| 1 | // Copyright (c) Facebook, Inc. and its affiliates. |
|---|---|
| 2 | // All rights reserved. |
| 3 | // |
| 4 | // Copyright 2019 Google LLC |
| 5 | // |
| 6 | // This source code is licensed under the BSD-style license found in the |
| 7 | // LICENSE file in the root directory of this source tree. |
| 8 | |
| 9 | #include <stdint.h> |
| 10 | #include <stddef.h> |
| 11 | #include <string.h> |
| 12 | |
| 13 | #include <fp16.h> |
| 14 | |
| 15 | #include <xnnpack/log.h> |
| 16 | #include <xnnpack/math.h> |
| 17 | #include <xnnpack/operator.h> |
| 18 | #include <xnnpack/pack.h> |
| 19 | #include <xnnpack/unaligned.h> |
| 20 | |
| 21 | |
| 22 | void xnn_pack_f32_gemm_goi_w( |
| 23 | size_t g, |
| 24 | size_t nc, |
| 25 | size_t kc, |
| 26 | size_t nr, |
| 27 | size_t kr, |
| 28 | size_t sr, |
| 29 | const float* k, |
| 30 | const float* b, |
| 31 | float* packed_weights, |
| 32 | size_t extra_bytes, |
| 33 | const void* params) |
| 34 | { |
| 35 | assert(g != 0); |
| 36 | assert(nr >= sr); |
| 37 | assert(k != NULL); |
| 38 | assert(packed_weights != NULL); |
| 39 | |
| 40 | const size_t skr = sr * kr; |
| 41 | do { |
| 42 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 43 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 44 | if XNN_LIKELY(b != NULL) { |
| 45 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 46 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 47 | } |
| 48 | } |
| 49 | packed_weights += nr; |
| 50 | |
| 51 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 52 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 53 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 54 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 55 | if (kc_idx < kc) { |
| 56 | packed_weights[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 57 | } |
| 58 | } |
| 59 | packed_weights += kr; |
| 60 | } |
| 61 | packed_weights += (nr - nr_block_size) * kr; |
| 62 | } |
| 63 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 64 | } |
| 65 | k += nc * kc; |
| 66 | if XNN_UNPREDICTABLE(b != NULL) { |
| 67 | b += nc; |
| 68 | } |
| 69 | } while (--g != 0); |
| 70 | } |
| 71 | |
| 72 | void xnn_pack_f16_gemm_goi_w( |
| 73 | size_t g, |
| 74 | size_t nc, |
| 75 | size_t kc, |
| 76 | size_t nr, |
| 77 | size_t kr, |
| 78 | size_t sr, |
| 79 | const uint16_t* k, |
| 80 | const uint16_t* b, |
| 81 | uint16_t* packed_weights, |
| 82 | size_t extra_bytes, |
| 83 | const void* params) |
| 84 | { |
| 85 | assert(g != 0); |
| 86 | assert(nr >= sr); |
| 87 | assert(k != NULL); |
| 88 | assert(packed_weights != NULL); |
| 89 | |
| 90 | const size_t skr = sr * kr; |
| 91 | do { |
| 92 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 93 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 94 | if XNN_LIKELY(b != NULL) { |
| 95 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 96 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 97 | } |
| 98 | } |
| 99 | packed_weights += nr; |
| 100 | |
| 101 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 102 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 103 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 104 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 105 | if (kc_idx < kc) { |
| 106 | packed_weights[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 107 | } |
| 108 | } |
| 109 | packed_weights += kr; |
| 110 | } |
| 111 | packed_weights += (nr - nr_block_size) * kr; |
| 112 | } |
| 113 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 114 | } |
| 115 | k += nc * kc; |
| 116 | if XNN_UNPREDICTABLE(b != NULL) { |
| 117 | b += nc; |
| 118 | } |
| 119 | } while (--g != 0); |
| 120 | } |
| 121 | |
| 122 | void xnn_pack_f32_to_f16_gemm_goi_w( |
| 123 | size_t g, |
| 124 | size_t nc, |
| 125 | size_t kc, |
| 126 | size_t nr, |
| 127 | size_t kr, |
| 128 | size_t sr, |
| 129 | const float* k, |
| 130 | const float* b, |
| 131 | uint16_t* packed_weights, |
| 132 | size_t extra_bytes, |
| 133 | const void* params) |
| 134 | { |
| 135 | assert(g != 0); |
| 136 | assert(nr >= sr); |
| 137 | assert(k != NULL); |
| 138 | assert(packed_weights != NULL); |
| 139 | |
| 140 | const size_t skr = sr * kr; |
| 141 | do { |
| 142 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 143 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 144 | if XNN_LIKELY(b != NULL) { |
| 145 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 146 | packed_weights[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 147 | } |
| 148 | } |
| 149 | packed_weights += nr; |
| 150 | |
| 151 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 152 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 153 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 154 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 155 | if (kc_idx < kc) { |
| 156 | packed_weights[kr_block_offset] = fp16_ieee_from_fp32_value(k[(nr_block_start + nr_block_offset) * kc + kc_idx]); |
| 157 | } |
| 158 | } |
| 159 | packed_weights += kr; |
| 160 | } |
| 161 | packed_weights += (nr - nr_block_size) * kr; |
| 162 | } |
| 163 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 164 | } |
| 165 | k += nc * kc; |
| 166 | if XNN_UNPREDICTABLE(b != NULL) { |
| 167 | b += nc; |
| 168 | } |
| 169 | } while (--g != 0); |
| 170 | } |
| 171 | |
| 172 | void xnn_pack_qu8_gemm_goi_w( |
| 173 | size_t g, |
| 174 | size_t nc, |
| 175 | size_t kc, |
| 176 | size_t nr, |
| 177 | size_t kr, |
| 178 | size_t sr, |
| 179 | const uint8_t* k, |
| 180 | const int32_t* b, |
| 181 | void* packed_weights, |
| 182 | size_t extra_bytes, |
| 183 | const struct xnn_qu8_packing_params* params) |
| 184 | { |
| 185 | assert(g != 0); |
| 186 | assert(nr >= sr); |
| 187 | assert(k != NULL); |
| 188 | assert(packed_weights != NULL); |
| 189 | |
| 190 | const size_t skr = sr * kr; |
| 191 | const int32_t izp = (int32_t) params->input_zero_point; |
| 192 | const int32_t bzp = (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
| 193 | do { |
| 194 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 195 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 196 | int32_t* packed_b = (int32_t*) packed_weights; |
| 197 | if XNN_LIKELY(b != NULL) { |
| 198 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 199 | unaligned_store_s32(packed_weights, bzp + b[nr_block_start + nr_block_offset]); |
| 200 | packed_weights = (int32_t*) packed_weights + 1; |
| 201 | } |
| 202 | } else { |
| 203 | size_t n = nr_block_size; |
| 204 | do { |
| 205 | unaligned_store_s32(packed_weights, bzp); |
| 206 | packed_weights = (int32_t*) packed_weights + 1; |
| 207 | } while (--n != 0); |
| 208 | } |
| 209 | packed_weights = (int32_t*) packed_weights + (nr - nr_block_size); |
| 210 | |
| 211 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 212 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 213 | int32_t ksum = 0; |
| 214 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 215 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 216 | if (kc_idx < kc) { |
| 217 | const uint8_t kv = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 218 | ksum += (int32_t) kv; |
| 219 | ((uint8_t*) packed_weights)[kr_block_offset] = kv; |
| 220 | } |
| 221 | } |
| 222 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 223 | packed_weights = (uint8_t*) packed_weights + kr; |
| 224 | } |
| 225 | packed_weights = (uint8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 226 | } |
| 227 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 228 | } |
| 229 | k += nc * kc; |
| 230 | if XNN_UNPREDICTABLE(b != NULL) { |
| 231 | b += nc; |
| 232 | } |
| 233 | } while (--g != 0); |
| 234 | } |
| 235 | |
| 236 | void xnn_pack_qs8_gemm_goi_w( |
| 237 | size_t g, |
| 238 | size_t nc, |
| 239 | size_t kc, |
| 240 | size_t nr, |
| 241 | size_t kr, |
| 242 | size_t sr, |
| 243 | const int8_t* k, |
| 244 | const int32_t* b, |
| 245 | void* packed_weights, |
| 246 | size_t extra_bytes, |
| 247 | const struct xnn_qs8_packing_params* params) |
| 248 | { |
| 249 | assert(g != 0); |
| 250 | assert(nr >= sr); |
| 251 | assert(k != NULL); |
| 252 | assert(packed_weights != NULL); |
| 253 | |
| 254 | const size_t skr = sr * kr; |
| 255 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 256 | do { |
| 257 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 258 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 259 | int32_t* packed_b = (int32_t*) packed_weights; |
| 260 | if XNN_LIKELY(b != NULL) { |
| 261 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 262 | unaligned_store_s32(packed_weights, b[nr_block_start + nr_block_offset]); |
| 263 | packed_weights = (int32_t*) packed_weights + 1; |
| 264 | } |
| 265 | } else { |
| 266 | size_t n = nr_block_size; |
| 267 | do { |
| 268 | unaligned_store_s32(packed_weights, 0); |
| 269 | packed_weights = (int32_t*) packed_weights + 1; |
| 270 | } while (--n != 0); |
| 271 | } |
| 272 | packed_weights = (int32_t*) packed_weights + (nr - nr_block_size); |
| 273 | |
| 274 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 275 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 276 | uint32_t ksum = 0; |
| 277 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 278 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 279 | if (kc_idx < kc) { |
| 280 | const int8_t kv = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 281 | ksum += (uint32_t) kv; |
| 282 | ((int8_t*) packed_weights)[kr_block_offset] = kv; |
| 283 | } |
| 284 | } |
| 285 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 286 | packed_weights = (int8_t*) packed_weights + kr; |
| 287 | } |
| 288 | packed_weights = (int8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 289 | } |
| 290 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 291 | } |
| 292 | k += nc * kc; |
| 293 | if XNN_UNPREDICTABLE(b != NULL) { |
| 294 | b += nc; |
| 295 | } |
| 296 | } while (--g != 0); |
| 297 | } |
| 298 | |
| 299 | void xnn_pack_qs8_gemm_xw_goi_w( |
| 300 | size_t g, |
| 301 | size_t nc, |
| 302 | size_t kc, |
| 303 | size_t nr, |
| 304 | size_t kr, |
| 305 | size_t sr, |
| 306 | const int8_t* k, |
| 307 | const int32_t* b, |
| 308 | void* packed_weights, |
| 309 | size_t extra_bytes, |
| 310 | const struct xnn_qs8_packing_params* params) |
| 311 | { |
| 312 | assert(g != 0); |
| 313 | assert(nr >= sr); |
| 314 | assert(k != NULL); |
| 315 | assert(packed_weights != NULL); |
| 316 | |
| 317 | const size_t skr = sr * kr; |
| 318 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 319 | do { |
| 320 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 321 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 322 | int32_t* packed_b = (int32_t*) packed_weights; |
| 323 | if XNN_LIKELY(b != NULL) { |
| 324 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 325 | unaligned_store_s32(packed_weights, b[nr_block_start + nr_block_offset]); |
| 326 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 327 | } |
| 328 | } else { |
| 329 | size_t n = nr_block_size; |
| 330 | do { |
| 331 | unaligned_store_s32(packed_weights, 0); |
| 332 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 333 | } while (--n != 0); |
| 334 | } |
| 335 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 336 | |
| 337 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 338 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 339 | uint32_t ksum = 0; |
| 340 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 341 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 342 | if (kc_idx < kc) { |
| 343 | const int8_t kv = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 344 | ksum += (uint32_t) kv; |
| 345 | ((int16_t*) packed_weights)[kr_block_offset] = (int16_t) kv; |
| 346 | } |
| 347 | } |
| 348 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 349 | packed_weights = (int16_t*) packed_weights + kr; |
| 350 | } |
| 351 | packed_weights = (int16_t*) packed_weights + (nr - nr_block_size) * kr; |
| 352 | } |
| 353 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 354 | } |
| 355 | k += nc * kc; |
| 356 | if XNN_UNPREDICTABLE(b != NULL) { |
| 357 | b += nc; |
| 358 | } |
| 359 | } while (--g != 0); |
| 360 | } |
| 361 | |
| 362 | void xnn_pack_f32_gemm_io_w( |
| 363 | size_t nc, |
| 364 | size_t kc, |
| 365 | size_t nr, |
| 366 | size_t kr, |
| 367 | size_t sr, |
| 368 | const float* k, |
| 369 | const float* b, |
| 370 | float* packed_weights, |
| 371 | const void* params) |
| 372 | { |
| 373 | assert(nr >= sr); |
| 374 | assert(k != NULL); |
| 375 | assert(packed_weights != NULL); |
| 376 | |
| 377 | const size_t skr = sr * kr; |
| 378 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 379 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 380 | if XNN_LIKELY(b != NULL) { |
| 381 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 382 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 383 | } |
| 384 | } |
| 385 | packed_weights += nr; |
| 386 | |
| 387 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 388 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 389 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 390 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 391 | if (kc_idx < kc) { |
| 392 | packed_weights[kr_block_offset] = k[kc_idx * nc + nr_block_start + nr_block_offset]; |
| 393 | } |
| 394 | } |
| 395 | packed_weights += kr; |
| 396 | } |
| 397 | packed_weights += (nr - nr_block_size) * kr; |
| 398 | } |
| 399 | } |
| 400 | } |
| 401 | |
| 402 | void xnn_pack_f16_gemm_io_w( |
| 403 | size_t nc, |
| 404 | size_t kc, |
| 405 | size_t nr, |
| 406 | size_t kr, |
| 407 | size_t sr, |
| 408 | const uint16_t* k, |
| 409 | const uint16_t* b, |
| 410 | uint16_t* packed_weights, |
| 411 | const void* params) |
| 412 | { |
| 413 | assert(nr >= sr); |
| 414 | assert(k != NULL); |
| 415 | assert(packed_weights != NULL); |
| 416 | |
| 417 | const size_t skr = sr * kr; |
| 418 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 419 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 420 | if XNN_LIKELY(b != NULL) { |
| 421 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 422 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 423 | } |
| 424 | } |
| 425 | packed_weights += nr; |
| 426 | |
| 427 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 428 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 429 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 430 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 431 | if (kc_idx < kc) { |
| 432 | packed_weights[kr_block_offset] = k[kc_idx * nc + nr_block_start + nr_block_offset]; |
| 433 | } |
| 434 | } |
| 435 | packed_weights += kr; |
| 436 | } |
| 437 | packed_weights += (nr - nr_block_size) * kr; |
| 438 | } |
| 439 | } |
| 440 | } |
| 441 | |
| 442 | void xnn_pack_f32_to_f16_gemm_io_w( |
| 443 | size_t nc, |
| 444 | size_t kc, |
| 445 | size_t nr, |
| 446 | size_t kr, |
| 447 | size_t sr, |
| 448 | const float* k, |
| 449 | const float* b, |
| 450 | uint16_t* packed_weights, |
| 451 | const void* params) |
| 452 | { |
| 453 | assert(nr >= sr); |
| 454 | assert(k != NULL); |
| 455 | assert(packed_weights != NULL); |
| 456 | |
| 457 | const size_t skr = sr * kr; |
| 458 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 459 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 460 | if XNN_LIKELY(b != NULL) { |
| 461 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 462 | packed_weights[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 463 | } |
| 464 | } |
| 465 | packed_weights += nr; |
| 466 | |
| 467 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 468 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 469 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 470 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 471 | if (kc_idx < kc) { |
| 472 | packed_weights[kr_block_offset] = fp16_ieee_from_fp32_value(k[kc_idx * nc + nr_block_start + nr_block_offset]); |
| 473 | } |
| 474 | } |
| 475 | packed_weights += kr; |
| 476 | } |
| 477 | packed_weights += (nr - nr_block_size) * kr; |
| 478 | } |
| 479 | } |
| 480 | } |
| 481 | |
| 482 | void xnn_pack_qu8_gemm_io_w( |
| 483 | size_t nc, |
| 484 | size_t kc, |
| 485 | size_t nr, |
| 486 | size_t kr, |
| 487 | size_t sr, |
| 488 | const uint8_t* k, |
| 489 | const int32_t* b, |
| 490 | void* packed_weights, |
| 491 | const struct xnn_qu8_packing_params* params) |
| 492 | { |
| 493 | assert(nr >= sr); |
| 494 | assert(k != NULL); |
| 495 | assert(packed_weights != NULL); |
| 496 | |
| 497 | const size_t skr = sr * kr; |
| 498 | const int32_t izp = (int32_t) params->input_zero_point; |
| 499 | const int32_t bzp = (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
| 500 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 501 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 502 | int32_t* packed_b = (int32_t*) packed_weights; |
| 503 | if XNN_LIKELY(b != NULL) { |
| 504 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 505 | unaligned_store_s32(packed_weights, bzp + b[nr_block_start + nr_block_offset]); |
| 506 | packed_weights = (int32_t*) packed_weights + 1; |
| 507 | } |
| 508 | } else { |
| 509 | size_t n = nr_block_size; |
| 510 | do { |
| 511 | unaligned_store_s32(packed_weights, bzp); |
| 512 | packed_weights = (int32_t*) packed_weights + 1; |
| 513 | } while (--n != 0); |
| 514 | } |
| 515 | packed_weights = (int32_t*) packed_weights + (nr - nr_block_size); |
| 516 | |
| 517 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 518 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 519 | int32_t ksum = 0; |
| 520 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 521 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 522 | if (kc_idx < kc) { |
| 523 | const uint8_t kv = k[kc_idx * nc + (nr_block_start + nr_block_offset)]; |
| 524 | ksum += (int32_t) kv; |
| 525 | ((uint8_t*) packed_weights)[kr_block_offset] = kv; |
| 526 | } |
| 527 | } |
| 528 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 529 | packed_weights = (uint8_t*) packed_weights + kr; |
| 530 | } |
| 531 | packed_weights = (uint8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 532 | } |
| 533 | } |
| 534 | } |
| 535 | |
| 536 | void xnn_pack_qs8_gemm_io_w( |
| 537 | size_t nc, |
| 538 | size_t kc, |
| 539 | size_t nr, |
| 540 | size_t kr, |
| 541 | size_t sr, |
| 542 | const int8_t* k, |
| 543 | const int32_t* b, |
| 544 | void* packed_weights, |
| 545 | const struct xnn_qs8_packing_params* params) |
| 546 | { |
| 547 | assert(nr >= sr); |
| 548 | assert(k != NULL); |
| 549 | assert(packed_weights != NULL); |
| 550 | |
| 551 | const size_t skr = sr * kr; |
| 552 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 553 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 554 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 555 | int32_t* packed_b = (int32_t*) packed_weights; |
| 556 | if XNN_LIKELY(b != NULL) { |
| 557 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 558 | unaligned_store_s32(packed_weights, b[nr_block_start + nr_block_offset]); |
| 559 | packed_weights = (int32_t*) packed_weights + 1; |
| 560 | } |
| 561 | } else { |
| 562 | size_t n = nr_block_size; |
| 563 | do { |
| 564 | unaligned_store_s32(packed_weights, 0); |
| 565 | packed_weights = (int32_t*) packed_weights + 1; |
| 566 | } while (--n != 0); |
| 567 | } |
| 568 | packed_weights = (uint32_t*) packed_weights + (nr - nr_block_size); |
| 569 | |
| 570 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 571 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 572 | uint32_t ksum = 0; |
| 573 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 574 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 575 | if (kc_idx < kc) { |
| 576 | const int8_t kv = k[kc_idx * nc + (nr_block_start + nr_block_offset)]; |
| 577 | ksum += (uint32_t) kv; |
| 578 | ((int8_t*) packed_weights)[kr_block_offset] = kv; |
| 579 | } |
| 580 | } |
| 581 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 582 | packed_weights = (int8_t*) packed_weights + kr; |
| 583 | } |
| 584 | packed_weights = (int8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 585 | } |
| 586 | } |
| 587 | } |
| 588 | |
| 589 | void xnn_pack_f32_conv_goki_w( |
| 590 | size_t g, |
| 591 | size_t nc, |
| 592 | size_t ks, |
| 593 | size_t kc, |
| 594 | size_t nr, |
| 595 | size_t kr, |
| 596 | size_t sr, |
| 597 | const float* k, |
| 598 | const float* b, |
| 599 | float* packed_weights, |
| 600 | size_t extra_bytes, |
| 601 | const void* params) |
| 602 | { |
| 603 | assert(g != 0); |
| 604 | assert(nr >= sr); |
| 605 | assert(k != NULL); |
| 606 | assert(packed_weights != NULL); |
| 607 | |
| 608 | const size_t skr = sr * kr; |
| 609 | do { |
| 610 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 611 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 612 | if XNN_LIKELY(b != NULL) { |
| 613 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 614 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 615 | } |
| 616 | } |
| 617 | packed_weights += nr; |
| 618 | |
| 619 | for (size_t ki = 0; ki < ks; ki++) { |
| 620 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 621 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 622 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 623 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 624 | if (kc_idx < kc) { |
| 625 | packed_weights[kr_block_offset] = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 626 | } |
| 627 | } |
| 628 | packed_weights += kr; |
| 629 | } |
| 630 | packed_weights += (nr - nr_block_size) * kr; |
| 631 | } |
| 632 | } |
| 633 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 634 | } |
| 635 | k += ks * kc * nc; |
| 636 | if XNN_UNPREDICTABLE(b != NULL) { |
| 637 | b += nc; |
| 638 | } |
| 639 | } while (--g != 0); |
| 640 | } |
| 641 | |
| 642 | void xnn_pack_f16_conv_goki_w( |
| 643 | size_t g, |
| 644 | size_t nc, |
| 645 | size_t ks, |
| 646 | size_t kc, |
| 647 | size_t nr, |
| 648 | size_t kr, |
| 649 | size_t sr, |
| 650 | const uint16_t* k, |
| 651 | const uint16_t* b, |
| 652 | uint16_t* packed_weights, |
| 653 | size_t extra_bytes, |
| 654 | const void* params) |
| 655 | { |
| 656 | assert(g != 0); |
| 657 | assert(nr >= sr); |
| 658 | assert(k != NULL); |
| 659 | assert(packed_weights != NULL); |
| 660 | |
| 661 | const size_t skr = sr * kr; |
| 662 | do { |
| 663 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 664 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 665 | if XNN_LIKELY(b != NULL) { |
| 666 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 667 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 668 | } |
| 669 | } |
| 670 | packed_weights += nr; |
| 671 | |
| 672 | for (size_t ki = 0; ki < ks; ki++) { |
| 673 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 674 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 675 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 676 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 677 | if (kc_idx < kc) { |
| 678 | packed_weights[kr_block_offset] = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 679 | } |
| 680 | } |
| 681 | packed_weights += kr; |
| 682 | } |
| 683 | packed_weights += (nr - nr_block_size) * kr; |
| 684 | } |
| 685 | } |
| 686 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 687 | } |
| 688 | k += ks * kc * nc; |
| 689 | if XNN_UNPREDICTABLE(b != NULL) { |
| 690 | b += nc; |
| 691 | } |
| 692 | } while (--g != 0); |
| 693 | } |
| 694 | |
| 695 | void xnn_pack_f32_to_f16_conv_goki_w( |
| 696 | size_t g, |
| 697 | size_t nc, |
| 698 | size_t ks, |
| 699 | size_t kc, |
| 700 | size_t nr, |
| 701 | size_t kr, |
| 702 | size_t sr, |
| 703 | const float* k, |
| 704 | const float* b, |
| 705 | uint16_t* packed_weights, |
| 706 | size_t extra_bytes, |
| 707 | const void* params) |
| 708 | { |
| 709 | assert(g != 0); |
| 710 | assert(nr >= sr); |
| 711 | assert(k != NULL); |
| 712 | assert(packed_weights != NULL); |
| 713 | |
| 714 | const size_t skr = sr * kr; |
| 715 | do { |
| 716 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 717 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 718 | if XNN_LIKELY(b != NULL) { |
| 719 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 720 | packed_weights[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 721 | } |
| 722 | } |
| 723 | packed_weights += nr; |
| 724 | |
| 725 | for (size_t ki = 0; ki < ks; ki++) { |
| 726 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 727 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 728 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 729 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 730 | if (kc_idx < kc) { |
| 731 | packed_weights[kr_block_offset] = fp16_ieee_from_fp32_value(k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]); |
| 732 | } |
| 733 | } |
| 734 | packed_weights += kr; |
| 735 | } |
| 736 | packed_weights += (nr - nr_block_size) * kr; |
| 737 | } |
| 738 | } |
| 739 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 740 | } |
| 741 | k += ks * kc * nc; |
| 742 | if XNN_UNPREDICTABLE(b != NULL) { |
| 743 | b += nc; |
| 744 | } |
| 745 | } while (--g != 0); |
| 746 | } |
| 747 | |
| 748 | void xnn_pack_qu8_conv_goki_w( |
| 749 | size_t g, |
| 750 | size_t nc, |
| 751 | size_t ks, |
| 752 | size_t kc, |
| 753 | size_t nr, |
| 754 | size_t kr, |
| 755 | size_t sr, |
| 756 | const uint8_t* k, |
| 757 | const int32_t* b, |
| 758 | void* packed_weights, |
| 759 | size_t extra_bytes, |
| 760 | const struct xnn_qu8_packing_params* params) |
| 761 | { |
| 762 | assert(g != 0); |
| 763 | assert(nr >= sr); |
| 764 | assert(k != NULL); |
| 765 | assert(packed_weights != NULL); |
| 766 | |
| 767 | const size_t skr = sr * kr; |
| 768 | const int32_t izp = (int32_t) params->input_zero_point; |
| 769 | const int32_t bzp = (int32_t) ks * (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
| 770 | do { |
| 771 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 772 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 773 | int32_t* packed_b = (int32_t*) packed_weights; |
| 774 | if XNN_LIKELY(b != NULL) { |
| 775 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 776 | unaligned_store_s32(packed_weights, bzp + b[nr_block_start + nr_block_offset]); |
| 777 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 778 | } |
| 779 | } else { |
| 780 | size_t n = nr_block_size; |
| 781 | do { |
| 782 | unaligned_store_s32(packed_weights, bzp); |
| 783 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 784 | } while (--n != 0); |
| 785 | } |
| 786 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 787 | |
| 788 | for (size_t ki = 0; ki < ks; ki++) { |
| 789 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 790 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 791 | int32_t ksum = 0; |
| 792 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 793 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 794 | if (kc_idx < kc) { |
| 795 | const uint8_t kv = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 796 | ksum += (int32_t) kv; |
| 797 | ((uint8_t*) packed_weights)[kr_block_offset] = kv; |
| 798 | } |
| 799 | } |
| 800 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 801 | packed_weights = (uint8_t*) packed_weights + kr; |
| 802 | } |
| 803 | packed_weights = (uint8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 804 | } |
| 805 | } |
| 806 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 807 | } |
| 808 | k += ks * kc * nc; |
| 809 | if XNN_UNPREDICTABLE(b != NULL) { |
| 810 | b += nc; |
| 811 | } |
| 812 | } while (--g != 0); |
| 813 | } |
| 814 | |
| 815 | void xnn_pack_qs8_conv_goki_w( |
| 816 | size_t g, |
| 817 | size_t nc, |
| 818 | size_t ks, |
| 819 | size_t kc, |
| 820 | size_t nr, |
| 821 | size_t kr, |
| 822 | size_t sr, |
| 823 | const int8_t* k, |
| 824 | const int32_t* b, |
| 825 | void* packed_weights, |
| 826 | size_t extra_bytes, |
| 827 | const struct xnn_qs8_packing_params* params) |
| 828 | { |
| 829 | assert(g != 0); |
| 830 | assert(nr >= sr); |
| 831 | assert(k != NULL); |
| 832 | assert(packed_weights != NULL); |
| 833 | |
| 834 | const size_t skr = sr * kr; |
| 835 | const uint32_t izp = (int32_t) params->input_zero_point; |
| 836 | do { |
| 837 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 838 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 839 | int32_t* packed_b = (int32_t*) packed_weights; |
| 840 | if XNN_LIKELY(b != NULL) { |
| 841 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 842 | unaligned_store_s32(packed_weights, b[nr_block_start + nr_block_offset]); |
| 843 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 844 | } |
| 845 | } else { |
| 846 | size_t n = nr_block_size; |
| 847 | do { |
| 848 | unaligned_store_s32(packed_weights, 0); |
| 849 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 850 | } while (--n != 0); |
| 851 | } |
| 852 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 853 | |
| 854 | for (size_t ki = 0; ki < ks; ki++) { |
| 855 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 856 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 857 | uint32_t ksum = 0; |
| 858 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 859 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 860 | if (kc_idx < kc) { |
| 861 | const int8_t kv = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 862 | ksum += (uint32_t) kv; |
| 863 | ((int8_t*) packed_weights)[kr_block_offset] = kv; |
| 864 | } |
| 865 | } |
| 866 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 867 | packed_weights = (int8_t*) packed_weights + kr; |
| 868 | } |
| 869 | packed_weights = (int8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 870 | } |
| 871 | } |
| 872 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 873 | } |
| 874 | k += ks * kc * nc; |
| 875 | if XNN_UNPREDICTABLE(b != NULL) { |
| 876 | b += nc; |
| 877 | } |
| 878 | } while (--g != 0); |
| 879 | } |
| 880 | |
| 881 | void xnn_pack_f32_conv_kgo_w( |
| 882 | size_t g, |
| 883 | size_t nc, |
| 884 | size_t ks, |
| 885 | size_t nr, |
| 886 | size_t kr, |
| 887 | size_t sr, |
| 888 | const float* k, |
| 889 | const float* b, |
| 890 | float* packed_weights, |
| 891 | size_t extra_bytes, |
| 892 | const void* params) |
| 893 | { |
| 894 | assert(g != 0); |
| 895 | assert(nr >= sr); |
| 896 | assert(k != NULL); |
| 897 | assert(packed_weights != NULL); |
| 898 | |
| 899 | for (size_t i = 0; i < g; i++) { |
| 900 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 901 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 902 | if XNN_LIKELY(b != NULL) { |
| 903 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 904 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 905 | } |
| 906 | } |
| 907 | packed_weights += nr; |
| 908 | |
| 909 | for (size_t ki = 0; ki < ks; ki++) { |
| 910 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 911 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 912 | packed_weights[nr_block_offset * kr] = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 913 | } |
| 914 | packed_weights += nr * kr; |
| 915 | } |
| 916 | } |
| 917 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 918 | } |
| 919 | k += nc; |
| 920 | if XNN_UNPREDICTABLE(b != NULL) { |
| 921 | b += nc; |
| 922 | } |
| 923 | } |
| 924 | } |
| 925 | |
| 926 | void xnn_pack_f16_conv_kgo_w( |
| 927 | size_t g, |
| 928 | size_t nc, |
| 929 | size_t ks, |
| 930 | size_t nr, |
| 931 | size_t kr, |
| 932 | size_t sr, |
| 933 | const uint16_t* k, |
| 934 | const uint16_t* b, |
| 935 | uint16_t* packed_weights, |
| 936 | size_t extra_bytes, |
| 937 | const void* params) |
| 938 | { |
| 939 | assert(g != 0); |
| 940 | assert(nr >= sr); |
| 941 | assert(k != NULL); |
| 942 | assert(packed_weights != NULL); |
| 943 | |
| 944 | for (size_t i = 0; i < g; i++) { |
| 945 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 946 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 947 | if XNN_LIKELY(b != NULL) { |
| 948 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 949 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 950 | } |
| 951 | } |
| 952 | packed_weights += nr; |
| 953 | |
| 954 | for (size_t ki = 0; ki < ks; ki++) { |
| 955 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 956 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 957 | packed_weights[nr_block_offset * kr] = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 958 | } |
| 959 | packed_weights += nr * kr; |
| 960 | } |
| 961 | } |
| 962 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 963 | } |
| 964 | k += nc; |
| 965 | if XNN_UNPREDICTABLE(b != NULL) { |
| 966 | b += nc; |
| 967 | } |
| 968 | } |
| 969 | } |
| 970 | |
| 971 | void xnn_pack_f32_to_f16_conv_kgo_w( |
| 972 | size_t g, |
| 973 | size_t nc, |
| 974 | size_t ks, |
| 975 | size_t nr, |
| 976 | size_t kr, |
| 977 | size_t sr, |
| 978 | const float* k, |
| 979 | const float* b, |
| 980 | uint16_t* packed_weights, |
| 981 | size_t extra_bytes, |
| 982 | const void* params) |
| 983 | { |
| 984 | assert(g != 0); |
| 985 | assert(nr >= sr); |
| 986 | assert(k != NULL); |
| 987 | assert(packed_weights != NULL); |
| 988 | |
| 989 | for (size_t i = 0; i < g; i++) { |
| 990 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 991 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 992 | if XNN_LIKELY(b != NULL) { |
| 993 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 994 | packed_weights[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 995 | } |
| 996 | } |
| 997 | packed_weights += nr; |
| 998 | |
| 999 | for (size_t ki = 0; ki < ks; ki++) { |
| 1000 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 1001 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 1002 | packed_weights[nr_block_offset * kr] = fp16_ieee_from_fp32_value(k[ki * g * nc + (nr_block_start + nr_block_offset)]); |
| 1003 | } |
| 1004 | packed_weights += nr * kr; |
| 1005 | } |
| 1006 | } |
| 1007 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 1008 | } |
| 1009 | k += nc; |
| 1010 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1011 | b += nc; |
| 1012 | } |
| 1013 | } |
| 1014 | } |
| 1015 | |
| 1016 | void xnn_pack_qu8_conv_kgo_w( |
| 1017 | size_t g, |
| 1018 | size_t nc, |
| 1019 | size_t ks, |
| 1020 | size_t nr, |
| 1021 | size_t kr, |
| 1022 | size_t sr, |
| 1023 | const uint8_t* k, |
| 1024 | const int32_t* b, |
| 1025 | void* packed_weights, |
| 1026 | size_t extra_bytes, |
| 1027 | const struct xnn_qu8_packing_params* params) |
| 1028 | { |
| 1029 | assert(g != 0); |
| 1030 | assert(nr >= sr); |
| 1031 | assert(k != NULL); |
| 1032 | assert(packed_weights != NULL); |
| 1033 | |
| 1034 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1035 | const int32_t bzp = (int32_t) ks * izp * (int32_t) params->kernel_zero_point; |
| 1036 | for (size_t i = 0; i < g; i++) { |
| 1037 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1038 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1039 | int32_t* packed_b = (int32_t*) packed_weights; |
| 1040 | if XNN_LIKELY(b != NULL) { |
| 1041 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1042 | unaligned_store_s32(packed_weights, bzp + b[nr_block_start + nr_block_offset]); |
| 1043 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1044 | } |
| 1045 | } else { |
| 1046 | size_t n = nr_block_size; |
| 1047 | do { |
| 1048 | unaligned_store_s32(packed_weights, bzp); |
| 1049 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1050 | } while (--n != 0); |
| 1051 | } |
| 1052 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 1053 | |
| 1054 | for (size_t ki = 0; ki < ks; ki++) { |
| 1055 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 1056 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 1057 | const uint8_t kv = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 1058 | ((uint8_t*) packed_weights)[nr_block_offset * kr] = kv; |
| 1059 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - (int32_t) kv * izp); |
| 1060 | } |
| 1061 | packed_weights = (uint8_t*) packed_weights + nr * kr; |
| 1062 | } |
| 1063 | } |
| 1064 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 1065 | } |
| 1066 | k += nc; |
| 1067 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1068 | b += nc; |
| 1069 | } |
| 1070 | } |
| 1071 | } |
| 1072 | |
| 1073 | void xnn_pack_qs8_conv_kgo_w( |
| 1074 | size_t g, |
| 1075 | size_t nc, |
| 1076 | size_t ks, |
| 1077 | size_t nr, |
| 1078 | size_t kr, |
| 1079 | size_t sr, |
| 1080 | const int8_t* k, |
| 1081 | const int32_t* b, |
| 1082 | void* packed_weights, |
| 1083 | size_t extra_bytes, |
| 1084 | const struct xnn_qs8_packing_params* params) |
| 1085 | { |
| 1086 | assert(g != 0); |
| 1087 | assert(nr >= sr); |
| 1088 | assert(k != NULL); |
| 1089 | assert(packed_weights != NULL); |
| 1090 | |
| 1091 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 1092 | for (size_t i = 0; i < g; i++) { |
| 1093 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1094 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1095 | int32_t* packed_b = (int32_t*) packed_weights; |
| 1096 | if XNN_LIKELY(b != NULL) { |
| 1097 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1098 | unaligned_store_s32(packed_weights, b[nr_block_start + nr_block_offset]); |
| 1099 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1100 | } |
| 1101 | } else { |
| 1102 | size_t n = nr_block_size; |
| 1103 | do { |
| 1104 | unaligned_store_s32(packed_weights, 0); |
| 1105 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1106 | } while (--n != 0); |
| 1107 | } |
| 1108 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 1109 | |
| 1110 | for (size_t ki = 0; ki < ks; ki++) { |
| 1111 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 1112 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 1113 | const int8_t kv = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 1114 | ((int8_t*) packed_weights)[nr_block_offset * kr] = kv; |
| 1115 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - (uint32_t) kv * izp); |
| 1116 | } |
| 1117 | packed_weights = (int8_t*) packed_weights + nr * kr; |
| 1118 | } |
| 1119 | } |
| 1120 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 1121 | } |
| 1122 | k += nc; |
| 1123 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1124 | b += nc; |
| 1125 | } |
| 1126 | } |
| 1127 | } |
| 1128 | |
| 1129 | void xnn_pack_f32_deconv_goki_w( |
| 1130 | size_t g, |
| 1131 | size_t nc, |
| 1132 | size_t kh, |
| 1133 | size_t kw, |
| 1134 | size_t kc, |
| 1135 | size_t sh, |
| 1136 | size_t sw, |
| 1137 | size_t nr, |
| 1138 | size_t kr, |
| 1139 | size_t sr, |
| 1140 | const float* k, |
| 1141 | const float* b, |
| 1142 | float* packed_weights, |
| 1143 | struct subconvolution_params* subconv_params, |
| 1144 | const void* params) |
| 1145 | { |
| 1146 | assert(g != 0); |
| 1147 | assert(nr >= sr); |
| 1148 | assert(k != NULL); |
| 1149 | assert(packed_weights != NULL); |
| 1150 | |
| 1151 | const size_t skr = sr * kr; |
| 1152 | for (size_t i = 0; i < g; i++) { |
| 1153 | for (size_t oy = 0; oy < sh; oy++) { |
| 1154 | for (size_t ox = 0; ox < sw; ox++) { |
| 1155 | if (i == 0) { |
| 1156 | (*subconv_params++).weights = packed_weights; |
| 1157 | } |
| 1158 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1159 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1160 | if XNN_LIKELY(b != NULL) { |
| 1161 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1162 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 1163 | } |
| 1164 | } |
| 1165 | packed_weights += nr; |
| 1166 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1167 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1168 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1169 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1170 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1171 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1172 | if (kc_idx < kc) { |
| 1173 | packed_weights[kr_block_offset] = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1174 | } |
| 1175 | } |
| 1176 | packed_weights += kr; |
| 1177 | } |
| 1178 | packed_weights += (nr - nr_block_size) * kr; |
| 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | } |
| 1183 | } |
| 1184 | } |
| 1185 | k += kh * kw * kc * nc; |
| 1186 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1187 | b += nc; |
| 1188 | } |
| 1189 | } |
| 1190 | } |
| 1191 | |
| 1192 | void xnn_pack_f16_deconv_goki_w( |
| 1193 | size_t g, |
| 1194 | size_t nc, |
| 1195 | size_t kh, |
| 1196 | size_t kw, |
| 1197 | size_t kc, |
| 1198 | size_t sh, |
| 1199 | size_t sw, |
| 1200 | size_t nr, |
| 1201 | size_t kr, |
| 1202 | size_t sr, |
| 1203 | const uint16_t* k, |
| 1204 | const uint16_t* b, |
| 1205 | uint16_t* packed_weights, |
| 1206 | struct subconvolution_params* subconv_params, |
| 1207 | const void* params) |
| 1208 | { |
| 1209 | assert(g != 0); |
| 1210 | assert(nr >= sr); |
| 1211 | assert(k != NULL); |
| 1212 | assert(packed_weights != NULL); |
| 1213 | |
| 1214 | const size_t skr = sr * kr; |
| 1215 | for (size_t i = 0; i < g; i++) { |
| 1216 | for (size_t oy = 0; oy < sh; oy++) { |
| 1217 | for (size_t ox = 0; ox < sw; ox++) { |
| 1218 | if (i == 0) { |
| 1219 | (*subconv_params++).weights = packed_weights; |
| 1220 | } |
| 1221 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1222 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1223 | if XNN_LIKELY(b != NULL) { |
| 1224 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1225 | packed_weights[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 1226 | } |
| 1227 | } |
| 1228 | packed_weights += nr; |
| 1229 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1230 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1231 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1232 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1233 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1234 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1235 | if (kc_idx < kc) { |
| 1236 | packed_weights[kr_block_offset] = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1237 | } |
| 1238 | } |
| 1239 | packed_weights += kr; |
| 1240 | } |
| 1241 | packed_weights += (nr - nr_block_size) * kr; |
| 1242 | } |
| 1243 | } |
| 1244 | } |
| 1245 | } |
| 1246 | } |
| 1247 | } |
| 1248 | k += kh * kw * kc * nc; |
| 1249 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1250 | b += nc; |
| 1251 | } |
| 1252 | } |
| 1253 | } |
| 1254 | |
| 1255 | void xnn_pack_f32_to_f16_deconv_goki_w( |
| 1256 | size_t g, |
| 1257 | size_t nc, |
| 1258 | size_t kh, |
| 1259 | size_t kw, |
| 1260 | size_t kc, |
| 1261 | size_t sh, |
| 1262 | size_t sw, |
| 1263 | size_t nr, |
| 1264 | size_t kr, |
| 1265 | size_t sr, |
| 1266 | const float* k, |
| 1267 | const float* b, |
| 1268 | uint16_t* packed_weights, |
| 1269 | struct subconvolution_params* subconv_params, |
| 1270 | const void* params) |
| 1271 | { |
| 1272 | assert(g != 0); |
| 1273 | assert(nr >= sr); |
| 1274 | assert(k != NULL); |
| 1275 | assert(packed_weights != NULL); |
| 1276 | |
| 1277 | const size_t skr = sr * kr; |
| 1278 | for (size_t i = 0; i < g; i++) { |
| 1279 | for (size_t oy = 0; oy < sh; oy++) { |
| 1280 | for (size_t ox = 0; ox < sw; ox++) { |
| 1281 | if (i == 0) { |
| 1282 | (*subconv_params++).weights = packed_weights; |
| 1283 | } |
| 1284 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1285 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1286 | if XNN_LIKELY(b != NULL) { |
| 1287 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1288 | packed_weights[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 1289 | } |
| 1290 | } |
| 1291 | packed_weights += nr; |
| 1292 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1293 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1294 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1295 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1296 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1297 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1298 | if (kc_idx < kc) { |
| 1299 | packed_weights[kr_block_offset] = fp16_ieee_from_fp32_value(k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]); |
| 1300 | } |
| 1301 | } |
| 1302 | packed_weights += kr; |
| 1303 | } |
| 1304 | packed_weights += (nr - nr_block_size) * kr; |
| 1305 | } |
| 1306 | } |
| 1307 | } |
| 1308 | } |
| 1309 | } |
| 1310 | } |
| 1311 | k += kh * kw * kc * nc; |
| 1312 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1313 | b += nc; |
| 1314 | } |
| 1315 | } |
| 1316 | } |
| 1317 | |
| 1318 | void xnn_pack_qs8_deconv_goki_w( |
| 1319 | size_t g, |
| 1320 | size_t nc, |
| 1321 | size_t kh, |
| 1322 | size_t kw, |
| 1323 | size_t kc, |
| 1324 | size_t sh, |
| 1325 | size_t sw, |
| 1326 | size_t nr, |
| 1327 | size_t kr, |
| 1328 | size_t sr, |
| 1329 | const int8_t* k, |
| 1330 | const int32_t* b, |
| 1331 | void* packed_weights, |
| 1332 | struct subconvolution_params* subconv_params, |
| 1333 | const struct xnn_qs8_packing_params* params) |
| 1334 | { |
| 1335 | assert(g != 0); |
| 1336 | assert(nr >= sr); |
| 1337 | assert(k != NULL); |
| 1338 | assert(packed_weights != NULL); |
| 1339 | |
| 1340 | const size_t skr = sr * kr; |
| 1341 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 1342 | for (size_t i = 0; i < g; i++) { |
| 1343 | for (size_t oy = 0; oy < sh; oy++) { |
| 1344 | for (size_t ox = 0; ox < sw; ox++) { |
| 1345 | if (i == 0) { |
| 1346 | (*subconv_params++).weights = packed_weights; |
| 1347 | } |
| 1348 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1349 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1350 | int32_t* packed_b = (int32_t*) packed_weights; |
| 1351 | if XNN_LIKELY(b != 0) { |
| 1352 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1353 | unaligned_store_s32(packed_weights, b[nr_block_start + nr_block_offset]); |
| 1354 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1355 | } |
| 1356 | } else { |
| 1357 | size_t n = nr_block_size; |
| 1358 | do { |
| 1359 | unaligned_store_s32(packed_weights, 0); |
| 1360 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1361 | } while (--n != 0); |
| 1362 | } |
| 1363 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 1364 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1365 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1366 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1367 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1368 | uint32_t ksum = 0; |
| 1369 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1370 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1371 | if (kc_idx < kc) { |
| 1372 | const int8_t kv = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1373 | ksum += (uint32_t) kv; |
| 1374 | ((int8_t*) packed_weights)[kr_block_offset] = kv; |
| 1375 | } |
| 1376 | } |
| 1377 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 1378 | packed_weights = (int8_t*) packed_weights + kr; |
| 1379 | } |
| 1380 | packed_weights = (int8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 1381 | } |
| 1382 | } |
| 1383 | } |
| 1384 | } |
| 1385 | } |
| 1386 | } |
| 1387 | k += kh * kw * kc * nc; |
| 1388 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1389 | b += nc; |
| 1390 | } |
| 1391 | } |
| 1392 | } |
| 1393 | |
| 1394 | void xnn_pack_qu8_deconv_goki_w( |
| 1395 | size_t g, |
| 1396 | size_t nc, |
| 1397 | size_t kh, |
| 1398 | size_t kw, |
| 1399 | size_t kc, |
| 1400 | size_t sh, |
| 1401 | size_t sw, |
| 1402 | size_t nr, |
| 1403 | size_t kr, |
| 1404 | size_t sr, |
| 1405 | const uint8_t* k, |
| 1406 | const int32_t* b, |
| 1407 | void* packed_weights, |
| 1408 | struct subconvolution_params* subconv_params, |
| 1409 | const struct xnn_qu8_packing_params* params) |
| 1410 | { |
| 1411 | assert(g != 0); |
| 1412 | assert(nr >= sr); |
| 1413 | assert(k != NULL); |
| 1414 | assert(packed_weights != NULL); |
| 1415 | |
| 1416 | const size_t skr = sr * kr; |
| 1417 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1418 | const int32_t kzp = (int32_t) params->kernel_zero_point; |
| 1419 | for (size_t i = 0; i < g; i++) { |
| 1420 | for (size_t oy = 0; oy < sh; oy++) { |
| 1421 | for (size_t ox = 0; ox < sw; ox++) { |
| 1422 | if (i == 0) { |
| 1423 | (*subconv_params++).weights = packed_weights; |
| 1424 | } |
| 1425 | const int32_t bzp = (int32_t) divide_round_up(kh - oy, sh) * (int32_t) divide_round_up(kw - ox, sw) * (int32_t) kc * izp * kzp; |
| 1426 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1427 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1428 | int32_t* packed_b = (int32_t*) packed_weights; |
| 1429 | if XNN_LIKELY(b != 0) { |
| 1430 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1431 | unaligned_store_s32(packed_weights, bzp + b[nr_block_start + nr_block_offset]); |
| 1432 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1433 | } |
| 1434 | } else { |
| 1435 | size_t n = nr_block_size; |
| 1436 | do { |
| 1437 | unaligned_store_s32(packed_weights, bzp); |
| 1438 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1439 | } while (--n != 0); |
| 1440 | } |
| 1441 | packed_weights = (void*) ((uintptr_t) packed_weights + (nr - nr_block_size) * sizeof(int32_t)); |
| 1442 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1443 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1444 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1445 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1446 | int32_t ksum = 0; |
| 1447 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1448 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1449 | if (kc_idx < kc) { |
| 1450 | const uint8_t kv = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1451 | ksum += (int32_t) kv; |
| 1452 | ((uint8_t*) packed_weights)[kr_block_offset] = kv; |
| 1453 | } |
| 1454 | } |
| 1455 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 1456 | packed_weights = (uint8_t*) packed_weights + kr; |
| 1457 | } |
| 1458 | packed_weights = (uint8_t*) packed_weights + (nr - nr_block_size) * kr; |
| 1459 | } |
| 1460 | } |
| 1461 | } |
| 1462 | } |
| 1463 | } |
| 1464 | } |
| 1465 | k += kh * kw * kc * nc; |
| 1466 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1467 | b += nc; |
| 1468 | } |
| 1469 | } |
| 1470 | } |
| 1471 | |
| 1472 | void xnn_pack_f32_dwconv_ghw_w( |
| 1473 | size_t primary_tile, |
| 1474 | size_t h, |
| 1475 | size_t w, |
| 1476 | size_t c, |
| 1477 | size_t cr, |
| 1478 | const float* k, |
| 1479 | const float* b, |
| 1480 | float* packed_weights, |
| 1481 | size_t extra_bytes, |
| 1482 | const void* params) |
| 1483 | { |
| 1484 | assert(primary_tile >= h * w); |
| 1485 | xnn_pack_f32_dwconv_multipass_ghw_w( |
| 1486 | primary_tile, |
| 1487 | /*middle_pass_tile=*/0, |
| 1488 | /*last_pass_tile=*/0, |
| 1489 | h, |
| 1490 | w, |
| 1491 | c, |
| 1492 | cr, |
| 1493 | cr, |
| 1494 | cr, |
| 1495 | k, |
| 1496 | b, |
| 1497 | packed_weights, |
| 1498 | extra_bytes, |
| 1499 | params); |
| 1500 | } |
| 1501 | |
| 1502 | // Helper function to advance x and y indices. |
| 1503 | inline static void advance_x_y(size_t h, size_t* x, size_t* y) { |
| 1504 | if (++*y == h) { |
| 1505 | *y = 0; |
| 1506 | ++*x; |
| 1507 | } |
| 1508 | } |
| 1509 | |
| 1510 | void xnn_pack_f32_dwconv_multipass_ghw_w( |
| 1511 | size_t first_pass_tile, |
| 1512 | size_t middle_pass_tile, |
| 1513 | size_t last_pass_tile, |
| 1514 | size_t h, |
| 1515 | size_t w, |
| 1516 | size_t c, |
| 1517 | size_t channel_tile, |
| 1518 | size_t channel_subtile, |
| 1519 | size_t channel_round, |
| 1520 | const float* k, |
| 1521 | const float* b, |
| 1522 | float* packed_weights, |
| 1523 | size_t extra_bytes, |
| 1524 | const void* params) |
| 1525 | { |
| 1526 | assert(k != NULL); |
| 1527 | assert(packed_weights != NULL); |
| 1528 | size_t kernel_size = h * w; |
| 1529 | if (middle_pass_tile == 0) { |
| 1530 | // Uni-pass DWCONV. |
| 1531 | assert(last_pass_tile == 0); |
| 1532 | } else { |
| 1533 | // Multi-pass DWCONV. |
| 1534 | assert(kernel_size > first_pass_tile); |
| 1535 | } |
| 1536 | |
| 1537 | // Stores the x and y index that should be processed next. |
| 1538 | size_t processed_x = 0; |
| 1539 | size_t processed_y = 0; |
| 1540 | size_t x = 0; |
| 1541 | size_t y = 0; |
| 1542 | // First and middle pass packs in sizes of channel_tile to tiled_c, then in sizes of channel_subtile. |
| 1543 | const size_t tiled_c = round_down_po2(round_up_po2(c, channel_round), channel_tile); |
| 1544 | |
| 1545 | // Pack in blocks of channel_tile, then in blocks of channel_subtile. |
| 1546 | { |
| 1547 | size_t cr_block_start = 0; |
| 1548 | for (; cr_block_start < tiled_c; cr_block_start += channel_tile) { |
| 1549 | const size_t cr_block_size = min(c - cr_block_start, channel_tile); |
| 1550 | if XNN_LIKELY(b != NULL) { |
| 1551 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1552 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 1553 | } |
| 1554 | } else { |
| 1555 | size_t n = cr_block_size; |
| 1556 | do { |
| 1557 | *packed_weights++ = 0.0f; |
| 1558 | } while (--n != 0); |
| 1559 | } |
| 1560 | packed_weights += channel_tile - cr_block_size; |
| 1561 | |
| 1562 | x = 0; |
| 1563 | y = 0; |
| 1564 | // kernel_size can be less than the first_pass_tile, in this case, pack up |
| 1565 | // to the smaller of the two. |
| 1566 | for (size_t i = 0; i < min(first_pass_tile, kernel_size); i++) { |
| 1567 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1568 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1569 | *packed_weights++ = kv; |
| 1570 | } |
| 1571 | packed_weights += channel_tile - cr_block_size; |
| 1572 | advance_x_y(h, &x, &y); |
| 1573 | } |
| 1574 | // And make sure to skip weights if kernel_size < first_pass_tile. |
| 1575 | packed_weights += doz(first_pass_tile, kernel_size) * cr_block_size; |
| 1576 | } |
| 1577 | |
| 1578 | for (; cr_block_start < c; cr_block_start += channel_subtile) { |
| 1579 | const size_t cr_block_size = min(c - cr_block_start, channel_subtile); |
| 1580 | if XNN_LIKELY(b != NULL) { |
| 1581 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1582 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 1583 | } |
| 1584 | } else { |
| 1585 | size_t n = cr_block_size; |
| 1586 | do { |
| 1587 | *packed_weights++ = 0.0f; |
| 1588 | } while (--n != 0); |
| 1589 | } |
| 1590 | packed_weights += channel_subtile - cr_block_size; |
| 1591 | |
| 1592 | x = 0; |
| 1593 | y = 0; |
| 1594 | // kernel_size can be less than the first_pass_tile, in this case, pack up |
| 1595 | // to the smaller of the two. |
| 1596 | for (size_t i = 0; i < min(first_pass_tile, kernel_size); i++) { |
| 1597 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1598 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1599 | *packed_weights++ = kv; |
| 1600 | } |
| 1601 | packed_weights += channel_subtile - cr_block_size; |
| 1602 | advance_x_y(h, &x, &y); |
| 1603 | } |
| 1604 | // And make sure to skip weights if kernel_size < first_pass_tile. |
| 1605 | packed_weights += doz(first_pass_tile, kernel_size) * cr_block_size; |
| 1606 | } |
| 1607 | } |
| 1608 | |
| 1609 | if (kernel_size <= first_pass_tile) { |
| 1610 | return; |
| 1611 | } |
| 1612 | |
| 1613 | kernel_size -= first_pass_tile; |
| 1614 | |
| 1615 | processed_x = x; |
| 1616 | processed_y = y; |
| 1617 | |
| 1618 | // Middle pass. (kernel_size / middle_pass_tile) blocks, within each block is |
| 1619 | // middle_pass_tile * cr weights. |
| 1620 | for (; kernel_size > last_pass_tile; kernel_size -= middle_pass_tile) { |
| 1621 | assert(kernel_size >= middle_pass_tile); |
| 1622 | size_t cr_block_start = 0; |
| 1623 | for (; cr_block_start < tiled_c; cr_block_start += channel_tile) { |
| 1624 | x = processed_x; |
| 1625 | y = processed_y; |
| 1626 | const size_t cr_block_size = min(c - cr_block_start, channel_tile); |
| 1627 | for (size_t j = 0; j < middle_pass_tile; j++) { |
| 1628 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1629 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1630 | *packed_weights++ = kv; |
| 1631 | } |
| 1632 | packed_weights += channel_tile - cr_block_size; |
| 1633 | advance_x_y(h, &x, &y); |
| 1634 | } |
| 1635 | } |
| 1636 | for (; cr_block_start < c; cr_block_start += channel_subtile) { |
| 1637 | x = processed_x; |
| 1638 | y = processed_y; |
| 1639 | const size_t cr_block_size = min(c - cr_block_start, channel_subtile); |
| 1640 | for (size_t j = 0; j < middle_pass_tile; j++) { |
| 1641 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1642 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1643 | *packed_weights++ = kv; |
| 1644 | } |
| 1645 | packed_weights += channel_subtile - cr_block_size; |
| 1646 | advance_x_y(h, &x, &y); |
| 1647 | } |
| 1648 | } |
| 1649 | processed_x = x; |
| 1650 | processed_y = y; |
| 1651 | } |
| 1652 | |
| 1653 | // Last pass. |
| 1654 | { |
| 1655 | assert(kernel_size <= last_pass_tile); |
| 1656 | size_t cr_block_start = 0; |
| 1657 | for (; cr_block_start < round_down_po2(c, channel_tile); cr_block_start += channel_tile) { |
| 1658 | // Last pass does not pack to rounded c, since it handles remainder. |
| 1659 | x = processed_x; |
| 1660 | y = processed_y; |
| 1661 | const size_t cr_block_size = min(c - cr_block_start, channel_tile); |
| 1662 | for (size_t i = 0; i < kernel_size; i++) { |
| 1663 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1664 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1665 | *packed_weights++ = kv; |
| 1666 | } |
| 1667 | packed_weights += channel_tile - cr_block_size; |
| 1668 | advance_x_y(h, &x, &y); |
| 1669 | } |
| 1670 | // Pad so that we can always read last_pass_tile weights in the last pass. |
| 1671 | packed_weights += (last_pass_tile - kernel_size) * channel_tile; |
| 1672 | // TODO(zhin): support extra bytes for channel_tile and subtile. |
| 1673 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 1674 | } |
| 1675 | for (; cr_block_start < c; cr_block_start += channel_subtile) { |
| 1676 | // Last pass does not pack to rounded c, since it handles remainder. |
| 1677 | x = processed_x; |
| 1678 | y = processed_y; |
| 1679 | const size_t cr_block_size = min(c - cr_block_start, channel_subtile); |
| 1680 | for (size_t i = 0; i < kernel_size; i++) { |
| 1681 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1682 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1683 | *packed_weights++ = kv; |
| 1684 | } |
| 1685 | packed_weights += channel_subtile - cr_block_size; |
| 1686 | advance_x_y(h, &x, &y); |
| 1687 | } |
| 1688 | // Pad so that we can always read last_pass_tile weights in the last pass. |
| 1689 | packed_weights += (last_pass_tile - kernel_size) * channel_subtile; |
| 1690 | // TODO(zhin): support extra bytes for channel_tile and subtile. |
| 1691 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 1692 | } |
| 1693 | } |
| 1694 | } |
| 1695 | |
| 1696 | void xnn_pack_f16_dwconv_ghw_w( |
| 1697 | size_t primary_tile, |
| 1698 | size_t h, |
| 1699 | size_t w, |
| 1700 | size_t c, |
| 1701 | size_t cr, |
| 1702 | const uint16_t* k, |
| 1703 | const uint16_t* b, |
| 1704 | uint16_t* packed_weights, |
| 1705 | size_t extra_bytes, |
| 1706 | const void* params) |
| 1707 | { |
| 1708 | assert(k != NULL); |
| 1709 | assert(packed_weights != NULL); |
| 1710 | |
| 1711 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1712 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1713 | if XNN_LIKELY(b != NULL) { |
| 1714 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1715 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 1716 | } |
| 1717 | } else { |
| 1718 | size_t n = cr_block_size; |
| 1719 | do { |
| 1720 | *packed_weights++ = 0; |
| 1721 | } while (--n != 0); |
| 1722 | } |
| 1723 | packed_weights += cr - cr_block_size; |
| 1724 | for (size_t x = 0; x < w; x++) { |
| 1725 | for (size_t y = 0; y < h; y++) { |
| 1726 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1727 | const uint16_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1728 | *packed_weights++ = kv; |
| 1729 | } |
| 1730 | packed_weights += cr - cr_block_size; |
| 1731 | } |
| 1732 | } |
| 1733 | packed_weights += (primary_tile - (h * w)) * cr_block_size; |
| 1734 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 1735 | } |
| 1736 | } |
| 1737 | |
| 1738 | void xnn_pack_f32_to_f16_dwconv_ghw_w( |
| 1739 | size_t primary_tile, |
| 1740 | size_t h, |
| 1741 | size_t w, |
| 1742 | size_t c, |
| 1743 | size_t cr, |
| 1744 | const float* k, |
| 1745 | const float* b, |
| 1746 | uint16_t* packed_weights, |
| 1747 | size_t extra_bytes, |
| 1748 | const void* params) |
| 1749 | { |
| 1750 | assert(k != NULL); |
| 1751 | assert(packed_weights != NULL); |
| 1752 | |
| 1753 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1754 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1755 | if XNN_LIKELY(b != NULL) { |
| 1756 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1757 | *packed_weights++ = fp16_ieee_from_fp32_value(b[cr_block_start + cr_block_offset]); |
| 1758 | } |
| 1759 | } else { |
| 1760 | size_t n = cr_block_size; |
| 1761 | do { |
| 1762 | *packed_weights++ = 0; |
| 1763 | } while (--n != 0); |
| 1764 | } |
| 1765 | packed_weights += cr - cr_block_size; |
| 1766 | for (size_t x = 0; x < w; x++) { |
| 1767 | for (size_t y = 0; y < h; y++) { |
| 1768 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1769 | const uint16_t kv = fp16_ieee_from_fp32_value(k[((cr_block_start + cr_block_offset) * h + y) * w + x]); |
| 1770 | *packed_weights++ = kv; |
| 1771 | } |
| 1772 | packed_weights += cr - cr_block_size; |
| 1773 | } |
| 1774 | } |
| 1775 | packed_weights += (primary_tile - (h * w)) * cr_block_size; |
| 1776 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 1777 | } |
| 1778 | } |
| 1779 | |
| 1780 | void xnn_pack_qu8_dwconv_ghw_w( |
| 1781 | size_t primary_tile, |
| 1782 | size_t h, |
| 1783 | size_t w, |
| 1784 | size_t c, |
| 1785 | size_t cr, |
| 1786 | const uint8_t* k, |
| 1787 | const int32_t* b, |
| 1788 | void* packed_weights, |
| 1789 | size_t extra_bytes, |
| 1790 | const struct xnn_qu8_packing_params* params) |
| 1791 | { |
| 1792 | assert(k != NULL); |
| 1793 | assert(packed_weights != NULL); |
| 1794 | |
| 1795 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1796 | const int32_t boff = (int32_t) h * (int32_t) w * izp * (int32_t) params->kernel_zero_point; |
| 1797 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1798 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1799 | int32_t* packed_b = (int32_t*) packed_weights; |
| 1800 | if XNN_LIKELY(b != NULL) { |
| 1801 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1802 | unaligned_store_s32(packed_weights, boff + b[cr_block_start + cr_block_offset]); |
| 1803 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1804 | } |
| 1805 | } else { |
| 1806 | size_t n = cr_block_size; |
| 1807 | do { |
| 1808 | unaligned_store_s32(packed_weights, boff); |
| 1809 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1810 | } while (--n != 0); |
| 1811 | } |
| 1812 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(int32_t)); |
| 1813 | for (size_t x = 0; x < w; x++) { |
| 1814 | for (size_t y = 0; y < h; y++) { |
| 1815 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1816 | const uint8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1817 | unaligned_indexed_store_s32(packed_b, cr_block_offset, unaligned_indexed_load_s32(packed_b, cr_block_offset) - (int32_t) kv * izp); |
| 1818 | *((uint8_t*) packed_weights) = kv; |
| 1819 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(uint8_t)); |
| 1820 | } |
| 1821 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(uint8_t)); |
| 1822 | } |
| 1823 | } |
| 1824 | packed_weights = (void*) ((uintptr_t) packed_weights + (primary_tile - (h * w)) * cr_block_size * sizeof(uint8_t)); |
| 1825 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 1826 | } |
| 1827 | } |
| 1828 | |
| 1829 | void xnn_pack_qs8_dwconv_ghw_w( |
| 1830 | size_t primary_tile, |
| 1831 | size_t h, |
| 1832 | size_t w, |
| 1833 | size_t c, |
| 1834 | size_t cr, |
| 1835 | const int8_t* k, |
| 1836 | const int32_t* b, |
| 1837 | void* packed_weights, |
| 1838 | size_t extra_bytes, |
| 1839 | const struct xnn_qs8_packing_params* params) |
| 1840 | { |
| 1841 | assert(k != NULL); |
| 1842 | assert(packed_weights != NULL); |
| 1843 | |
| 1844 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 1845 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1846 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1847 | int32_t* packed_b = (int32_t*) packed_weights; |
| 1848 | if XNN_LIKELY(b != NULL) { |
| 1849 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1850 | unaligned_store_s32(packed_weights, b[cr_block_start + cr_block_offset]); |
| 1851 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1852 | } |
| 1853 | } else { |
| 1854 | size_t n = cr_block_size; |
| 1855 | do { |
| 1856 | unaligned_store_s32(packed_weights, 0); |
| 1857 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 1858 | } while (--n != 0); |
| 1859 | } |
| 1860 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(int32_t)); |
| 1861 | for (size_t x = 0; x < w; x++) { |
| 1862 | for (size_t y = 0; y < h; y++) { |
| 1863 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1864 | const int8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1865 | unaligned_indexed_store_u32(packed_b, cr_block_offset, unaligned_indexed_load_u32(packed_b, cr_block_offset) - (uint32_t) kv * izp); |
| 1866 | *((int8_t*) packed_weights) = kv; |
| 1867 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int8_t)); |
| 1868 | } |
| 1869 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(int8_t)); |
| 1870 | } |
| 1871 | } |
| 1872 | packed_weights = (void*) ((uintptr_t) packed_weights + (primary_tile - (h * w)) * cr_block_size * sizeof(int8_t)); |
| 1873 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 1874 | } |
| 1875 | } |
| 1876 | |
| 1877 | void xnn_pack_f32_dwconv_hwg_w( |
| 1878 | size_t primary_tile, |
| 1879 | size_t h, |
| 1880 | size_t w, |
| 1881 | size_t c, |
| 1882 | size_t cr, |
| 1883 | const float* k, |
| 1884 | const float* b, |
| 1885 | float* packed_weights, |
| 1886 | size_t extra_bytes, |
| 1887 | const void* params) |
| 1888 | { |
| 1889 | assert(primary_tile >= h * w); |
| 1890 | xnn_pack_f32_dwconv_multipass_hwg_w( |
| 1891 | primary_tile, |
| 1892 | /*middle_pass_tile=*/0, |
| 1893 | /*last_pass_tile=*/0, |
| 1894 | h, |
| 1895 | w, |
| 1896 | c, |
| 1897 | cr, |
| 1898 | cr, |
| 1899 | cr, |
| 1900 | k, |
| 1901 | b, |
| 1902 | packed_weights, |
| 1903 | extra_bytes, |
| 1904 | params); |
| 1905 | } |
| 1906 | |
| 1907 | void xnn_pack_f32_dwconv_multipass_hwg_w( |
| 1908 | size_t first_pass_tile, |
| 1909 | size_t middle_pass_tile, |
| 1910 | size_t last_pass_tile, |
| 1911 | size_t h, |
| 1912 | size_t w, |
| 1913 | size_t c, |
| 1914 | size_t channel_tile, |
| 1915 | size_t channel_subtile, |
| 1916 | size_t channel_round, |
| 1917 | const float* k, |
| 1918 | const float* b, |
| 1919 | float* packed_weights, |
| 1920 | size_t extra_bytes, |
| 1921 | const void* params) |
| 1922 | { |
| 1923 | assert(k != NULL); |
| 1924 | assert(packed_weights != NULL); |
| 1925 | size_t kernel_size = h * w; |
| 1926 | if (middle_pass_tile == 0) { |
| 1927 | // Uni-pass DWCONV. |
| 1928 | assert(last_pass_tile == 0); |
| 1929 | } else { |
| 1930 | // Multi-pass DWCONV. |
| 1931 | assert(kernel_size > first_pass_tile); |
| 1932 | } |
| 1933 | |
| 1934 | // Stores the x and y index that should be processed next. |
| 1935 | size_t processed_x = 0; |
| 1936 | size_t processed_y = 0; |
| 1937 | size_t x = 0; |
| 1938 | size_t y = 0; |
| 1939 | // First and middle pass packs in sizes of channel_tile to tiled_c, then in sizes of channel_subtile. |
| 1940 | const size_t tiled_c = round_down_po2(round_up_po2(c, channel_round), channel_tile); |
| 1941 | |
| 1942 | // Pack in blocks of channel_tile, then in blocks of channel_subtile. |
| 1943 | { |
| 1944 | size_t cr_block_start = 0; |
| 1945 | for (; cr_block_start < tiled_c; cr_block_start += channel_tile) { |
| 1946 | const size_t cr_block_size = min(c - cr_block_start, channel_tile); |
| 1947 | if XNN_LIKELY(b != NULL) { |
| 1948 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1949 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 1950 | } |
| 1951 | } else { |
| 1952 | size_t n = cr_block_size; |
| 1953 | do { |
| 1954 | *packed_weights++ = 0.0f; |
| 1955 | } while (--n != 0); |
| 1956 | } |
| 1957 | packed_weights += channel_tile - cr_block_size; |
| 1958 | |
| 1959 | x = processed_x; |
| 1960 | y = processed_y; |
| 1961 | // kernel_size can be less than the first_pass_tile, in this case, pack up |
| 1962 | // to the smaller of the two. |
| 1963 | for (size_t i = 0; i < min(first_pass_tile, kernel_size); i++) { |
| 1964 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1965 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1966 | *packed_weights++ = kv; |
| 1967 | } |
| 1968 | packed_weights += channel_tile - cr_block_size; |
| 1969 | if (++y == h) { |
| 1970 | y = 0; |
| 1971 | x++; |
| 1972 | } |
| 1973 | } |
| 1974 | // And make sure to skip weights if kernel_size < first_pass_tile. |
| 1975 | packed_weights += doz(first_pass_tile, kernel_size) * cr_block_size; |
| 1976 | } |
| 1977 | for (; cr_block_start < c; cr_block_start += channel_subtile) { |
| 1978 | const size_t cr_block_size = min(c - cr_block_start, channel_subtile); |
| 1979 | if XNN_LIKELY(b != NULL) { |
| 1980 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1981 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 1982 | } |
| 1983 | } else { |
| 1984 | size_t n = cr_block_size; |
| 1985 | do { |
| 1986 | *packed_weights++ = 0.0f; |
| 1987 | } while (--n != 0); |
| 1988 | } |
| 1989 | packed_weights += channel_subtile - cr_block_size; |
| 1990 | |
| 1991 | x = processed_x; |
| 1992 | y = processed_y; |
| 1993 | // kernel_size can be less than the first_pass_tile, in this case, pack up |
| 1994 | // to the smaller of the two. |
| 1995 | for (size_t i = 0; i < min(first_pass_tile, kernel_size); i++) { |
| 1996 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1997 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1998 | *packed_weights++ = kv; |
| 1999 | } |
| 2000 | packed_weights += channel_subtile - cr_block_size; |
| 2001 | if (++y == h) { |
| 2002 | y = 0; |
| 2003 | x++; |
| 2004 | } |
| 2005 | } |
| 2006 | // And make sure to skip weights if kernel_size < first_pass_tile. |
| 2007 | packed_weights += doz(first_pass_tile, kernel_size) * cr_block_size; |
| 2008 | } |
| 2009 | } |
| 2010 | |
| 2011 | if (kernel_size <= first_pass_tile) { |
| 2012 | return; |
| 2013 | } |
| 2014 | |
| 2015 | kernel_size -= first_pass_tile; |
| 2016 | |
| 2017 | processed_x = x; |
| 2018 | processed_y = y; |
| 2019 | |
| 2020 | // Middle pass. (kernel_size / middle_pass_tile) blocks, within each block is |
| 2021 | // middle_pass_tile * cr weights. |
| 2022 | for (; kernel_size > last_pass_tile; kernel_size -= middle_pass_tile) { |
| 2023 | assert(kernel_size >= middle_pass_tile); |
| 2024 | size_t cr_block_start = 0; |
| 2025 | for (; cr_block_start < tiled_c; cr_block_start += channel_tile) { |
| 2026 | x = processed_x; |
| 2027 | y = processed_y; |
| 2028 | const size_t cr_block_size = min(c - cr_block_start, channel_tile); |
| 2029 | for (size_t j = 0; j < middle_pass_tile; j++) { |
| 2030 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2031 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2032 | *packed_weights++ = kv; |
| 2033 | } |
| 2034 | packed_weights += channel_tile - cr_block_size; |
| 2035 | if (++y == h) { |
| 2036 | y = 0; |
| 2037 | x++; |
| 2038 | } |
| 2039 | } |
| 2040 | } |
| 2041 | for (; cr_block_start < c; cr_block_start += channel_subtile) { |
| 2042 | x = processed_x; |
| 2043 | y = processed_y; |
| 2044 | const size_t cr_block_size = min(c - cr_block_start, channel_subtile); |
| 2045 | for (size_t j = 0; j < middle_pass_tile; j++) { |
| 2046 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2047 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2048 | *packed_weights++ = kv; |
| 2049 | } |
| 2050 | packed_weights += channel_subtile - cr_block_size; |
| 2051 | if (++y == h) { |
| 2052 | y = 0; |
| 2053 | x++; |
| 2054 | } |
| 2055 | } |
| 2056 | } |
| 2057 | processed_x = x; |
| 2058 | processed_y = y; |
| 2059 | } |
| 2060 | |
| 2061 | // Last pass. |
| 2062 | { |
| 2063 | assert(kernel_size <= last_pass_tile); |
| 2064 | size_t cr_block_start = 0; |
| 2065 | for (; cr_block_start < round_down_po2(c, channel_tile); cr_block_start += channel_tile) { |
| 2066 | // for (; cr_block_start <= c - channel_tile; cr_block_start += channel_tile) { |
| 2067 | x = processed_x; |
| 2068 | y = processed_y; |
| 2069 | const size_t cr_block_size = min(c - cr_block_start, channel_tile); |
| 2070 | for (size_t i = 0; i < kernel_size; i++) { |
| 2071 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2072 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2073 | *packed_weights++ = kv; |
| 2074 | } |
| 2075 | packed_weights += channel_tile - cr_block_size; |
| 2076 | if (++y == h) { |
| 2077 | y = 0; |
| 2078 | x++; |
| 2079 | } |
| 2080 | } |
| 2081 | // Pad so that we can always read last_pass_tile weights in the last pass. |
| 2082 | packed_weights += (last_pass_tile - kernel_size) * channel_tile; |
| 2083 | // TODO(zhin): support extra bytes for channel_tile and subtile. |
| 2084 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 2085 | } |
| 2086 | for (; cr_block_start < c; cr_block_start += channel_subtile) { |
| 2087 | x = processed_x; |
| 2088 | y = processed_y; |
| 2089 | const size_t cr_block_size = min(c - cr_block_start, channel_subtile); |
| 2090 | for (size_t i = 0; i < kernel_size; i++) { |
| 2091 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2092 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2093 | *packed_weights++ = kv; |
| 2094 | } |
| 2095 | packed_weights += channel_subtile - cr_block_size; |
| 2096 | if (++y == h) { |
| 2097 | y = 0; |
| 2098 | x++; |
| 2099 | } |
| 2100 | } |
| 2101 | // Pad so that we can always read last_pass_tile weights in the last pass. |
| 2102 | packed_weights += (last_pass_tile - kernel_size) * channel_subtile; |
| 2103 | // TODO(zhin): support extra bytes for channel_tile and subtile. |
| 2104 | packed_weights = (float*) ((uintptr_t) packed_weights + extra_bytes); |
| 2105 | } |
| 2106 | } |
| 2107 | } |
| 2108 | |
| 2109 | void xnn_pack_f16_dwconv_hwg_w( |
| 2110 | size_t primary_tile, |
| 2111 | size_t h, |
| 2112 | size_t w, |
| 2113 | size_t c, |
| 2114 | size_t cr, |
| 2115 | const uint16_t* k, |
| 2116 | const uint16_t* b, |
| 2117 | uint16_t* packed_weights, |
| 2118 | size_t extra_bytes, |
| 2119 | const void* params) |
| 2120 | { |
| 2121 | assert(k != NULL); |
| 2122 | assert(packed_weights != NULL); |
| 2123 | |
| 2124 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2125 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2126 | if XNN_LIKELY(b != NULL) { |
| 2127 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2128 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 2129 | } |
| 2130 | } else { |
| 2131 | size_t n = cr_block_size; |
| 2132 | do { |
| 2133 | *packed_weights++ = 0; |
| 2134 | } while (--n != 0); |
| 2135 | } |
| 2136 | packed_weights += cr - cr_block_size; |
| 2137 | for (size_t x = 0; x < w; x++) { |
| 2138 | for (size_t y = 0; y < h; y++) { |
| 2139 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2140 | const uint16_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2141 | *packed_weights++ = kv; |
| 2142 | } |
| 2143 | packed_weights += cr - cr_block_size; |
| 2144 | } |
| 2145 | } |
| 2146 | packed_weights += (primary_tile - (h * w)) * cr_block_size; |
| 2147 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 2148 | } |
| 2149 | } |
| 2150 | |
| 2151 | void xnn_pack_f32_to_f16_dwconv_hwg_w( |
| 2152 | size_t primary_tile, |
| 2153 | size_t h, |
| 2154 | size_t w, |
| 2155 | size_t c, |
| 2156 | size_t cr, |
| 2157 | const float* k, |
| 2158 | const float* b, |
| 2159 | uint16_t* packed_weights, |
| 2160 | size_t extra_bytes, |
| 2161 | const void* params) |
| 2162 | { |
| 2163 | assert(k != NULL); |
| 2164 | assert(packed_weights != NULL); |
| 2165 | |
| 2166 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2167 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2168 | if XNN_LIKELY(b != NULL) { |
| 2169 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2170 | *packed_weights++ = fp16_ieee_from_fp32_value(b[cr_block_start + cr_block_offset]); |
| 2171 | } |
| 2172 | } else { |
| 2173 | size_t n = cr_block_size; |
| 2174 | do { |
| 2175 | *packed_weights++ = 0; |
| 2176 | } while (--n != 0); |
| 2177 | } |
| 2178 | packed_weights += cr - cr_block_size; |
| 2179 | for (size_t x = 0; x < w; x++) { |
| 2180 | for (size_t y = 0; y < h; y++) { |
| 2181 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2182 | const uint16_t kv = fp16_ieee_from_fp32_value(k[(y * w + x) * c + (cr_block_start + cr_block_offset)]); |
| 2183 | *packed_weights++ = kv; |
| 2184 | } |
| 2185 | packed_weights += cr - cr_block_size; |
| 2186 | } |
| 2187 | } |
| 2188 | packed_weights += (primary_tile - (h * w)) * cr_block_size; |
| 2189 | packed_weights = (uint16_t*) ((uintptr_t) packed_weights + extra_bytes); |
| 2190 | } |
| 2191 | } |
| 2192 | |
| 2193 | void xnn_pack_qu8_dwconv_hwg_w( |
| 2194 | size_t primary_tile, |
| 2195 | size_t h, |
| 2196 | size_t w, |
| 2197 | size_t c, |
| 2198 | size_t cr, |
| 2199 | const uint8_t* k, |
| 2200 | const int32_t* b, |
| 2201 | void* packed_weights, |
| 2202 | size_t extra_bytes, |
| 2203 | const struct xnn_qu8_packing_params* params) |
| 2204 | { |
| 2205 | assert(k != NULL); |
| 2206 | assert(packed_weights != NULL); |
| 2207 | |
| 2208 | const int32_t izp = (int32_t) params->input_zero_point; |
| 2209 | const int32_t boff = (int32_t) h * (int32_t) w * izp * (int32_t) params->kernel_zero_point; |
| 2210 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2211 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2212 | int32_t* packed_b = (int32_t*) packed_weights; |
| 2213 | if XNN_LIKELY(b != NULL) { |
| 2214 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2215 | unaligned_store_s32(packed_weights, boff + b[cr_block_start + cr_block_offset]); |
| 2216 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 2217 | } |
| 2218 | } else { |
| 2219 | size_t n = cr_block_size; |
| 2220 | do { |
| 2221 | unaligned_store_s32(packed_weights, boff); |
| 2222 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 2223 | } while (--n != 0); |
| 2224 | } |
| 2225 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(int32_t)); |
| 2226 | for (size_t x = 0; x < w; x++) { |
| 2227 | for (size_t y = 0; y < h; y++) { |
| 2228 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2229 | const uint8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2230 | unaligned_indexed_store_s32(packed_b, cr_block_offset, unaligned_indexed_load_s32(packed_b, cr_block_offset) - (int32_t) kv * izp); |
| 2231 | *((uint8_t*) packed_weights) = kv; |
| 2232 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(uint8_t)); |
| 2233 | } |
| 2234 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(uint8_t)); |
| 2235 | } |
| 2236 | } |
| 2237 | packed_weights = (void*) ((uintptr_t) packed_weights + (primary_tile - (h * w)) * cr_block_size * sizeof(uint8_t)); |
| 2238 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 2239 | } |
| 2240 | } |
| 2241 | |
| 2242 | void xnn_pack_qs8_dwconv_hwg_w( |
| 2243 | size_t primary_tile, |
| 2244 | size_t h, |
| 2245 | size_t w, |
| 2246 | size_t c, |
| 2247 | size_t cr, |
| 2248 | const int8_t* k, |
| 2249 | const int32_t* b, |
| 2250 | void* packed_weights, |
| 2251 | size_t extra_bytes, |
| 2252 | const struct xnn_qs8_packing_params* params) |
| 2253 | { |
| 2254 | assert(k != NULL); |
| 2255 | assert(packed_weights != NULL); |
| 2256 | |
| 2257 | const uint32_t izp = (int32_t) params->input_zero_point; |
| 2258 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2259 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2260 | int32_t* packed_b = (int32_t*) packed_weights; |
| 2261 | if XNN_LIKELY(b != NULL) { |
| 2262 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2263 | unaligned_store_s32(packed_weights, b[cr_block_start + cr_block_offset]); |
| 2264 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 2265 | } |
| 2266 | } else { |
| 2267 | size_t n = cr_block_size; |
| 2268 | do { |
| 2269 | unaligned_store_s32(packed_weights, 0); |
| 2270 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int32_t)); |
| 2271 | } while (--n != 0); |
| 2272 | } |
| 2273 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(int32_t)); |
| 2274 | for (size_t x = 0; x < w; x++) { |
| 2275 | for (size_t y = 0; y < h; y++) { |
| 2276 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2277 | const int8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 2278 | unaligned_indexed_store_u32(packed_b, cr_block_offset, unaligned_indexed_load_u32(packed_b, cr_block_offset) - (uint32_t) kv * izp); |
| 2279 | *((int8_t*) packed_weights) = kv; |
| 2280 | packed_weights = (void*) ((uintptr_t) packed_weights + sizeof(int8_t)); |
| 2281 | } |
| 2282 | packed_weights = (void*) ((uintptr_t) packed_weights + (cr - cr_block_size) * sizeof(int8_t)); |
| 2283 | } |
| 2284 | } |
| 2285 | packed_weights = (void*) ((uintptr_t) packed_weights + (primary_tile - (h * w)) * cr_block_size * sizeof(int8_t)); |
| 2286 | packed_weights = (void*) ((uintptr_t) packed_weights + extra_bytes); |
| 2287 | } |
| 2288 | } |
| 2289 | |
| 2290 | void xnn_pack_f32_gemminc_goi_w( |
| 2291 | size_t g, |
| 2292 | size_t nc, |
| 2293 | size_t kc, |
| 2294 | size_t nr, |
| 2295 | size_t kr, |
| 2296 | size_t sr, |
| 2297 | const float* k, |
| 2298 | float* packed_weights, |
| 2299 | const void* params) |
| 2300 | { |
| 2301 | assert(g != 0); |
| 2302 | assert(nr >= sr); |
| 2303 | assert(k != NULL); |
| 2304 | assert(packed_weights != NULL); |
| 2305 | |
| 2306 | const size_t skr = sr * kr; |
| 2307 | do { |
| 2308 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 2309 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 2310 | |
| 2311 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 2312 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 2313 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 2314 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 2315 | if (kc_idx < kc) { |
| 2316 | packed_weights[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 2317 | } |
| 2318 | } |
| 2319 | packed_weights += kr; |
| 2320 | } |
| 2321 | packed_weights += (nr - nr_block_size) * kr; |
| 2322 | } |
| 2323 | } |
| 2324 | k += nc * kc; |
| 2325 | } while (--g != 0); |
| 2326 | } |
| 2327 | |
| 2328 | void xnn_pack_f16_gemminc_goi_w( |
| 2329 | size_t g, |
| 2330 | size_t nc, |
| 2331 | size_t kc, |
| 2332 | size_t nr, |
| 2333 | size_t kr, |
| 2334 | size_t sr, |
| 2335 | const uint16_t* k, |
| 2336 | uint16_t* packed_weights, |
| 2337 | const void* params) |
| 2338 | { |
| 2339 | assert(g != 0); |
| 2340 | assert(nr >= sr); |
| 2341 | assert(k != NULL); |
| 2342 | assert(packed_weights != NULL); |
| 2343 | |
| 2344 | const size_t skr = sr * kr; |
| 2345 | do { |
| 2346 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 2347 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 2348 | |
| 2349 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 2350 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 2351 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 2352 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 2353 | if (kc_idx < kc) { |
| 2354 | packed_weights[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 2355 | } |
| 2356 | } |
| 2357 | packed_weights += kr; |
| 2358 | } |
| 2359 | packed_weights += (nr - nr_block_size) * kr; |
| 2360 | } |
| 2361 | } |
| 2362 | k += nc * kc; |
| 2363 | } while (--g != 0); |
| 2364 | } |
| 2365 | |
| 2366 | void xnn_pack_f32_dconv_oki_w( |
| 2367 | size_t nc, |
| 2368 | size_t kc, |
| 2369 | size_t nr, |
| 2370 | size_t kh, |
| 2371 | size_t kw, |
| 2372 | const float* k, |
| 2373 | const float* b, |
| 2374 | float* packed_weights, |
| 2375 | const void* params) |
| 2376 | { |
| 2377 | assert(k != NULL); |
| 2378 | assert(packed_weights != NULL); |
| 2379 | |
| 2380 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 2381 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 2382 | if XNN_LIKELY(b != NULL) { |
| 2383 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 2384 | *packed_weights++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 2385 | } |
| 2386 | } else { |
| 2387 | size_t n = nr; |
| 2388 | do { |
| 2389 | *packed_weights++ = 0.0f; |
| 2390 | } while (--n != 0); |
| 2391 | } |
| 2392 | |
| 2393 | for (size_t kx = 0; kx < kw; kx++) { |
| 2394 | for (size_t c = 0; c < kc; c++) { |
| 2395 | for (size_t ky = 0; ky < kh; ky++) { |
| 2396 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 2397 | *packed_weights++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 2398 | } |
| 2399 | } |
| 2400 | } |
| 2401 | } |
| 2402 | if XNN_UNPREDICTABLE(b != NULL) { |
| 2403 | b += nr; |
| 2404 | } |
| 2405 | } |
| 2406 | } |
| 2407 | |
| 2408 | void xnn_pack_f32_to_f16_dconv_oki_w( |
| 2409 | size_t nc, |
| 2410 | size_t kc, |
| 2411 | size_t nr, |
| 2412 | size_t kh, |
| 2413 | size_t kw, |
| 2414 | const float* k, |
| 2415 | const float* b, |
| 2416 | uint16_t* packed_weights, |
| 2417 | const void* params) |
| 2418 | { |
| 2419 | assert(k != NULL); |
| 2420 | assert(packed_weights != NULL); |
| 2421 | |
| 2422 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 2423 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 2424 | if XNN_LIKELY(b != NULL) { |
| 2425 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 2426 | *packed_weights++ = fp16_ieee_from_fp32_value(b[min(nr_block_offset, nr_block_size - 1)]); |
| 2427 | } |
| 2428 | } else { |
| 2429 | size_t n = nr; |
| 2430 | do { |
| 2431 | *packed_weights++ = 0; |
| 2432 | } while (--n != 0); |
| 2433 | } |
| 2434 | |
| 2435 | for (size_t kx = 0; kx < kw; kx++) { |
| 2436 | for (size_t c = 0; c < kc; c++) { |
| 2437 | for (size_t ky = 0; ky < kh; ky++) { |
| 2438 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 2439 | *packed_weights++ = fp16_ieee_from_fp32_value(k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]); |
| 2440 | } |
| 2441 | } |
| 2442 | } |
| 2443 | } |
| 2444 | if XNN_UNPREDICTABLE(b != NULL) { |
| 2445 | b += nr; |
| 2446 | } |
| 2447 | } |
| 2448 | } |
| 2449 | |
| 2450 | void xnn_pack_f16_dconv_oki_w( |
| 2451 | size_t nc, |
| 2452 | size_t kc, |
| 2453 | size_t nr, |
| 2454 | size_t kh, |
| 2455 | size_t kw, |
| 2456 | const uint16_t* k, |
| 2457 | const uint16_t* b, |
| 2458 | uint16_t* packed_weights, |
| 2459 | const void* params) |
| 2460 | { |
| 2461 | assert(k != NULL); |
| 2462 | assert(packed_weights != NULL); |
| 2463 | |
| 2464 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 2465 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 2466 | if XNN_LIKELY(b != NULL) { |
| 2467 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 2468 | *packed_weights++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 2469 | } |
| 2470 | } else { |
| 2471 | size_t n = nr; |
| 2472 | do { |
| 2473 | *packed_weights++ = 0; |
| 2474 | } while (--n != 0); |
| 2475 | } |
| 2476 | |
| 2477 | for (size_t kx = 0; kx < kw; kx++) { |
| 2478 | for (size_t c = 0; c < kc; c++) { |
| 2479 | for (size_t ky = 0; ky < kh; ky++) { |
| 2480 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 2481 | *packed_weights++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 2482 | } |
| 2483 | } |
| 2484 | } |
| 2485 | } |
| 2486 | if XNN_UNPREDICTABLE(b != NULL) { |
| 2487 | b += nr; |
| 2488 | } |
| 2489 | } |
| 2490 | } |
| 2491 | |
| 2492 | void xnn_pack_f32_chw_dwconv_ghw_w( |
| 2493 | size_t kernel_size, |
| 2494 | size_t groups, |
| 2495 | const float* k, |
| 2496 | const float* b, |
| 2497 | float* packed_weights, |
| 2498 | const void* params) |
| 2499 | { |
| 2500 | assert(k != NULL); |
| 2501 | assert(packed_weights != NULL); |
| 2502 | |
| 2503 | for (size_t g = 0; g < groups; g++) { |
| 2504 | if XNN_LIKELY(b != NULL) { |
| 2505 | *packed_weights = *b++; |
| 2506 | } else { |
| 2507 | *packed_weights = 0.0f; |
| 2508 | } |
| 2509 | packed_weights += 1; |
| 2510 | for (size_t i = 0; i < kernel_size; i++) { |
| 2511 | *packed_weights++ = k[g * kernel_size + i]; |
| 2512 | } |
| 2513 | } |
| 2514 | } |
| 2515 | |
| 2516 | void xnn_pack_f32_to_f16_chw_dwconv_ghw_w( |
| 2517 | size_t kernel_size, |
| 2518 | size_t groups, |
| 2519 | const float* k, |
| 2520 | const float* b, |
| 2521 | uint16_t* packed_weights, |
| 2522 | const void* params) |
| 2523 | { |
| 2524 | assert(k != NULL); |
| 2525 | assert(packed_weights != NULL); |
| 2526 | |
| 2527 | for (size_t g = 0; g < groups; g++) { |
| 2528 | if XNN_LIKELY(b != NULL) { |
| 2529 | *packed_weights = fp16_ieee_from_fp32_value(*b++); |
| 2530 | } else { |
| 2531 | *packed_weights = 0; |
| 2532 | } |
| 2533 | packed_weights += 1; |
| 2534 | for (size_t i = 0; i < kernel_size; i++) { |
| 2535 | *packed_weights++ = fp16_ieee_from_fp32_value(k[g * kernel_size + i]); |
| 2536 | } |
| 2537 | } |
| 2538 | } |
| 2539 | |
| 2540 | void xnn_pack_f16_chw_dwconv_ghw_w( |
| 2541 | size_t kernel_size, |
| 2542 | size_t groups, |
| 2543 | const uint16_t* k, |
| 2544 | const uint16_t* b, |
| 2545 | uint16_t* packed_weights, |
| 2546 | const void* params) |
| 2547 | { |
| 2548 | assert(k != NULL); |
| 2549 | assert(packed_weights != NULL); |
| 2550 | |
| 2551 | for (size_t g = 0; g < groups; g++) { |
| 2552 | if XNN_LIKELY(b != NULL) { |
| 2553 | *packed_weights = *b++; |
| 2554 | } else { |
| 2555 | *packed_weights = 0; |
| 2556 | } |
| 2557 | packed_weights += 1; |
| 2558 | for (size_t i = 0; i < kernel_size; i++) { |
| 2559 | *packed_weights++ = k[g * kernel_size + i]; |
| 2560 | } |
| 2561 | } |
| 2562 | } |
| 2563 | |
| 2564 | void xnn_pack_f32_chw_dwconv_hwg_w( |
| 2565 | size_t kernel_size, |
| 2566 | size_t groups, |
| 2567 | const float* k, |
| 2568 | const float* b, |
| 2569 | float* packed_weights, |
| 2570 | const void* params) |
| 2571 | { |
| 2572 | assert(k != NULL); |
| 2573 | assert(packed_weights != NULL); |
| 2574 | |
| 2575 | for (size_t g = 0; g < groups; g++) { |
| 2576 | if XNN_LIKELY(b != NULL) { |
| 2577 | *packed_weights = *b++; |
| 2578 | } else { |
| 2579 | *packed_weights = 0.0f; |
| 2580 | } |
| 2581 | packed_weights += 1; |
| 2582 | for (size_t i = 0; i < kernel_size; i++) { |
| 2583 | *packed_weights++ = k[i * groups + g]; |
| 2584 | } |
| 2585 | } |
| 2586 | } |
| 2587 | |
| 2588 | void xnn_pack_f16_chw_dwconv_hwg_w( |
| 2589 | size_t kernel_size, |
| 2590 | size_t groups, |
| 2591 | const uint16_t* k, |
| 2592 | const uint16_t* b, |
| 2593 | uint16_t* packed_weights, |
| 2594 | const void* params) |
| 2595 | { |
| 2596 | assert(k != NULL); |
| 2597 | assert(packed_weights != NULL); |
| 2598 | |
| 2599 | for (size_t g = 0; g < groups; g++) { |
| 2600 | if XNN_LIKELY(b != NULL) { |
| 2601 | *packed_weights = *b++; |
| 2602 | } else { |
| 2603 | *packed_weights = 0; |
| 2604 | } |
| 2605 | packed_weights += 1; |
| 2606 | for (size_t i = 0; i < kernel_size; i++) { |
| 2607 | *packed_weights++ = k[i * groups + g]; |
| 2608 | } |
| 2609 | } |
| 2610 | } |
| 2611 | |
| 2612 | void xnn_pack_f32_to_f16_chw_dwconv_hwg_w( |
| 2613 | size_t kernel_size, |
| 2614 | size_t groups, |
| 2615 | const float* k, |
| 2616 | const float* b, |
| 2617 | uint16_t* packed_weights, |
| 2618 | const void* params) |
| 2619 | { |
| 2620 | assert(k != NULL); |
| 2621 | assert(packed_weights != NULL); |
| 2622 | |
| 2623 | for (size_t g = 0; g < groups; g++) { |
| 2624 | if XNN_LIKELY(b != NULL) { |
| 2625 | *packed_weights = fp16_ieee_from_fp32_value(*b++); |
| 2626 | } else { |
| 2627 | *packed_weights = 0; |
| 2628 | } |
| 2629 | packed_weights += 1; |
| 2630 | for (size_t i = 0; i < kernel_size; i++) { |
| 2631 | *packed_weights++ = fp16_ieee_from_fp32_value(k[i * groups + g]); |
| 2632 | } |
| 2633 | } |
| 2634 | } |
| 2635 | |
| 2636 | |
| 2637 | void xnn_pack_f32_vmulcaddc_w( |
| 2638 | size_t c, |
| 2639 | size_t cr, |
| 2640 | const float* s, |
| 2641 | const float* b, |
| 2642 | float* packed_weights, |
| 2643 | const void* params) |
| 2644 | { |
| 2645 | assert(s != NULL); |
| 2646 | assert(packed_weights != NULL); |
| 2647 | |
| 2648 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2649 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2650 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2651 | *packed_weights++ = s[cr_block_start + cr_block_offset]; |
| 2652 | } |
| 2653 | packed_weights += cr - cr_block_size; |
| 2654 | if XNN_LIKELY(b != NULL) { |
| 2655 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2656 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 2657 | } |
| 2658 | } else { |
| 2659 | size_t n = cr_block_size; |
| 2660 | do { |
| 2661 | *packed_weights++ = 0.0f; |
| 2662 | } while (--n != 0); |
| 2663 | } |
| 2664 | packed_weights += cr - cr_block_size; |
| 2665 | } |
| 2666 | } |
| 2667 | |
| 2668 | void xnn_pack_f16_vmulcaddc_w( |
| 2669 | size_t c, |
| 2670 | size_t cr, |
| 2671 | const uint16_t* s, |
| 2672 | const uint16_t* b, |
| 2673 | uint16_t* packed_weights, |
| 2674 | const void* params) |
| 2675 | { |
| 2676 | assert(s != NULL); |
| 2677 | assert(packed_weights != NULL); |
| 2678 | |
| 2679 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2680 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2681 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2682 | *packed_weights++ = s[cr_block_start + cr_block_offset]; |
| 2683 | } |
| 2684 | packed_weights += cr - cr_block_size; |
| 2685 | if XNN_LIKELY(b != NULL) { |
| 2686 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2687 | *packed_weights++ = b[cr_block_start + cr_block_offset]; |
| 2688 | } |
| 2689 | } else { |
| 2690 | size_t n = cr_block_size; |
| 2691 | do { |
| 2692 | *packed_weights++ = 0; |
| 2693 | } while (--n != 0); |
| 2694 | } |
| 2695 | packed_weights += cr - cr_block_size; |
| 2696 | } |
| 2697 | } |
| 2698 | |
| 2699 | void xnn_pack_f32_to_f16_vmulcaddc_w( |
| 2700 | size_t c, |
| 2701 | size_t cr, |
| 2702 | const float* s, |
| 2703 | const float* b, |
| 2704 | uint16_t* packed_weights, |
| 2705 | const void* params) |
| 2706 | { |
| 2707 | assert(s != NULL); |
| 2708 | assert(packed_weights != NULL); |
| 2709 | |
| 2710 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2711 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2712 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2713 | *packed_weights++ = fp16_ieee_from_fp32_value(s[cr_block_start + cr_block_offset]); |
| 2714 | } |
| 2715 | packed_weights += cr - cr_block_size; |
| 2716 | if XNN_LIKELY(b != NULL) { |
| 2717 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2718 | *packed_weights++ = fp16_ieee_from_fp32_value(b[cr_block_start + cr_block_offset]); |
| 2719 | } |
| 2720 | } else { |
| 2721 | size_t n = cr_block_size; |
| 2722 | do { |
| 2723 | *packed_weights++ = 0; |
| 2724 | } while (--n != 0); |
| 2725 | } |
| 2726 | packed_weights += cr - cr_block_size; |
| 2727 | } |
| 2728 | } |
| 2729 | |
| 2730 | void xnn_pack_f32_prelu_w( |
| 2731 | size_t c, |
| 2732 | const float* s, |
| 2733 | float* packed_weights) |
| 2734 | { |
| 2735 | assert(s != NULL); |
| 2736 | assert(packed_weights != NULL); |
| 2737 | |
| 2738 | memcpy(packed_weights, s, c * sizeof(float)); |
| 2739 | } |
| 2740 | |
| 2741 | void xnn_pack_f16_prelu_w( |
| 2742 | size_t c, |
| 2743 | const uint16_t* s, |
| 2744 | uint16_t* packed_weights) |
| 2745 | { |
| 2746 | assert(s != NULL); |
| 2747 | assert(packed_weights != NULL); |
| 2748 | |
| 2749 | memcpy(packed_weights, s, c * sizeof(uint16_t)); |
| 2750 | } |
| 2751 | |
| 2752 | void xnn_pack_f32_to_f16_prelu_w( |
| 2753 | size_t c, |
| 2754 | const float* s, |
| 2755 | uint16_t* packed_weights) |
| 2756 | { |
| 2757 | assert(s != NULL); |
| 2758 | assert(packed_weights != NULL); |
| 2759 | |
| 2760 | do { |
| 2761 | *packed_weights++ = fp16_ieee_from_fp32_value(*s++); |
| 2762 | } while (--c != 0); |
| 2763 | } |
| 2764 | |
| 2765 | void xnn_analyze_f32_spmm( |
| 2766 | size_t group_output_channels, |
| 2767 | size_t group_input_channels, |
| 2768 | const float* kernel, |
| 2769 | struct xnn_spmm_packing_params* params) |
| 2770 | { |
| 2771 | assert(kernel != NULL); |
| 2772 | assert(params != NULL); |
| 2773 | |
| 2774 | // Count number of non-zero values. |
| 2775 | size_t num_nonzeroes = 0; |
| 2776 | size_t num_nonzero_blocks2 = 0; |
| 2777 | size_t num_nonzero_blocks4 = 0; |
| 2778 | for (size_t oc = 0; oc < round_down_po2(group_output_channels, 4); oc += 4) { |
| 2779 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2780 | const size_t row0_nonzero = (size_t) (kernel[oc * group_input_channels + ic] != 0.0f); |
| 2781 | const size_t row1_nonzero = (size_t) (kernel[(oc + 1) * group_input_channels + ic] != 0.0f); |
| 2782 | const size_t row2_nonzero = (size_t) (kernel[(oc + 2) * group_input_channels + ic] != 0.0f); |
| 2783 | const size_t row3_nonzero = (size_t) (kernel[(oc + 3) * group_input_channels + ic] != 0.0f); |
| 2784 | num_nonzeroes += row0_nonzero + row1_nonzero + row2_nonzero + row3_nonzero; |
| 2785 | num_nonzero_blocks2 += (row0_nonzero | row1_nonzero) + (row2_nonzero | row3_nonzero); |
| 2786 | num_nonzero_blocks4 += (row0_nonzero | row1_nonzero | row2_nonzero | row3_nonzero); |
| 2787 | } |
| 2788 | } |
| 2789 | const size_t num_block4_nonzeroes = num_nonzeroes; |
| 2790 | for (size_t oc = round_down_po2(group_output_channels, 4); oc < round_down_po2(group_output_channels, 2); oc += 2) { |
| 2791 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2792 | const size_t row0_nonzero = (size_t) (kernel[oc * group_input_channels + ic] != 0.0f); |
| 2793 | const size_t row1_nonzero = (size_t) (kernel[(oc + 1) * group_input_channels + ic] != 0.0f); |
| 2794 | num_nonzeroes += row0_nonzero + row1_nonzero; |
| 2795 | num_nonzero_blocks2 += (row0_nonzero | row1_nonzero); |
| 2796 | } |
| 2797 | } |
| 2798 | const size_t num_block2_nonzeroes = num_nonzeroes; |
| 2799 | for (size_t oc = round_down_po2(group_output_channels, 2); oc < group_output_channels; oc++) { |
| 2800 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2801 | num_nonzeroes += (size_t) (kernel[oc * group_input_channels + ic] != 0.0f); |
| 2802 | } |
| 2803 | } |
| 2804 | params->num_nonzeroes = num_nonzeroes; |
| 2805 | params->num_nonzero_blocks2 = num_nonzero_blocks2; |
| 2806 | params->num_nonzero_blocks4 = num_nonzero_blocks4; |
| 2807 | params->num_block2_nonzeroes = num_block2_nonzeroes; |
| 2808 | params->num_block4_nonzeroes = num_block4_nonzeroes; |
| 2809 | } |
| 2810 | |
| 2811 | void xnn_analyze_f16_spmm( |
| 2812 | size_t group_output_channels, |
| 2813 | size_t group_input_channels, |
| 2814 | const uint16_t* kernel, |
| 2815 | struct xnn_spmm_packing_params* params) |
| 2816 | { |
| 2817 | assert(kernel != NULL); |
| 2818 | assert(params != NULL); |
| 2819 | |
| 2820 | // Count number of non-zero values. |
| 2821 | size_t num_nonzeroes = 0; |
| 2822 | size_t num_nonzero_blocks2 = 0; |
| 2823 | size_t num_nonzero_blocks4 = 0; |
| 2824 | for (size_t oc = 0; oc < round_down_po2(group_output_channels, 4); oc += 4) { |
| 2825 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2826 | const size_t row0_nonzero = (size_t) (kernel[oc * group_input_channels + ic] != 0); |
| 2827 | const size_t row1_nonzero = (size_t) (kernel[(oc + 1) * group_input_channels + ic] != 0); |
| 2828 | const size_t row2_nonzero = (size_t) (kernel[(oc + 2) * group_input_channels + ic] != 0); |
| 2829 | const size_t row3_nonzero = (size_t) (kernel[(oc + 3) * group_input_channels + ic] != 0); |
| 2830 | num_nonzeroes += row0_nonzero + row1_nonzero + row2_nonzero + row3_nonzero; |
| 2831 | num_nonzero_blocks2 += (row0_nonzero | row1_nonzero) + (row2_nonzero | row3_nonzero); |
| 2832 | num_nonzero_blocks4 += (row0_nonzero | row1_nonzero | row2_nonzero | row3_nonzero); |
| 2833 | } |
| 2834 | } |
| 2835 | const size_t num_block4_nonzeroes = num_nonzeroes; |
| 2836 | for (size_t oc = round_down_po2(group_output_channels, 4); oc < round_down_po2(group_output_channels, 2); oc += 2) { |
| 2837 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2838 | const size_t row0_nonzero = (size_t) (kernel[oc * group_input_channels + ic] != 0); |
| 2839 | const size_t row1_nonzero = (size_t) (kernel[(oc + 1) * group_input_channels + ic] != 0); |
| 2840 | num_nonzeroes += row0_nonzero + row1_nonzero; |
| 2841 | num_nonzero_blocks2 += (row0_nonzero | row1_nonzero); |
| 2842 | } |
| 2843 | } |
| 2844 | const size_t num_block2_nonzeroes = num_nonzeroes; |
| 2845 | for (size_t oc = round_down_po2(group_output_channels, 2); oc < group_output_channels; oc++) { |
| 2846 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2847 | num_nonzeroes += (size_t) (kernel[oc * group_input_channels + ic] != 0); |
| 2848 | } |
| 2849 | } |
| 2850 | params->num_nonzeroes = num_nonzeroes; |
| 2851 | params->num_nonzero_blocks2 = num_nonzero_blocks2; |
| 2852 | params->num_nonzero_blocks4 = num_nonzero_blocks4; |
| 2853 | params->num_block2_nonzeroes = num_block2_nonzeroes; |
| 2854 | params->num_block4_nonzeroes = num_block4_nonzeroes; |
| 2855 | } |
| 2856 | |
| 2857 | enum xnn_status xnn_pack_f32_spmm( |
| 2858 | size_t group_output_channels, |
| 2859 | size_t output_channels_block_size, |
| 2860 | size_t group_input_channels, |
| 2861 | const float* kernel, |
| 2862 | const float* bias, |
| 2863 | int32_t* input_channel_diffs, |
| 2864 | uint32_t* output_channel_nonzeros, |
| 2865 | float* nonzero_values, |
| 2866 | size_t* first_input_channel) |
| 2867 | { |
| 2868 | size_t first_ic = 0, last_ic = 0; |
| 2869 | bool first_nonzero = true; |
| 2870 | for (size_t ocb = 0; ocb < round_down_po2(group_output_channels, output_channels_block_size); ocb += output_channels_block_size) { |
| 2871 | if XNN_LIKELY(bias != NULL) { |
| 2872 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2873 | *nonzero_values++ = bias[ocb + oco]; |
| 2874 | } |
| 2875 | } else { |
| 2876 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2877 | *nonzero_values++ = 0.0f; |
| 2878 | } |
| 2879 | } |
| 2880 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2881 | bool is_nonzero_block = false; |
| 2882 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2883 | is_nonzero_block |= (kernel[(ocb + oco) * group_input_channels + ic] != 0.0f); |
| 2884 | } |
| 2885 | if (is_nonzero_block) { |
| 2886 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2887 | *nonzero_values++ = kernel[(ocb + oco) * group_input_channels + ic]; |
| 2888 | } |
| 2889 | if (first_nonzero) { |
| 2890 | first_ic = ic; |
| 2891 | } else { |
| 2892 | const int64_t diff = (int64_t) ((uint64_t) ic - (uint64_t) last_ic) * (int64_t) sizeof(float); |
| 2893 | if (diff != (int64_t) (int32_t) diff) { |
| 2894 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 2895 | "scaled difference in input channels exceeds int32_t range"); |
| 2896 | return xnn_status_unsupported_parameter; |
| 2897 | } |
| 2898 | *input_channel_diffs++ = (int32_t) diff; |
| 2899 | } |
| 2900 | first_nonzero = false; |
| 2901 | last_ic = ic; |
| 2902 | *output_channel_nonzeros += 1; |
| 2903 | } |
| 2904 | } |
| 2905 | output_channel_nonzeros += 1; |
| 2906 | } |
| 2907 | for (size_t oc = round_down_po2(group_output_channels, output_channels_block_size); oc < group_output_channels; oc++) { |
| 2908 | if XNN_LIKELY(bias != NULL) { |
| 2909 | *nonzero_values++ = bias[oc]; |
| 2910 | } else { |
| 2911 | *nonzero_values++ = 0.0f; |
| 2912 | } |
| 2913 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2914 | const float weight = kernel[oc * group_input_channels + ic]; |
| 2915 | if (weight != 0.0f) { |
| 2916 | *nonzero_values++ = weight; |
| 2917 | if (first_nonzero) { |
| 2918 | first_ic = ic; |
| 2919 | } else { |
| 2920 | const int64_t diff = (int64_t) ((uint64_t) ic - (uint64_t) last_ic) * (int64_t) sizeof(float); |
| 2921 | if (diff != (int64_t) (int32_t) diff) { |
| 2922 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 2923 | "scaled difference in input channels exceeds int32_t range"); |
| 2924 | return xnn_status_unsupported_parameter; |
| 2925 | } |
| 2926 | *input_channel_diffs++ = (int32_t) diff; |
| 2927 | } |
| 2928 | first_nonzero = false; |
| 2929 | last_ic = ic; |
| 2930 | *output_channel_nonzeros += 1; |
| 2931 | } |
| 2932 | } |
| 2933 | output_channel_nonzeros += 1; |
| 2934 | } |
| 2935 | // If there are any non-zero elements, we have to return to the initial input channel. |
| 2936 | if (!first_nonzero) { |
| 2937 | const int64_t diff = (int64_t) ((uint64_t) first_ic - (uint64_t) last_ic) * (int64_t) sizeof(float); |
| 2938 | if (diff != (int64_t) (int32_t) diff) { |
| 2939 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 2940 | "scaled difference in input channels exceeds int32_t range"); |
| 2941 | return xnn_status_unsupported_parameter; |
| 2942 | } |
| 2943 | *input_channel_diffs++ = (int32_t) diff; |
| 2944 | } |
| 2945 | *first_input_channel = first_ic; |
| 2946 | return xnn_status_success; |
| 2947 | } |
| 2948 | |
| 2949 | |
| 2950 | enum xnn_status xnn_pack_f32_to_f16_spmm( |
| 2951 | size_t group_output_channels, |
| 2952 | size_t output_channels_block_size, |
| 2953 | size_t group_input_channels, |
| 2954 | const float* kernel, |
| 2955 | const float* bias, |
| 2956 | int32_t* input_channel_diffs, |
| 2957 | uint32_t* output_channel_nonzeros, |
| 2958 | uint16_t* nonzero_values, // fp16 values |
| 2959 | size_t* first_input_channel) |
| 2960 | { |
| 2961 | size_t first_ic = 0, last_ic = 0; |
| 2962 | bool first_nonzero = true; |
| 2963 | for (size_t ocb = 0; ocb < round_down_po2(group_output_channels, output_channels_block_size); ocb += output_channels_block_size) { |
| 2964 | if XNN_LIKELY(bias != NULL) { |
| 2965 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2966 | *nonzero_values++ = fp16_ieee_from_fp32_value(bias[ocb + oco]); |
| 2967 | } |
| 2968 | } else { |
| 2969 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2970 | *nonzero_values++ = 0; |
| 2971 | } |
| 2972 | } |
| 2973 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 2974 | bool is_nonzero_block = false; |
| 2975 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2976 | is_nonzero_block |= (kernel[(ocb + oco) * group_input_channels + ic] != 0.0f); |
| 2977 | } |
| 2978 | if (is_nonzero_block) { |
| 2979 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 2980 | *nonzero_values++ = fp16_ieee_from_fp32_value(kernel[(ocb + oco) * group_input_channels + ic]); |
| 2981 | } |
| 2982 | if (first_nonzero) { |
| 2983 | first_ic = ic; |
| 2984 | } else { |
| 2985 | const int64_t diff = (int64_t) ((uint64_t) ic - (uint64_t) last_ic) * (int64_t) sizeof(uint16_t); |
| 2986 | if (diff != (int64_t) (int32_t) diff) { |
| 2987 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 2988 | "scaled difference in input channels exceeds int32_t range"); |
| 2989 | return xnn_status_unsupported_parameter; |
| 2990 | } |
| 2991 | *input_channel_diffs++ = (int32_t) diff; |
| 2992 | } |
| 2993 | first_nonzero = false; |
| 2994 | last_ic = ic; |
| 2995 | *output_channel_nonzeros += 1; |
| 2996 | } |
| 2997 | } |
| 2998 | output_channel_nonzeros += 1; |
| 2999 | } |
| 3000 | for (size_t oc = round_down_po2(group_output_channels, output_channels_block_size); oc < group_output_channels; oc++) { |
| 3001 | if XNN_LIKELY(bias != NULL) { |
| 3002 | *nonzero_values++ = fp16_ieee_from_fp32_value(bias[oc]); |
| 3003 | } else { |
| 3004 | *nonzero_values++ = 0; |
| 3005 | } |
| 3006 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 3007 | const float weight = kernel[oc * group_input_channels + ic]; |
| 3008 | if (weight != 0.0f) { |
| 3009 | *nonzero_values++ = fp16_ieee_from_fp32_value(weight); |
| 3010 | if (first_nonzero) { |
| 3011 | first_ic = ic; |
| 3012 | } else { |
| 3013 | const int64_t diff = (int64_t) ((uint64_t) ic - (uint64_t) last_ic) * (int64_t) sizeof(uint16_t); |
| 3014 | if (diff != (int64_t) (int32_t) diff) { |
| 3015 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 3016 | "scaled difference in input channels exceeds int32_t range"); |
| 3017 | return xnn_status_unsupported_parameter; |
| 3018 | } |
| 3019 | *input_channel_diffs++ = (int32_t) diff; |
| 3020 | } |
| 3021 | first_nonzero = false; |
| 3022 | last_ic = ic; |
| 3023 | *output_channel_nonzeros += 1; |
| 3024 | } |
| 3025 | } |
| 3026 | output_channel_nonzeros += 1; |
| 3027 | } |
| 3028 | // If there are any non-zero elements, we have to return to the initial input channel. |
| 3029 | if (!first_nonzero) { |
| 3030 | const int64_t diff = (int64_t) ((uint64_t) first_ic - (uint64_t) last_ic) * (int64_t) sizeof(uint16_t); |
| 3031 | if (diff != (int64_t) (int32_t) diff) { |
| 3032 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 3033 | "scaled difference in input channels exceeds int32_t range"); |
| 3034 | return xnn_status_unsupported_parameter; |
| 3035 | } |
| 3036 | *input_channel_diffs++ = (int32_t) diff; |
| 3037 | } |
| 3038 | *first_input_channel = first_ic; |
| 3039 | return xnn_status_success; |
| 3040 | } |
| 3041 | |
| 3042 | enum xnn_status xnn_pack_f16_spmm( |
| 3043 | size_t group_output_channels, |
| 3044 | size_t output_channels_block_size, |
| 3045 | size_t group_input_channels, |
| 3046 | const uint16_t* kernel, // fp16 values |
| 3047 | const uint16_t* bias, // fp16 values |
| 3048 | int32_t* input_channel_diffs, |
| 3049 | uint32_t* output_channel_nonzeros, |
| 3050 | uint16_t* nonzero_values, // fp16 values |
| 3051 | size_t* first_input_channel) |
| 3052 | { |
| 3053 | size_t first_ic = 0, last_ic = 0; |
| 3054 | bool first_nonzero = true; |
| 3055 | for (size_t ocb = 0; ocb < round_down_po2(group_output_channels, output_channels_block_size); ocb += output_channels_block_size) { |
| 3056 | if XNN_LIKELY(bias != NULL) { |
| 3057 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 3058 | *nonzero_values++ = bias[ocb + oco]; |
| 3059 | } |
| 3060 | } else { |
| 3061 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 3062 | *nonzero_values++ = 0; |
| 3063 | } |
| 3064 | } |
| 3065 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 3066 | bool is_nonzero_block = false; |
| 3067 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 3068 | is_nonzero_block |= (kernel[(ocb + oco) * group_input_channels + ic] != 0); |
| 3069 | } |
| 3070 | if (is_nonzero_block) { |
| 3071 | for (size_t oco = 0; oco < output_channels_block_size; oco++) { |
| 3072 | *nonzero_values++ = kernel[(ocb + oco) * group_input_channels + ic]; |
| 3073 | } |
| 3074 | if (first_nonzero) { |
| 3075 | first_ic = ic; |
| 3076 | } else { |
| 3077 | const int64_t diff = (int64_t) ((uint64_t) ic - (uint64_t) last_ic) * (int64_t) sizeof(uint16_t); |
| 3078 | if (diff != (int64_t) (int32_t) diff) { |
| 3079 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 3080 | "scaled difference in input channels exceeds int32_t range"); |
| 3081 | return xnn_status_unsupported_parameter; |
| 3082 | } |
| 3083 | *input_channel_diffs++ = (int32_t) diff; |
| 3084 | } |
| 3085 | first_nonzero = false; |
| 3086 | last_ic = ic; |
| 3087 | *output_channel_nonzeros += 1; |
| 3088 | } |
| 3089 | } |
| 3090 | output_channel_nonzeros += 1; |
| 3091 | } |
| 3092 | for (size_t oc = round_down_po2(group_output_channels, output_channels_block_size); oc < group_output_channels; oc++) { |
| 3093 | if XNN_LIKELY(bias != NULL) { |
| 3094 | *nonzero_values++ = bias[oc]; |
| 3095 | } else { |
| 3096 | *nonzero_values++ = 0; |
| 3097 | } |
| 3098 | for (size_t ic = 0; ic < group_input_channels; ic++) { |
| 3099 | const float weight = kernel[oc * group_input_channels + ic]; |
| 3100 | if (weight != 0) { |
| 3101 | *nonzero_values++ = weight; |
| 3102 | if (first_nonzero) { |
| 3103 | first_ic = ic; |
| 3104 | } else { |
| 3105 | const int64_t diff = (int64_t) ((uint64_t) ic - (uint64_t) last_ic) * (int64_t) sizeof(uint16_t); |
| 3106 | if (diff != (int64_t) (int32_t) diff) { |
| 3107 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 3108 | "scaled difference in input channels exceeds int32_t range"); |
| 3109 | return xnn_status_unsupported_parameter; |
| 3110 | } |
| 3111 | *input_channel_diffs++ = (int32_t) diff; |
| 3112 | } |
| 3113 | first_nonzero = false; |
| 3114 | last_ic = ic; |
| 3115 | *output_channel_nonzeros += 1; |
| 3116 | } |
| 3117 | } |
| 3118 | output_channel_nonzeros += 1; |
| 3119 | } |
| 3120 | // If there are any non-zero elements, we have to return to the initial input channel. |
| 3121 | if (!first_nonzero) { |
| 3122 | const int64_t diff = (int64_t) ((uint64_t) first_ic - (uint64_t) last_ic) * (int64_t) sizeof(uint16_t); |
| 3123 | if (diff != (int64_t) (int32_t) diff) { |
| 3124 | xnn_log_error("failed to convert kernel to sparse representation: " |
| 3125 | "scaled difference in input channels exceeds int32_t range"); |
| 3126 | return xnn_status_unsupported_parameter; |
| 3127 | } |
| 3128 | *input_channel_diffs++ = (int32_t) diff; |
| 3129 | } |
| 3130 | *first_input_channel = first_ic; |
| 3131 | return xnn_status_success; |
| 3132 | } |
| 3133 |
Generated on 2023-Feb-12 from project pytorch revision 2c76838d7ff
Powered by 
Code Browser 2.1
Generator usage only permitted with license.