Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2022 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 | */ |
Ramy Elgammal | df6a3b0 | 2022-11-30 16:23:10 +0000 | [diff] [blame] | 24 | #include "src/runtime/heuristics/direct_conv/ClDirectConvDefaultConfigValhall.h" |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 25 | |
| 26 | #include "arm_compute/core/CL/CLHelpers.h" |
| 27 | #include "arm_compute/core/CL/CLKernelLibrary.h" |
| 28 | #include "arm_compute/core/GPUTarget.h" |
| 29 | #include "arm_compute/core/TensorInfo.h" |
| 30 | #include "arm_compute/core/TensorShape.h" |
| 31 | #include "arm_compute/core/utils/misc/ShapeCalculator.h" |
| 32 | #include <utility> |
| 33 | |
| 34 | namespace arm_compute |
| 35 | { |
| 36 | namespace cl_direct_conv |
| 37 | { |
| 38 | using namespace arm_compute::misc::shape_calculator; |
| 39 | |
| 40 | ClDirectConvDefaultConfigValhall::ClDirectConvDefaultConfigValhall(GPUTarget gpu) |
| 41 | : IClDirectConvKernelConfig(gpu) |
| 42 | { |
| 43 | } |
| 44 | |
| 45 | DirectConvComputeKernelInfo ClDirectConvDefaultConfigValhall::configure(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) |
| 46 | { |
| 47 | using ConfigurationFunctionExecutorPtr = DirectConvComputeKernelInfo (ClDirectConvDefaultConfigValhall::*)(const ITensorInfo * src, const ITensorInfo * wei, const PadStrideInfo & conv_info); |
| 48 | |
| 49 | ClDirectConvConfigArray<ConfigurationFunctionExecutorPtr> configs_G78(&ClDirectConvDefaultConfigValhall::configure_G78_f32, |
| 50 | &ClDirectConvDefaultConfigValhall::configure_G78_f16, |
| 51 | &ClDirectConvDefaultConfigValhall::configure_G78_u8); |
| 52 | |
| 53 | ClDirectConvConfigArray<ConfigurationFunctionExecutorPtr> configs_G57(&ClDirectConvDefaultConfigValhall::configure_G57_f32, |
| 54 | &ClDirectConvDefaultConfigValhall::configure_G57_f16, |
| 55 | &ClDirectConvDefaultConfigValhall::configure_G78_u8); |
| 56 | |
| 57 | ConfigurationFunctionExecutorPtr func = nullptr; |
| 58 | switch(_target) |
| 59 | { |
| 60 | case GPUTarget::G57: |
| 61 | func = configs_G57.get_function(src->data_type()); |
| 62 | break; |
| 63 | case GPUTarget::G78: |
| 64 | default: |
| 65 | func = configs_G78.get_function(src->data_type()); |
| 66 | break; |
| 67 | } |
| 68 | |
| 69 | ARM_COMPUTE_ERROR_ON_MSG(func == nullptr, "Data type not supported for direct convolution"); |
| 70 | return (this->*func)(src, wei, conv_info); |
| 71 | } |
| 72 | |
| 73 | DirectConvComputeKernelInfo ClDirectConvDefaultConfigValhall::configure_G78_f32(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) |
| 74 | { |
| 75 | DirectConvComputeKernelInfo desc; |
| 76 | |
| 77 | if(src->data_layout() == DataLayout::NHWC) |
| 78 | { |
| 79 | // Get the output shape |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 80 | const TensorShape wei_shape = wei->tensor_shape(); |
| 81 | const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); |
| 82 | const bool export_weights_to_cl_image = export_to_cl_image(wei); |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 83 | |
| 84 | const int32_t ofm = dst_shape[0]; |
| 85 | const int32_t m = dst_shape[1] * dst_shape[2]; |
| 86 | const bool is_pointwise = (wei_shape[1] == wei_shape[2]) && wei_shape[1] == 1; |
| 87 | |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 88 | desc.export_weights_to_cl_image = export_weights_to_cl_image; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 89 | |
| 90 | if(dst_shape[0] <= 4) |
| 91 | { |
| 92 | if(is_pointwise) |
| 93 | { |
| 94 | if(ofm == 4) |
| 95 | { |
| 96 | desc.m0 = 1; |
| 97 | desc.n0 = 4; |
| 98 | desc.k0 = 16; |
| 99 | } |
| 100 | else |
| 101 | { |
| 102 | desc.m0 = 1; |
| 103 | desc.n0 = 1; |
| 104 | desc.k0 = 16; |
| 105 | } |
| 106 | } |
| 107 | else |
| 108 | { |
| 109 | desc.m0 = 1; |
| 110 | desc.n0 = 2; |
| 111 | desc.k0 = 16; |
| 112 | } |
| 113 | } |
| 114 | else |
| 115 | { |
| 116 | if(m < 64) |
| 117 | { |
| 118 | desc.m0 = 1; |
| 119 | desc.n0 = 1; |
| 120 | desc.k0 = 16; |
| 121 | } |
| 122 | else |
| 123 | { |
| 124 | desc.m0 = 4; |
| 125 | desc.n0 = 4; |
| 126 | desc.k0 = 4; |
| 127 | } |
| 128 | } |
| 129 | } |
| 130 | |
| 131 | return desc; |
| 132 | } |
| 133 | |
| 134 | DirectConvComputeKernelInfo ClDirectConvDefaultConfigValhall::configure_G78_f16(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) |
| 135 | { |
| 136 | DirectConvComputeKernelInfo desc; |
| 137 | |
| 138 | if(src->data_layout() == DataLayout::NHWC) |
| 139 | { |
| 140 | // Get the output shape |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 141 | const TensorShape wei_shape = wei->tensor_shape(); |
| 142 | const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); |
| 143 | const bool export_weights_to_cl_image = export_to_cl_image(wei); |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 144 | |
| 145 | const int32_t ofm = dst_shape[0]; |
| 146 | const int32_t m = dst_shape[1] * dst_shape[2]; |
Gian Marco Iodice | 3394f3e | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 147 | const int32_t k = wei_shape[0]; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 148 | const bool is_pointwise = (wei_shape[1] == wei_shape[2]) && wei_shape[1] == 1; |
| 149 | |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 150 | desc.export_weights_to_cl_image = export_weights_to_cl_image; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 151 | |
| 152 | if(dst_shape[0] <= 4) |
| 153 | { |
Gian Marco Iodice | 3394f3e | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 154 | // k0 should be as larger as possible. However, we should avoid |
| 155 | // having left-over for loops that make the implementation slower. |
| 156 | if((k % 16) == 0) |
| 157 | { |
| 158 | desc.k0 = 16; |
| 159 | } |
| 160 | else if((k % 8) == 0) |
| 161 | { |
| 162 | desc.k0 = 8; |
| 163 | } |
| 164 | else |
| 165 | { |
| 166 | desc.k0 = 4; |
| 167 | } |
| 168 | |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 169 | if(is_pointwise) |
| 170 | { |
| 171 | if(ofm == 4) |
| 172 | { |
| 173 | desc.m0 = 1; |
| 174 | desc.n0 = 4; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 175 | } |
| 176 | else |
| 177 | { |
| 178 | desc.m0 = 1; |
| 179 | desc.n0 = 1; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 180 | } |
| 181 | } |
| 182 | else |
| 183 | { |
| 184 | desc.m0 = 1; |
| 185 | desc.n0 = dst_shape[0]; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 186 | } |
| 187 | } |
| 188 | else |
| 189 | { |
| 190 | if(m < 64) |
| 191 | { |
| 192 | desc.m0 = 1; |
| 193 | desc.n0 = 1; |
Gian Marco Iodice | 3394f3e | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 194 | if((k % 16) == 0) |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 195 | { |
Gian Marco Iodice | 3394f3e | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 196 | desc.k0 = 16; |
| 197 | } |
| 198 | else if((k % 8) == 0) |
| 199 | { |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 200 | desc.k0 = 8; |
| 201 | } |
| 202 | else |
| 203 | { |
Gian Marco Iodice | 3394f3e | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 204 | desc.k0 = 4; |
| 205 | } |
| 206 | } |
| 207 | else |
| 208 | { |
| 209 | if(ofm >= 16) |
| 210 | { |
| 211 | if(m / 6 > 24000) |
| 212 | { |
| 213 | desc.m0 = 6; |
| 214 | } |
| 215 | else |
| 216 | { |
| 217 | desc.m0 = 5; |
| 218 | } |
| 219 | desc.n0 = 8; |
| 220 | desc.k0 = 4; |
| 221 | } |
| 222 | else |
| 223 | { |
| 224 | desc.m0 = 2; |
| 225 | desc.n0 = 8; |
| 226 | if((k % 16) == 0) |
| 227 | { |
| 228 | desc.k0 = 16; |
| 229 | } |
| 230 | else if((k % 8) == 0) |
| 231 | { |
| 232 | desc.k0 = 8; |
| 233 | } |
| 234 | else |
| 235 | { |
| 236 | desc.k0 = 4; |
| 237 | } |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 238 | } |
| 239 | } |
| 240 | } |
| 241 | } |
| 242 | |
| 243 | return desc; |
| 244 | } |
| 245 | |
| 246 | DirectConvComputeKernelInfo ClDirectConvDefaultConfigValhall::configure_G78_u8(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) |
| 247 | { |
| 248 | DirectConvComputeKernelInfo desc; |
| 249 | |
| 250 | if(src->data_layout() == DataLayout::NHWC) |
| 251 | { |
| 252 | // Get the output shape |
| 253 | TensorShape output_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); |
| 254 | |
| 255 | desc.n0 = 4; |
| 256 | |
| 257 | if(output_shape[0] > 16) |
| 258 | { |
| 259 | desc.m0 = 4; |
| 260 | } |
| 261 | |
| 262 | desc.k0 = 16; |
| 263 | |
| 264 | desc.export_weights_to_cl_image = false; |
| 265 | } |
| 266 | |
| 267 | return desc; |
| 268 | } |
| 269 | |
| 270 | DirectConvComputeKernelInfo ClDirectConvDefaultConfigValhall::configure_G57_f32(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) |
| 271 | { |
| 272 | DirectConvComputeKernelInfo desc; |
| 273 | |
| 274 | if(src->data_layout() == DataLayout::NHWC) |
| 275 | { |
| 276 | // Get the output shape |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 277 | const TensorShape wei_shape = wei->tensor_shape(); |
| 278 | const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); |
| 279 | const bool export_weights_to_cl_image = export_to_cl_image(wei); |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 280 | |
| 281 | const int32_t m = dst_shape[1] * dst_shape[2]; |
| 282 | const bool is_pointwise = (wei_shape[1] == wei_shape[2]) && wei_shape[1] == 1; |
| 283 | |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 284 | desc.export_weights_to_cl_image = export_weights_to_cl_image; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 285 | |
| 286 | if(dst_shape[0] <= 4) |
| 287 | { |
| 288 | if(is_pointwise) |
| 289 | { |
| 290 | desc.m0 = 1; |
| 291 | desc.n0 = 1; |
| 292 | desc.k0 = 16; |
| 293 | } |
| 294 | else |
| 295 | { |
| 296 | desc.m0 = 1; |
| 297 | desc.n0 = dst_shape[0]; |
| 298 | desc.k0 = 16; |
| 299 | } |
| 300 | } |
| 301 | else |
| 302 | { |
| 303 | if(m < 64) |
| 304 | { |
| 305 | if(m == 1) |
| 306 | { |
| 307 | desc.m0 = 1; |
| 308 | desc.n0 = 1; |
| 309 | desc.k0 = 16; |
| 310 | } |
| 311 | else |
| 312 | { |
| 313 | desc.m0 = 4; |
| 314 | desc.n0 = 2; |
| 315 | desc.k0 = 8; |
| 316 | } |
| 317 | } |
| 318 | else |
| 319 | { |
| 320 | desc.m0 = 4; |
| 321 | desc.n0 = 4; |
| 322 | desc.k0 = 4; |
| 323 | } |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | return desc; |
| 328 | } |
| 329 | |
| 330 | DirectConvComputeKernelInfo ClDirectConvDefaultConfigValhall::configure_G57_f16(const ITensorInfo *src, const ITensorInfo *wei, const PadStrideInfo &conv_info) |
| 331 | { |
| 332 | DirectConvComputeKernelInfo desc; |
| 333 | |
| 334 | if(src->data_layout() == DataLayout::NHWC) |
| 335 | { |
| 336 | // Get the output shape |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 337 | const TensorShape wei_shape = wei->tensor_shape(); |
| 338 | const TensorShape dst_shape = misc::shape_calculator::compute_deep_convolution_shape(*src, *wei, conv_info); |
| 339 | const bool export_weights_to_cl_image = export_to_cl_image(wei); |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 340 | |
| 341 | const int32_t ofm = dst_shape[0]; |
| 342 | const int32_t m = dst_shape[1] * dst_shape[2]; |
| 343 | const bool is_pointwise = (wei_shape[1] == wei_shape[2]) && wei_shape[1] == 1; |
| 344 | |
Gian Marco Iodice | ad9a7ed | 2022-09-16 14:14:21 +0100 | [diff] [blame] | 345 | desc.export_weights_to_cl_image = export_weights_to_cl_image; |
Gian Marco Iodice | 2cc50b3 | 2022-05-30 14:41:49 +0100 | [diff] [blame] | 346 | |
| 347 | if(dst_shape[0] <= 4) |
| 348 | { |
| 349 | if(is_pointwise) |
| 350 | { |
| 351 | desc.m0 = 2; |
| 352 | desc.n0 = 1; |
| 353 | desc.k0 = 16; |
| 354 | } |
| 355 | else |
| 356 | { |
| 357 | desc.m0 = 1; |
| 358 | desc.n0 = dst_shape[0]; |
| 359 | desc.k0 = 16; |
| 360 | } |
| 361 | } |
| 362 | else |
| 363 | { |
| 364 | if(m < 64) |
| 365 | { |
| 366 | if(m == 1) |
| 367 | { |
| 368 | desc.m0 = 1; |
| 369 | desc.n0 = 1; |
| 370 | desc.k0 = 16; |
| 371 | } |
| 372 | else |
| 373 | { |
| 374 | desc.m0 = 4; |
| 375 | desc.n0 = 2; |
| 376 | desc.k0 = 8; |
| 377 | } |
| 378 | } |
| 379 | else |
| 380 | { |
| 381 | if(ofm > 16) |
| 382 | { |
| 383 | desc.m0 = 4; |
| 384 | desc.n0 = 8; |
| 385 | desc.k0 = 8; |
| 386 | } |
| 387 | else |
| 388 | { |
| 389 | desc.m0 = 8; |
| 390 | desc.n0 = 4; |
| 391 | desc.k0 = 4; |
| 392 | } |
| 393 | } |
| 394 | } |
| 395 | } |
| 396 | |
| 397 | return desc; |
| 398 | } |
| 399 | } // namespace opencl |
| 400 | } // namespace arm_compute |