo *output, GEMMKernelInfo gemm_kernel_info, bool reinterpret_input_as_3d) |
| 151 | { |
| 152 | // Validate GEMMLHSMatrixInfo and GEMMRHSMatrixInfo for reshaped kernel |
| 153 | TensorInfo tmp_a_info{}; |
| 154 | TensorInfo tmp_b_info{}; |
| 155 | |
| 156 | // Validate reshape LHS kernel |
| 157 | auto_init_if_empty(tmp_a_info, a->clone()->set_tensor_shape(compute_lhs_reshaped_shape(*a, lhs_info, reinterpret_input_as_3d))); |
| 158 | if(!bool(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, reinterpret_input_as_3d))) |
| 159 | { |
| 160 | return false; |
| 161 | } |
| 162 | |
| 163 | // Validate reshape RHS kernel |
| 164 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 165 | if(!bool(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info))) |
| 166 | { |
| 167 | return false; |
| 168 | } |
| 169 | // Validate mm kernel |
| 170 | gemm_kernel_info.lhs_info = lhs_info; |
| 171 | gemm_kernel_info.rhs_info = rhs_info; |
| 172 | if(!bool(ClGemmMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, 1.f, 0.f, lhs_info, rhs_info, gemm_kernel_info))) |
| 173 | { |
| 174 | return false; |
| 175 | } |
| 176 | return true; |
| 177 | } |
| 178 | |
| 179 | //Automatically select between mlgo (prioritized) and default heuristics for reshaped kernel configs |
| 180 | inline std::pair<GEMMLHSMatrixInfo, GEMMRHSMatrixInfo> auto_select_gemm_config_reshaped(auto_heuristics::CommonQuery query, GEMMKernelInfo kernel_info, const ITensorInfo *a, const ITensorInfo *b, |
| 181 | const ITensorInfo *c, const ITensorInfo *output, bool reinterpret_input_as_3d) |
| 182 | { |
| 183 | auto config = auto_heuristics::select_mlgo_gemm_config_reshaped(query); |
| 184 | if(config) |
| 185 | { |
| 186 | if(validate_lhs_rhs_info_reshaped(config.lhs_info, config.rhs_info, a, b, c, output, kernel_info, reinterpret_input_as_3d)) |
| 187 | { |
| 188 | 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()); |
| 189 | return { config.lhs_info, config.rhs_info }; |
| 190 | } |
| 191 | } |
| 192 | config = auto_heuristics::select_default_gemm_config_reshaped(query); |
| 193 | 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()); |
| 194 | return { config.lhs_info, config.rhs_info }; |
| 195 | } |
| 196 | } // namespace |
| 197 | |
| 198 | ClGemm::ClGemm() |
| 199 | : _mm_kernel(std::make_unique<ClGemmMatrixMultiplyKernel>()), |
| 200 | _reshape_lhs_kernel(std::make_unique<ClGemmReshapeLhsMatrixKernel>()), |
| 201 | _reshape_rhs_kernel(std::make_unique<ClGemmReshapeRhsMatrixKernel>()), |
| 202 | _mm_reshaped_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedKernel>()), |
| 203 | _mm_reshaped_only_rhs_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedOnlyRhsKernel>()), |
| 204 | _mm_reshaped_only_rhs_fallback_kernel(std::make_unique<ClGemmMatrixMultiplyReshapedOnlyRhsKernel>()), |
| 205 | _tmp_a(), |
| 206 | _tmp_b(), |
| 207 | _reshape_b_only_on_first_run(false), |
| 208 | _gemm_kernel_type(CLGEMMKernelType::NATIVE_V1), |
| 209 | _aux_mem(AuxTensorIdx::Count) |
| 210 | { |
| 211 | } |
| 212 | |
| 213 | void ClGemm::configure_native_v1(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 214 | const GEMMInfo &gemm_info) |
| 215 | { |
| 216 | const unsigned int m = gemm_info.reinterpret_input_as_3d() ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 217 | const unsigned int n = b->dimension(0); |
| 218 | const unsigned int k = a->dimension(0); |
| 219 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 220 | |
| 221 | // Set the target for the kernels |
| 222 | _mm_kernel->set_target(gpu_target); |
| 223 | |
| 224 | GEMMReshapeInfo reshape_info(m, n, k, 1, 1, gemm_info.depth_output_gemm3d(), gemm_info.reinterpret_input_as_3d(), gemm_info.broadcast_bias()); |
| 225 | |
| 226 | // Configure and tune matrix multiply kernel |
| 227 | _mm_kernel->configure(compile_context, a, b, c, output, alpha, beta, false, reshape_info, gemm_info.fp_mixed_precision(), gemm_info.activation_info()); |
| 228 | |
| 229 | // Tune kernel statically |
| 230 | CLScheduler::get().tune_kernel_static(*_mm_kernel); |
| 231 | } |
| 232 | |
| 233 | void ClGemm::configure_reshaped_v1(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 234 | const GEMMInfo &gemm_info) |
| 235 | { |
| 236 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 237 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 238 | const unsigned int n = b->dimension(0); |
| 239 | const unsigned int k = a->dimension(0); |
| 240 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 241 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 242 | int mult_transpose1xW_width = 1; |
| 243 | int mult_interleave4x4_height = 1; |
| 244 | |
| 245 | // Set the target for the kernels |
| 246 | _reshape_lhs_kernel->set_target(gpu_target); |
| 247 | _mm_kernel->set_target(gpu_target); |
| 248 | |
| 249 | if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST) |
| 250 | { |
| 251 | mult_transpose1xW_width = 4; |
| 252 | mult_interleave4x4_height = 2; |
| 253 | } |
| 254 | |
| 255 | GEMMRHSMatrixInfo rhs_info; |
| 256 | rhs_info.n0 = 16 / b->element_size(); |
| 257 | rhs_info.k0 = 1; |
| 258 | rhs_info.h0 = mult_transpose1xW_width; |
| 259 | rhs_info.interleave = false; |
| 260 | rhs_info.transpose = false; |
| 261 | |
| 262 | GEMMLHSMatrixInfo lhs_info; |
| 263 | lhs_info.m0 = 4; |
| 264 | lhs_info.k0 = 4; |
| 265 | lhs_info.v0 = mult_interleave4x4_height; |
| 266 | lhs_info.interleave = true; |
| 267 | lhs_info.transpose = true; |
| 268 | |
| 269 | GEMMReshapeInfo reshape_info(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, false, gemm_info.broadcast_bias()); |
| 270 | |
| 271 | // Configure interleave kernel |
| 272 | _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, reinterpret_input_as_3d); |
| 273 | |
| 274 | // Configure transpose kernel |
| 275 | _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info); |
| 276 | |
| 277 | // Configure and tune matrix multiply kernel |
| 278 | _mm_kernel->configure(compile_context, &_tmp_a, &_tmp_b, c, output, alpha, beta, true, reshape_info, gemm_info.fp_mixed_precision(), gemm_info.activation_info()); |
| 279 | |
| 280 | CLScheduler::get().tune_kernel_static(*_mm_kernel); |
| 281 | |
| 282 | // Request memory for LHS and RHS reshape matrix |
| 283 | _aux_mem[LhsReshape] = MemoryInfo(offset_int_vec(LhsReshape), MemoryLifetime::Temporary, _tmp_a.total_size()); |
| 284 | _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); |
| 285 | } |
| 286 | |
| 287 | void ClGemm::configure_reshaped_v2(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 288 | const GEMMInfo &gemm_info) |
| 289 | { |
| 290 | DataType data_type = a->data_type(); |
| 291 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 292 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 293 | const unsigned int n = b->dimension(0); |
| 294 | const unsigned int k = a->dimension(0); |
| 295 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 296 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 297 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 298 | bool broadcast_bias = gemm_info.broadcast_bias(); |
| 299 | |
| 300 | GEMMKernelInfo kernel_info; |
| 301 | kernel_info.m = m; |
| 302 | kernel_info.n = n; |
| 303 | kernel_info.k = k; |
| 304 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 305 | kernel_info.reinterpret_input_as_3d = false; |
| 306 | kernel_info.broadcast_bias = broadcast_bias; |
| 307 | kernel_info.activation_info = gemm_info.activation_info(); |
| 308 | |
| 309 | // Set the target for the kernels |
| 310 | _reshape_lhs_kernel->set_target(gpu_target); |
| 311 | _mm_kernel->set_target(gpu_target); |
| 312 | |
| 313 | GEMMLHSMatrixInfo lhs_info{}; |
| 314 | GEMMRHSMatrixInfo rhs_info{}; |
| 315 | |
| 316 | // Pick up the GEMM configuration |
| 317 | 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, |
| 318 | c, output, gemm_info.reinterpret_input_as_3d()); |
| 319 | |
| 320 | _reshape_lhs_kernel->configure(compile_context, a, &_tmp_a, lhs_info, gemm_info.reinterpret_input_as_3d()); |
| 321 | _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info); |
| 322 | |
| 323 | // Configure and tune matrix multiply kernel |
| 324 | _mm_reshaped_kernel->configure(compile_context, &_tmp_a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); |
| 325 | |
| 326 | // Request memory for LHS and RHS reshape matrix |
| 327 | _aux_mem[LhsReshape] = MemoryInfo(offset_int_vec(LhsReshape), MemoryLifetime::Temporary, _tmp_a.total_size()); |
| 328 | _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); |
| 329 | } |
| 330 | |
| 331 | void ClGemm::configure_reshaped_only_rhs(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, |
| 332 | const GEMMInfo &gemm_info) |
| 333 | { |
| 334 | DataType data_type = a->data_type(); |
| 335 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 336 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 337 | const unsigned int n = b->dimension(0); |
| 338 | const unsigned int k = a->dimension(0); |
| 339 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 340 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 341 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 342 | bool broadcast_bias = gemm_info.broadcast_bias(); |
| 343 | |
| 344 | GEMMKernelInfo kernel_info; |
| 345 | kernel_info.m = m; |
| 346 | kernel_info.n = n; |
| 347 | kernel_info.k = k; |
| 348 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 349 | kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 350 | kernel_info.broadcast_bias = broadcast_bias; |
| 351 | kernel_info.activation_info = gemm_info.activation_info(); |
| 352 | |
| 353 | // Set the target for the kernels |
| 354 | _mm_kernel->set_target(gpu_target); |
| 355 | |
| 356 | GEMMLHSMatrixInfo lhs_info{}; |
| 357 | GEMMRHSMatrixInfo rhs_info{}; |
| 358 | |
| 359 | // Pick up the GEMM configuration |
| 360 | 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); |
| 361 | |
| 362 | // Transpose matrix |
| 363 | _reshape_rhs_kernel->configure(compile_context, b, &_tmp_b, rhs_info); |
| 364 | |
| 365 | // Configure two variants of CLGEMMMatrixMultiplyReshapedOnlyRHSKernel (has_pad_y = false/true) |
| 366 | // During the prepare stage we check the padding requirement for the lhs and dst tensors. If they do not have |
| 367 | // pad y, we dispatch CLGEMMMatrixMultiplyReshapedOnlyRHSKernel with has_pad_y = false |
| 368 | |
| 369 | // Configure matrix multiply kernel with no y padding support |
| 370 | kernel_info.has_pad_y = false; |
| 371 | _mm_reshaped_only_rhs_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); |
| 372 | |
| 373 | // Configure matrix multiply kernel with y padding support |
| 374 | kernel_info.has_pad_y = true; |
| 375 | _mm_reshaped_only_rhs_fallback_kernel->configure(compile_context, a, &_tmp_b, c, output, alpha, beta, lhs_info, rhs_info, kernel_info); |
| 376 | |
| 377 | // Request memory for RHS reshape matrix |
| 378 | _aux_mem[RhsReshape] = MemoryInfo(offset_int_vec(RhsReshape), _reshape_b_only_on_first_run ? MemoryLifetime::Persistent : MemoryLifetime::Temporary, _tmp_b.total_size()); |
| 379 | } |
| 380 | |
| 381 | Status ClGemm::validate_native_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 382 | { |
| 383 | ARM_COMPUTE_UNUSED(alpha); |
| 384 | ARM_COMPUTE_UNUSED(output); |
| 385 | |
| 386 | // Get the GPU target |
| 387 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 388 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 389 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 390 | const unsigned int n = b->dimension(0); |
| 391 | const unsigned int k = a->dimension(0); |
| 392 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 393 | |
| 394 | const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, 1, 1, depth_output_gemm3d, reinterpret_input_as_3d, gemm_info.broadcast_bias()); |
| 395 | |
| 396 | // Validate matrix multiply |
| 397 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyKernel::validate(a, b, c, output, alpha, beta, |
| 398 | false, reshape_info, gpu_target, gemm_info.fp_mixed_precision(), gemm_info.activation_info())); |
| 399 | |
| 400 | return Status{}; |
| 401 | } |
| 402 | |
| 403 | Status ClGemm::validate_reshaped_v1(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 404 | { |
| 405 | ARM_COMPUTE_UNUSED(alpha); |
| 406 | ARM_COMPUTE_UNUSED(output); |
| 407 | |
| 408 | TensorInfo tmp_a_info{}; |
| 409 | TensorInfo tmp_b_info{}; |
| 410 | |
| 411 | // Get the GPU target |
| 412 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 413 | const unsigned int m = gemm_info.reinterpret_input_as_3d() ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 414 | const unsigned int n = b->dimension(0); |
| 415 | const unsigned int k = a->dimension(0); |
| 416 | int mult_transpose1xW_width = 1; |
| 417 | int mult_interleave4x4_height = 1; |
| 418 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 419 | |
| 420 | if(get_arch_from_target(gpu_target) == GPUTarget::BIFROST) |
| 421 | { |
| 422 | mult_transpose1xW_width = 4; |
| 423 | mult_interleave4x4_height = 2; |
| 424 | } |
| 425 | |
| 426 | GEMMRHSMatrixInfo rhs_info; |
| 427 | rhs_info.n0 = 16 / b->element_size(); |
| 428 | rhs_info.k0 = 1; |
| 429 | rhs_info.h0 = mult_transpose1xW_width; |
| 430 | rhs_info.interleave = false; |
| 431 | rhs_info.transpose = false; |
| 432 | |
| 433 | GEMMLHSMatrixInfo lhs_info; |
| 434 | lhs_info.m0 = 4; |
| 435 | lhs_info.k0 = 4; |
| 436 | lhs_info.v0 = mult_interleave4x4_height; |
| 437 | lhs_info.interleave = true; |
| 438 | lhs_info.transpose = true; |
| 439 | |
| 440 | const GEMMReshapeInfo reshape_info = GEMMReshapeInfo(m, n, k, mult_transpose1xW_width, mult_interleave4x4_height, depth_output_gemm3d, false, gemm_info.broadcast_bias()); |
| 441 | |
| 442 | // Validate interleave kernel |
| 443 | 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()))); |
| 444 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d())); |
| 445 | |
| 446 | // Validate transpose kernel |
| 447 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 448 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)); |
| 449 | |
| 450 | // Validate matrix multiply |
| 451 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyKernel::validate(&tmp_a_info, &tmp_b_info, c, output, alpha, beta, |
| 452 | true, reshape_info, gpu_target, gemm_info.fp_mixed_precision(), gemm_info.activation_info())); |
| 453 | |
| 454 | return Status{}; |
| 455 | } |
| 456 | |
| 457 | Status ClGemm::validate_reshaped(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 458 | { |
| 459 | ARM_COMPUTE_UNUSED(alpha); |
| 460 | ARM_COMPUTE_UNUSED(output); |
| 461 | |
| 462 | TensorInfo tmp_a_info{}; |
| 463 | TensorInfo tmp_b_info{}; |
| 464 | |
| 465 | // Get the GPU target |
| 466 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 467 | DataType data_type = a->data_type(); |
| 468 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 469 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 470 | const unsigned int n = b->dimension(0); |
| 471 | const unsigned int k = a->dimension(0); |
| 472 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 473 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 474 | const bool broadcast_bias = gemm_info.broadcast_bias(); |
| 475 | |
| 476 | GEMMKernelInfo kernel_info; |
| 477 | kernel_info.m = m; |
| 478 | kernel_info.n = n; |
| 479 | kernel_info.k = k; |
| 480 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 481 | kernel_info.reinterpret_input_as_3d = false; |
| 482 | kernel_info.broadcast_bias = broadcast_bias; |
| 483 | kernel_info.activation_info = gemm_info.activation_info(); |
| 484 | |
| 485 | GEMMLHSMatrixInfo lhs_info; |
| 486 | GEMMRHSMatrixInfo rhs_info; |
| 487 | |
| 488 | // Pick up the GEMM configuration |
| 489 | // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails |
| 490 | const auto gemm_config = select_default_gemm_config_reshaped(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); |
| 491 | lhs_info = gemm_config.lhs_info; |
| 492 | rhs_info = gemm_config.rhs_info; |
| 493 | |
| 494 | 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()))); |
| 495 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeLhsMatrixKernel::validate(a, &tmp_a_info, lhs_info, gemm_info.reinterpret_input_as_3d())); |
| 496 | |
| 497 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 498 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)); |
| 499 | |
| 500 | // Validate matrix multiply |
| 501 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedKernel::validate(&tmp_a_info, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 502 | |
| 503 | return Status{}; |
| 504 | } |
| 505 | |
| 506 | 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) |
| 507 | { |
| 508 | ARM_COMPUTE_UNUSED(alpha); |
| 509 | ARM_COMPUTE_UNUSED(output); |
| 510 | |
| 511 | TensorInfo tmp_b_info{}; |
| 512 | |
| 513 | // Get the GPU target |
| 514 | const GPUTarget gpu_target = CLScheduler::get().target(); |
| 515 | const DataType data_type = a->data_type(); |
| 516 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 517 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 518 | const unsigned int n = b->dimension(0); |
| 519 | const unsigned int k = a->dimension(0); |
| 520 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 521 | const int depth_output_gemm3d = gemm_info.depth_output_gemm3d(); |
| 522 | const bool broadcast_bias = gemm_info.broadcast_bias(); |
| 523 | |
| 524 | GEMMKernelInfo kernel_info; |
| 525 | kernel_info.m = m; |
| 526 | kernel_info.n = n; |
| 527 | kernel_info.k = k; |
| 528 | kernel_info.depth_output_gemm3d = depth_output_gemm3d; |
| 529 | kernel_info.reinterpret_input_as_3d = reinterpret_input_as_3d; |
| 530 | kernel_info.broadcast_bias = broadcast_bias; |
| 531 | kernel_info.activation_info = gemm_info.activation_info(); |
| 532 | |
| 533 | GEMMLHSMatrixInfo lhs_info; |
| 534 | GEMMRHSMatrixInfo rhs_info; |
| 535 | |
| 536 | // Pick up the GEMM configuration |
| 537 | // NOTE: No need to validate mlgo configurations as they automatically fall back to default heuristics if validation fails |
| 538 | const auto gemm_config = select_default_gemm_config_reshaped_only_rhs(auto_heuristics::CommonQuery{ gpu_target, data_type, m, n, k, batch_size }); |
| 539 | lhs_info = gemm_config.lhs_info; |
| 540 | rhs_info = gemm_config.rhs_info; |
| 541 | |
| 542 | auto_init_if_empty(tmp_b_info, b->clone()->set_tensor_shape(compute_rhs_reshaped_shape(*b, rhs_info))); |
| 543 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmReshapeRhsMatrixKernel::validate(b, &tmp_b_info, rhs_info)); |
| 544 | |
| 545 | // Validate matrix multiply |
| 546 | kernel_info.has_pad_y = false; |
| 547 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 548 | |
| 549 | kernel_info.has_pad_y = true; |
| 550 | ARM_COMPUTE_RETURN_ON_ERROR(ClGemmMatrixMultiplyReshapedOnlyRhsKernel::validate(a, &tmp_b_info, c, output, alpha, beta, lhs_info, rhs_info, kernel_info)); |
| 551 | |
| 552 | return Status{}; |
| 553 | } |
| 554 | |
| 555 | void ClGemm::configure(const CLCompileContext &compile_context, ITensorInfo *a, ITensorInfo *b, ITensorInfo *c, ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 556 | { |
| 557 | ARM_COMPUTE_ERROR_ON_NULLPTR(a, b, output); |
| 558 | |
| 559 | // Perform validation step |
| 560 | ARM_COMPUTE_ERROR_THROW_ON(validate(a, b, c, output, alpha, beta, gemm_info)); |
| 561 | |
| 562 | // Check if we need to reshape the matrix B only on the first run |
| 563 | _reshape_b_only_on_first_run = gemm_info.reshape_b_only_on_first_run(); |
| 564 | |
| 565 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 566 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 567 | const unsigned int n = b->dimension(0); |
| 568 | const unsigned int k = a->dimension(0); |
| 569 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 570 | |
| 571 | // Select GEMMType |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 572 | _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, |
| 573 | gemm_info.constant_weights()); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 574 | |
| 575 | const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); |
| 576 | |
| 577 | ITensorInfo *c_to_use = fuse_add_c ? c : nullptr; |
| 578 | |
| 579 | switch(_gemm_kernel_type) |
| 580 | { |
| 581 | case CLGEMMKernelType::NATIVE_V1: |
| 582 | { |
| 583 | configure_native_v1(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
| 584 | break; |
| 585 | } |
| 586 | case CLGEMMKernelType::RESHAPED_V1: |
| 587 | { |
| 588 | configure_reshaped_v1(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
| 589 | break; |
| 590 | } |
| 591 | case CLGEMMKernelType::RESHAPED: |
| 592 | { |
| 593 | configure_reshaped_v2(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
| 594 | break; |
| 595 | } |
| 596 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 597 | { |
| 598 | configure_reshaped_only_rhs(compile_context, a, b, c_to_use, output, alpha, beta, gemm_info); |
| 599 | break; |
| 600 | } |
| 601 | default: |
| 602 | { |
| 603 | ARM_COMPUTE_ERROR("GEMMType not supported"); |
| 604 | } |
| 605 | } |
| 606 | } |
| 607 | |
| 608 | Status ClGemm::validate(const ITensorInfo *a, const ITensorInfo *b, const ITensorInfo *c, const ITensorInfo *output, float alpha, float beta, const GEMMInfo &gemm_info) |
| 609 | { |
| 610 | // Get the GPU target |
| 611 | bool reinterpret_input_as_3d = gemm_info.reinterpret_input_as_3d(); |
| 612 | const unsigned int m = reinterpret_input_as_3d ? (a->dimension(1) * a->dimension(2)) : a->dimension(1); |
| 613 | const unsigned int n = b->dimension(0); |
| 614 | const unsigned int k = a->dimension(0); |
| 615 | const unsigned int batch_size = reinterpret_input_as_3d ? a->dimension(3) : a->dimension(2); |
| 616 | |
| 617 | // Select GEMMType |
| 618 | CLGEMMKernelType gemm_kernel_type = auto_select_gemm_kernel(auto_heuristics::CommonQuery |
| 619 | { |
| 620 | CLScheduler::get().target(), a->data_type(), m, n, k, batch_size, |
| 621 | }, |
Giorgio Arena | 4403ed3 | 2021-05-17 13:03:50 +0100 | [diff] [blame] | 622 | gemm_info.reshape_b_only_on_first_run(), gemm_info.constant_weights()); |
Georgios Pinitas | 856f66e | 2021-04-22 21:13:21 +0100 | [diff] [blame] | 623 | |
| 624 | const bool fuse_add_c = (!(helpers::float_ops::is_zero(beta)) && c != nullptr); |
| 625 | |
| 626 | const ITensorInfo *c_to_use = fuse_add_c ? c : nullptr; |
| 627 | |
| 628 | switch(gemm_kernel_type) |
| 629 | { |
| 630 | case CLGEMMKernelType::NATIVE_V1: |
| 631 | { |
| 632 | ARM_COMPUTE_RETURN_ON_ERROR(validate_native_v1(a, b, c_to_use, output, alpha, beta, gemm_info)); |
| 633 | break; |
| 634 | } |
| 635 | case CLGEMMKernelType::RESHAPED_V1: |
| 636 | { |
| 637 | ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_v1(a, b, c_to_use, output, alpha, beta, gemm_info)); |
| 638 | break; |
| 639 | } |
| 640 | case CLGEMMKernelType::RESHAPED: |
| 641 | { |
| 642 | ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped(a, b, c_to_use, output, alpha, beta, gemm_info)); |
| 643 | break; |
| 644 | } |
| 645 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 646 | { |
| 647 | ARM_COMPUTE_RETURN_ON_ERROR(validate_reshaped_only_rhs(a, b, c_to_use, output, alpha, beta, gemm_info)); |
| 648 | break; |
| 649 | } |
| 650 | default: |
| 651 | { |
| 652 | ARM_COMPUTE_RETURN_ERROR_MSG("GEMMType not supported"); |
| 653 | } |
| 654 | } |
| 655 | |
| 656 | return Status{}; |
| 657 | } |
| 658 | |
| 659 | void ClGemm::run(ITensorPack &tensors) |
| 660 | { |
| 661 | const ITensor *lhs = tensors.get_const_tensor(ACL_SRC_0); |
| 662 | const ITensor *rhs = tensors.get_const_tensor(ACL_SRC_1); |
| 663 | const ITensor *src2 = tensors.get_const_tensor(ACL_SRC_2); |
| 664 | ITensor *dst = tensors.get_tensor(ACL_DST); |
| 665 | |
| 666 | ARM_COMPUTE_ERROR_ON_NULLPTR(lhs, dst); |
| 667 | |
| 668 | CLAuxTensorHandler lhs_reshaped(offset_int_vec(LhsReshape), _tmp_a, tensors, true); |
| 669 | CLAuxTensorHandler rhs_reshaped(offset_int_vec(RhsReshape), _tmp_b, tensors, true); |
| 670 | |
| 671 | // Prepare the consts if needed |
| 672 | prepare(tensors); |
| 673 | |
| 674 | // Run matrix multiply kernel |
| 675 | switch(_gemm_kernel_type) |
| 676 | { |
| 677 | case CLGEMMKernelType::NATIVE_V1: |
| 678 | { |
| 679 | CLScheduler::get().enqueue_op(*_mm_kernel, tensors, true); |
| 680 | break; |
| 681 | } |
| 682 | case CLGEMMKernelType::RESHAPED_V1: |
| 683 | case CLGEMMKernelType::RESHAPED: |
| 684 | { |
| 685 | // Run interleave kernel |
| 686 | ITensorPack reshape_lhs_pack{ { ACL_SRC, lhs }, { ACL_DST, lhs_reshaped.get() } }; |
| 687 | CLScheduler::get().enqueue_op(*_reshape_lhs_kernel, reshape_lhs_pack, false); |
| 688 | |
| 689 | if(!_reshape_b_only_on_first_run) |
| 690 | { |
| 691 | // Run transpose kernel |
| 692 | ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } }; |
| 693 | CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false); |
| 694 | } |
| 695 | |
| 696 | ITensorPack gemm_reshaped_pack{ { ACL_SRC_0, lhs_reshaped.get() }, { ACL_SRC_1, rhs_reshaped.get() }, { ACL_SRC_2, src2 }, { ACL_DST, dst } }; |
| 697 | if(_gemm_kernel_type == CLGEMMKernelType::RESHAPED) |
| 698 | { |
| 699 | CLScheduler::get().enqueue_op(*_mm_reshaped_kernel, gemm_reshaped_pack, true); |
| 700 | } |
| 701 | else |
| 702 | { |
| 703 | CLScheduler::get().enqueue_op(*_mm_kernel, gemm_reshaped_pack, true); |
| 704 | } |
| 705 | break; |
| 706 | } |
| 707 | case CLGEMMKernelType::RESHAPED_ONLY_RHS: |
| 708 | { |
| 709 | if(!_reshape_b_only_on_first_run) |
| 710 | { |
| 711 | // Run transpose kernel |
| 712 | ITensorPack reshape_rhs_pack{ { ACL_SRC, rhs }, { ACL_DST, rhs_reshaped.get() } }; |
| 713 | CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, false); |
| 714 | } |
| 715 | // In case of RESHAPED_ONLY_RHS, we need to check the padding requirement |
| 716 | // Check if the lhs or dst tensors have padding |
| 717 | const unsigned int cross_plane_pad_lhs = lhs->info()->padding().top + lhs->info()->padding().bottom; |
| 718 | const unsigned int cross_plane_pad_dst = dst->info()->padding().top + dst->info()->padding().bottom; |
| 719 | bool has_pad_y = (cross_plane_pad_lhs != 0) || (cross_plane_pad_dst != 0); |
| 720 | |
| 721 | ITensorPack gemm_reshaped_onlyrhs_pack{ { ACL_SRC_0, lhs }, { ACL_SRC_1, rhs_reshaped.get() }, { ACL_SRC_2, src2 }, { ACL_DST, dst } }; |
| 722 | if(has_pad_y) |
| 723 | { |
| 724 | CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_fallback_kernel, gemm_reshaped_onlyrhs_pack, true); |
| 725 | } |
| 726 | else |
| 727 | { |
| 728 | CLScheduler::get().enqueue_op(*_mm_reshaped_only_rhs_kernel, gemm_reshaped_onlyrhs_pack, true); |
| 729 | } |
| 730 | break; |
| 731 | } |
| 732 | default: |
| 733 | { |
| 734 | ARM_COMPUTE_ERROR("GEMMType not supported"); |
| 735 | } |
| 736 | } |
| 737 | } |
| 738 | |
| 739 | void ClGemm::prepare(ITensorPack &constants) |
| 740 | { |
| 741 | const ITensor *src1 = constants.get_const_tensor(ACL_SRC_1); |
| 742 | ICLTensor *rhs_aux = utils::cast::polymorphic_downcast<ICLTensor *>(constants.get_tensor(offset_int_vec(RhsReshape))); |
| 743 | |
| 744 | // If memory for RHS is persistent and src1 is provided re-transform else assume that RHS is transformed |
| 745 | if((_aux_mem[AuxTensorIdx::RhsReshape].lifetime == MemoryLifetime::Persistent) && (src1 != nullptr && rhs_aux != nullptr) && rhs_aux) |
| 746 | { |
| 747 | CLAuxTensorHandler rhs_reshaped(_tmp_b, *rhs_aux); |
| 748 | ARM_COMPUTE_ERROR_ON(rhs_reshaped.get()->cl_buffer().get() == nullptr); |
| 749 | |
| 750 | ITensorPack reshape_rhs_pack{ { ACL_SRC, src1 }, { ACL_DST, rhs_reshaped.get() } }; |
| 751 | CLScheduler::get().enqueue_op(*_reshape_rhs_kernel, reshape_rhs_pack, true); |
| 752 | } |
| 753 | } |
| 754 | |
| 755 | experimental::MemoryRequirements ClGemm::workspace() const |
| 756 | { |
| 757 | return _aux_mem; |
| 758 | } |
| 759 | } // namespace opencl |
| 760 | } // namespace arm_compute |