| 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/math.h> |
| 16 | #include <xnnpack/operator.h> |
| 17 | #include <xnnpack/pack.h> |
| 18 | #include <xnnpack/unaligned.h> |
| 19 | |
| 20 | |
| 21 | void xnn_pack_f32_gemm_goi_w( |
| 22 | size_t g, |
| 23 | size_t nc, |
| 24 | size_t kc, |
| 25 | size_t nr, |
| 26 | size_t kr, |
| 27 | size_t sr, |
| 28 | const float* k, |
| 29 | const float* b, |
| 30 | float* packed_w, |
| 31 | size_t extra_bytes, |
| 32 | const void* params) |
| 33 | { |
| 34 | assert(nr >= sr); |
| 35 | |
| 36 | const size_t skr = sr * kr; |
| 37 | do { |
| 38 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 39 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 40 | if XNN_LIKELY(b != NULL) { |
| 41 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 42 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 43 | } |
| 44 | } |
| 45 | packed_w += nr; |
| 46 | |
| 47 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 48 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 49 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 50 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 51 | if (kc_idx < kc) { |
| 52 | packed_w[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 53 | } |
| 54 | } |
| 55 | packed_w += kr; |
| 56 | } |
| 57 | packed_w += (nr - nr_block_size) * kr; |
| 58 | } |
| 59 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 60 | } |
| 61 | k += nc * kc; |
| 62 | if XNN_UNPREDICTABLE(b != NULL) { |
| 63 | b += nc; |
| 64 | } |
| 65 | } while (--g != 0); |
| 66 | } |
| 67 | |
| 68 | void xnn_pack_f16_gemm_goi_w( |
| 69 | size_t g, |
| 70 | size_t nc, |
| 71 | size_t kc, |
| 72 | size_t nr, |
| 73 | size_t kr, |
| 74 | size_t sr, |
| 75 | const uint16_t* k, |
| 76 | const uint16_t* b, |
| 77 | uint16_t* packed_w, |
| 78 | size_t extra_bytes, |
| 79 | const void* params) |
| 80 | { |
| 81 | assert(nr >= sr); |
| 82 | |
| 83 | const size_t skr = sr * kr; |
| 84 | do { |
| 85 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 86 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 87 | if XNN_LIKELY(b != NULL) { |
| 88 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 89 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 90 | } |
| 91 | } |
| 92 | packed_w += nr; |
| 93 | |
| 94 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 95 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 96 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 97 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 98 | if (kc_idx < kc) { |
| 99 | packed_w[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 100 | } |
| 101 | } |
| 102 | packed_w += kr; |
| 103 | } |
| 104 | packed_w += (nr - nr_block_size) * kr; |
| 105 | } |
| 106 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 107 | } |
| 108 | k += nc * kc; |
| 109 | if XNN_UNPREDICTABLE(b != NULL) { |
| 110 | b += nc; |
| 111 | } |
| 112 | } while (--g != 0); |
| 113 | } |
| 114 | |
| 115 | void xnn_pack_f32_to_f16_gemm_goi_w( |
| 116 | size_t g, |
| 117 | size_t nc, |
| 118 | size_t kc, |
| 119 | size_t nr, |
| 120 | size_t kr, |
| 121 | size_t sr, |
| 122 | const float* k, |
| 123 | const float* b, |
| 124 | uint16_t* packed_w, |
| 125 | size_t extra_bytes, |
| 126 | const void* params) |
| 127 | { |
| 128 | assert(nr >= sr); |
| 129 | |
| 130 | const size_t skr = sr * kr; |
| 131 | do { |
| 132 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 133 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 134 | if XNN_LIKELY(b != NULL) { |
| 135 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 136 | packed_w[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 137 | } |
| 138 | } |
| 139 | packed_w += nr; |
| 140 | |
| 141 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 142 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 143 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 144 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 145 | if (kc_idx < kc) { |
| 146 | packed_w[kr_block_offset] = fp16_ieee_from_fp32_value(k[(nr_block_start + nr_block_offset) * kc + kc_idx]); |
| 147 | } |
| 148 | } |
| 149 | packed_w += kr; |
| 150 | } |
| 151 | packed_w += (nr - nr_block_size) * kr; |
| 152 | } |
| 153 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 154 | } |
| 155 | k += nc * kc; |
| 156 | if XNN_UNPREDICTABLE(b != NULL) { |
| 157 | b += nc; |
| 158 | } |
| 159 | } while (--g != 0); |
| 160 | } |
| 161 | |
| 162 | void xnn_pack_qu8_gemm_goi_w( |
| 163 | size_t g, |
| 164 | size_t nc, |
| 165 | size_t kc, |
| 166 | size_t nr, |
| 167 | size_t kr, |
| 168 | size_t sr, |
| 169 | const uint8_t* k, |
| 170 | const int32_t* b, |
| 171 | void* packed_w, |
| 172 | size_t extra_bytes, |
| 173 | const struct xnn_qu8_packing_params* params) |
| 174 | { |
| 175 | assert(nr >= sr); |
| 176 | |
| 177 | const size_t skr = sr * kr; |
| 178 | const int32_t izp = (int32_t) params->input_zero_point; |
| 179 | const int32_t bzp = (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
| 180 | do { |
| 181 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 182 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 183 | int32_t* packed_b = (int32_t*) packed_w; |
| 184 | if XNN_LIKELY(b != NULL) { |
| 185 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 186 | unaligned_store_s32(packed_w, bzp + b[nr_block_start + nr_block_offset]); |
| 187 | packed_w = (int32_t*) packed_w + 1; |
| 188 | } |
| 189 | } else { |
| 190 | size_t n = nr_block_size; |
| 191 | do { |
| 192 | unaligned_store_s32(packed_w, bzp); |
| 193 | packed_w = (int32_t*) packed_w + 1; |
| 194 | } while (--n != 0); |
| 195 | } |
| 196 | packed_w = (int32_t*) packed_w + (nr - nr_block_size); |
| 197 | |
| 198 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 199 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 200 | int32_t ksum = 0; |
| 201 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 202 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 203 | if (kc_idx < kc) { |
| 204 | const uint8_t kv = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 205 | ksum += (int32_t) kv; |
| 206 | ((uint8_t*) packed_w)[kr_block_offset] = kv; |
| 207 | } |
| 208 | } |
| 209 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 210 | packed_w = (uint8_t*) packed_w + kr; |
| 211 | } |
| 212 | packed_w = (uint8_t*) packed_w + (nr - nr_block_size) * kr; |
| 213 | } |
| 214 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 215 | } |
| 216 | k += nc * kc; |
| 217 | if XNN_UNPREDICTABLE(b != NULL) { |
| 218 | b += nc; |
| 219 | } |
| 220 | } while (--g != 0); |
| 221 | } |
| 222 | |
| 223 | void xnn_pack_qs8_gemm_goi_w( |
| 224 | size_t g, |
| 225 | size_t nc, |
| 226 | size_t kc, |
| 227 | size_t nr, |
| 228 | size_t kr, |
| 229 | size_t sr, |
| 230 | const int8_t* k, |
| 231 | const int32_t* b, |
| 232 | void* packed_w, |
| 233 | size_t extra_bytes, |
| 234 | const struct xnn_qs8_packing_params* params) |
| 235 | { |
| 236 | assert(nr >= sr); |
| 237 | |
| 238 | const size_t skr = sr * kr; |
| 239 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 240 | do { |
| 241 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 242 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 243 | int32_t* packed_b = (int32_t*) packed_w; |
| 244 | if XNN_LIKELY(b != NULL) { |
| 245 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 246 | unaligned_store_s32(packed_w, b[nr_block_start + nr_block_offset]); |
| 247 | packed_w = (int32_t*) packed_w + 1; |
| 248 | } |
| 249 | } else { |
| 250 | size_t n = nr_block_size; |
| 251 | do { |
| 252 | unaligned_store_s32(packed_w, 0); |
| 253 | packed_w = (int32_t*) packed_w + 1; |
| 254 | } while (--n != 0); |
| 255 | } |
| 256 | packed_w = (int32_t*) packed_w + (nr - nr_block_size); |
| 257 | |
| 258 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 259 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 260 | uint32_t ksum = 0; |
| 261 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 262 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 263 | if (kc_idx < kc) { |
| 264 | const int8_t kv = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 265 | ksum += (uint32_t) kv; |
| 266 | ((int8_t*) packed_w)[kr_block_offset] = kv; |
| 267 | } |
| 268 | } |
| 269 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 270 | packed_w = (int8_t*) packed_w + kr; |
| 271 | } |
| 272 | packed_w = (int8_t*) packed_w + (nr - nr_block_size) * kr; |
| 273 | } |
| 274 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 275 | } |
| 276 | k += nc * kc; |
| 277 | if XNN_UNPREDICTABLE(b != NULL) { |
| 278 | b += nc; |
| 279 | } |
| 280 | } while (--g != 0); |
| 281 | } |
| 282 | |
| 283 | void xnn_pack_qs8_gemm_xw_goi_w( |
| 284 | size_t g, |
| 285 | size_t nc, |
| 286 | size_t kc, |
| 287 | size_t nr, |
| 288 | size_t kr, |
| 289 | size_t sr, |
| 290 | const int8_t* k, |
| 291 | const int32_t* b, |
| 292 | void* packed_w, |
| 293 | size_t extra_bytes, |
| 294 | const struct xnn_qs8_packing_params* params) |
| 295 | { |
| 296 | assert(nr >= sr); |
| 297 | |
| 298 | const size_t skr = sr * kr; |
| 299 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 300 | do { |
| 301 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 302 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 303 | int32_t* packed_b = (int32_t*) packed_w; |
| 304 | if XNN_LIKELY(b != NULL) { |
| 305 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 306 | unaligned_store_s32(packed_w, b[nr_block_start + nr_block_offset]); |
| 307 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 308 | } |
| 309 | } else { |
| 310 | size_t n = nr_block_size; |
| 311 | do { |
| 312 | unaligned_store_s32(packed_w, 0); |
| 313 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 314 | } while (--n != 0); |
| 315 | } |
| 316 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 317 | |
| 318 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 319 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 320 | uint32_t ksum = 0; |
| 321 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 322 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 323 | if (kc_idx < kc) { |
| 324 | const int8_t kv = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 325 | ksum += (uint32_t) kv; |
| 326 | ((int16_t*) packed_w)[kr_block_offset] = (int16_t) kv; |
| 327 | } |
| 328 | } |
| 329 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 330 | packed_w = (int16_t*) packed_w + kr; |
| 331 | } |
| 332 | packed_w = (int16_t*) packed_w + (nr - nr_block_size) * kr; |
| 333 | } |
| 334 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 335 | } |
| 336 | k += nc * kc; |
| 337 | if XNN_UNPREDICTABLE(b != NULL) { |
| 338 | b += nc; |
| 339 | } |
| 340 | } while (--g != 0); |
| 341 | } |
| 342 | |
| 343 | void xnn_pack_f32_gemm_io_w( |
| 344 | size_t nc, |
| 345 | size_t kc, |
| 346 | size_t nr, |
| 347 | size_t kr, |
| 348 | size_t sr, |
| 349 | const float* k, |
| 350 | const float* b, |
| 351 | float* packed_w, |
| 352 | const void* params) |
| 353 | { |
| 354 | assert(nr >= sr); |
| 355 | |
| 356 | const size_t skr = sr * kr; |
| 357 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 358 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 359 | if XNN_LIKELY(b != NULL) { |
| 360 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 361 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 362 | } |
| 363 | } |
| 364 | packed_w += nr; |
| 365 | |
| 366 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 367 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 368 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 369 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 370 | if (kc_idx < kc) { |
| 371 | packed_w[kr_block_offset] = k[kc_idx * nc + nr_block_start + nr_block_offset]; |
| 372 | } |
| 373 | } |
| 374 | packed_w += kr; |
| 375 | } |
| 376 | packed_w += (nr - nr_block_size) * kr; |
| 377 | } |
| 378 | } |
| 379 | } |
| 380 | |
| 381 | void xnn_pack_f16_gemm_io_w( |
| 382 | size_t nc, |
| 383 | size_t kc, |
| 384 | size_t nr, |
| 385 | size_t kr, |
| 386 | size_t sr, |
| 387 | const uint16_t* k, |
| 388 | const uint16_t* b, |
| 389 | uint16_t* packed_w, |
| 390 | const void* params) |
| 391 | { |
| 392 | assert(nr >= sr); |
| 393 | |
| 394 | const size_t skr = sr * kr; |
| 395 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 396 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 397 | if XNN_LIKELY(b != NULL) { |
| 398 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 399 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 400 | } |
| 401 | } |
| 402 | packed_w += nr; |
| 403 | |
| 404 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 405 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 406 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 407 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 408 | if (kc_idx < kc) { |
| 409 | packed_w[kr_block_offset] = k[kc_idx * nc + nr_block_start + nr_block_offset]; |
| 410 | } |
| 411 | } |
| 412 | packed_w += kr; |
| 413 | } |
| 414 | packed_w += (nr - nr_block_size) * kr; |
| 415 | } |
| 416 | } |
| 417 | } |
| 418 | |
| 419 | void xnn_pack_f32_to_f16_gemm_io_w( |
| 420 | size_t nc, |
| 421 | size_t kc, |
| 422 | size_t nr, |
| 423 | size_t kr, |
| 424 | size_t sr, |
| 425 | const float* k, |
| 426 | const float* b, |
| 427 | uint16_t* packed_w, |
| 428 | const void* params) |
| 429 | { |
| 430 | assert(nr >= sr); |
| 431 | |
| 432 | const size_t skr = sr * kr; |
| 433 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 434 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 435 | if XNN_LIKELY(b != NULL) { |
| 436 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 437 | packed_w[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 438 | } |
| 439 | } |
| 440 | packed_w += nr; |
| 441 | |
| 442 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 443 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 444 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 445 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 446 | if (kc_idx < kc) { |
| 447 | packed_w[kr_block_offset] = fp16_ieee_from_fp32_value(k[kc_idx * nc + nr_block_start + nr_block_offset]); |
| 448 | } |
| 449 | } |
| 450 | packed_w += kr; |
| 451 | } |
| 452 | packed_w += (nr - nr_block_size) * kr; |
| 453 | } |
| 454 | } |
| 455 | } |
| 456 | |
| 457 | void xnn_pack_qu8_gemm_io_w( |
| 458 | size_t nc, |
| 459 | size_t kc, |
| 460 | size_t nr, |
| 461 | size_t kr, |
| 462 | size_t sr, |
| 463 | const uint8_t* k, |
| 464 | const int32_t* b, |
| 465 | void* packed_w, |
| 466 | const struct xnn_qu8_packing_params* params) |
| 467 | { |
| 468 | assert(nr >= sr); |
| 469 | |
| 470 | const size_t skr = sr * kr; |
| 471 | const int32_t izp = (int32_t) params->input_zero_point; |
| 472 | const int32_t bzp = (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
| 473 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 474 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 475 | int32_t* packed_b = (int32_t*) packed_w; |
| 476 | if XNN_LIKELY(b != NULL) { |
| 477 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 478 | unaligned_store_s32(packed_w, bzp + b[nr_block_start + nr_block_offset]); |
| 479 | packed_w = (int32_t*) packed_w + 1; |
| 480 | } |
| 481 | } else { |
| 482 | size_t n = nr_block_size; |
| 483 | do { |
| 484 | unaligned_store_s32(packed_w, bzp); |
| 485 | packed_w = (int32_t*) packed_w + 1; |
| 486 | } while (--n != 0); |
| 487 | } |
| 488 | packed_w = (int32_t*) packed_w + (nr - nr_block_size); |
| 489 | |
| 490 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 491 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 492 | int32_t ksum = 0; |
| 493 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 494 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 495 | if (kc_idx < kc) { |
| 496 | const uint8_t kv = k[kc_idx * nc + (nr_block_start + nr_block_offset)]; |
| 497 | ksum += (int32_t) kv; |
| 498 | ((uint8_t*) packed_w)[kr_block_offset] = kv; |
| 499 | } |
| 500 | } |
| 501 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 502 | packed_w = (uint8_t*) packed_w + kr; |
| 503 | } |
| 504 | packed_w = (uint8_t*) packed_w + (nr - nr_block_size) * kr; |
| 505 | } |
| 506 | } |
| 507 | } |
| 508 | |
| 509 | void xnn_pack_qs8_gemm_io_w( |
| 510 | size_t nc, |
| 511 | size_t kc, |
| 512 | size_t nr, |
| 513 | size_t kr, |
| 514 | size_t sr, |
| 515 | const int8_t* k, |
| 516 | const int32_t* b, |
| 517 | void* packed_w, |
| 518 | const struct xnn_qs8_packing_params* params) |
| 519 | { |
| 520 | assert(nr >= sr); |
| 521 | |
| 522 | const size_t skr = sr * kr; |
| 523 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 524 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 525 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 526 | int32_t* packed_b = (int32_t*) packed_w; |
| 527 | if XNN_LIKELY(b != NULL) { |
| 528 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 529 | unaligned_store_s32(packed_w, b[nr_block_start + nr_block_offset]); |
| 530 | packed_w = (int32_t*) packed_w + 1; |
| 531 | } |
| 532 | } else { |
| 533 | size_t n = nr_block_size; |
| 534 | do { |
| 535 | unaligned_store_s32(packed_w, 0); |
| 536 | packed_w = (int32_t*) packed_w + 1; |
| 537 | } while (--n != 0); |
| 538 | } |
| 539 | packed_w = (uint32_t*) packed_w + (nr - nr_block_size); |
| 540 | |
| 541 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 542 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 543 | uint32_t ksum = 0; |
| 544 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 545 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 546 | if (kc_idx < kc) { |
| 547 | const int8_t kv = k[kc_idx * nc + (nr_block_start + nr_block_offset)]; |
| 548 | ksum += (uint32_t) kv; |
| 549 | ((int8_t*) packed_w)[kr_block_offset] = kv; |
| 550 | } |
| 551 | } |
| 552 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 553 | packed_w = (int8_t*) packed_w + kr; |
| 554 | } |
| 555 | packed_w = (int8_t*) packed_w + (nr - nr_block_size) * kr; |
| 556 | } |
| 557 | } |
| 558 | } |
| 559 | |
| 560 | void xnn_pack_f32_conv_goki_w( |
| 561 | size_t g, |
| 562 | size_t nc, |
| 563 | size_t ks, |
| 564 | size_t kc, |
| 565 | size_t nr, |
| 566 | size_t kr, |
| 567 | size_t sr, |
| 568 | const float* k, |
| 569 | const float* b, |
| 570 | float* packed_w, |
| 571 | size_t extra_bytes, |
| 572 | const void* params) |
| 573 | { |
| 574 | assert(nr >= sr); |
| 575 | |
| 576 | const size_t skr = sr * kr; |
| 577 | do { |
| 578 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 579 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 580 | if XNN_LIKELY(b != NULL) { |
| 581 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 582 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 583 | } |
| 584 | } |
| 585 | packed_w += nr; |
| 586 | |
| 587 | for (size_t ki = 0; ki < ks; ki++) { |
| 588 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 589 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 590 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 591 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 592 | if (kc_idx < kc) { |
| 593 | packed_w[kr_block_offset] = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 594 | } |
| 595 | } |
| 596 | packed_w += kr; |
| 597 | } |
| 598 | packed_w += (nr - nr_block_size) * kr; |
| 599 | } |
| 600 | } |
| 601 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 602 | } |
| 603 | k += ks * kc * nc; |
| 604 | if XNN_UNPREDICTABLE(b != NULL) { |
| 605 | b += nc; |
| 606 | } |
| 607 | } while (--g != 0); |
| 608 | } |
| 609 | |
| 610 | void xnn_pack_f16_conv_goki_w( |
| 611 | size_t g, |
| 612 | size_t nc, |
| 613 | size_t ks, |
| 614 | size_t kc, |
| 615 | size_t nr, |
| 616 | size_t kr, |
| 617 | size_t sr, |
| 618 | const uint16_t* k, |
| 619 | const uint16_t* b, |
| 620 | uint16_t* packed_w, |
| 621 | size_t extra_bytes, |
| 622 | const void* params) |
| 623 | { |
| 624 | assert(nr >= sr); |
| 625 | |
| 626 | const size_t skr = sr * kr; |
| 627 | do { |
| 628 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 629 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 630 | if XNN_LIKELY(b != NULL) { |
| 631 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 632 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 633 | } |
| 634 | } |
| 635 | packed_w += nr; |
| 636 | |
| 637 | for (size_t ki = 0; ki < ks; ki++) { |
| 638 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 639 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 640 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 641 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 642 | if (kc_idx < kc) { |
| 643 | packed_w[kr_block_offset] = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 644 | } |
| 645 | } |
| 646 | packed_w += kr; |
| 647 | } |
| 648 | packed_w += (nr - nr_block_size) * kr; |
| 649 | } |
| 650 | } |
| 651 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 652 | } |
| 653 | k += ks * kc * nc; |
| 654 | if XNN_UNPREDICTABLE(b != NULL) { |
| 655 | b += nc; |
| 656 | } |
| 657 | } while (--g != 0); |
| 658 | } |
| 659 | |
| 660 | void xnn_pack_f32_to_f16_conv_goki_w( |
| 661 | size_t g, |
| 662 | size_t nc, |
| 663 | size_t ks, |
| 664 | size_t kc, |
| 665 | size_t nr, |
| 666 | size_t kr, |
| 667 | size_t sr, |
| 668 | const float* k, |
| 669 | const float* b, |
| 670 | uint16_t* packed_w, |
| 671 | size_t extra_bytes, |
| 672 | const void* params) |
| 673 | { |
| 674 | assert(nr >= sr); |
| 675 | |
| 676 | const size_t skr = sr * kr; |
| 677 | do { |
| 678 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 679 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 680 | if XNN_LIKELY(b != NULL) { |
| 681 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 682 | packed_w[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 683 | } |
| 684 | } |
| 685 | packed_w += nr; |
| 686 | |
| 687 | for (size_t ki = 0; ki < ks; ki++) { |
| 688 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 689 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 690 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 691 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 692 | if (kc_idx < kc) { |
| 693 | packed_w[kr_block_offset] = fp16_ieee_from_fp32_value(k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]); |
| 694 | } |
| 695 | } |
| 696 | packed_w += kr; |
| 697 | } |
| 698 | packed_w += (nr - nr_block_size) * kr; |
| 699 | } |
| 700 | } |
| 701 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 702 | } |
| 703 | k += ks * kc * nc; |
| 704 | if XNN_UNPREDICTABLE(b != NULL) { |
| 705 | b += nc; |
| 706 | } |
| 707 | } while (--g != 0); |
| 708 | } |
| 709 | |
| 710 | void xnn_pack_qu8_conv_goki_w( |
| 711 | size_t g, |
| 712 | size_t nc, |
| 713 | size_t ks, |
| 714 | size_t kc, |
| 715 | size_t nr, |
| 716 | size_t kr, |
| 717 | size_t sr, |
| 718 | const uint8_t* k, |
| 719 | const int32_t* b, |
| 720 | void* packed_w, |
| 721 | size_t extra_bytes, |
| 722 | const struct xnn_qu8_packing_params* params) |
| 723 | { |
| 724 | assert(nr >= sr); |
| 725 | |
| 726 | const size_t skr = sr * kr; |
| 727 | const int32_t izp = (int32_t) params->input_zero_point; |
| 728 | const int32_t bzp = (int32_t) ks * (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
| 729 | do { |
| 730 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 731 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 732 | int32_t* packed_b = (int32_t*) packed_w; |
| 733 | if XNN_LIKELY(b != NULL) { |
| 734 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 735 | unaligned_store_s32(packed_w, bzp + b[nr_block_start + nr_block_offset]); |
| 736 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 737 | } |
| 738 | } else { |
| 739 | size_t n = nr_block_size; |
| 740 | do { |
| 741 | unaligned_store_s32(packed_w, bzp); |
| 742 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 743 | } while (--n != 0); |
| 744 | } |
| 745 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 746 | |
| 747 | for (size_t ki = 0; ki < ks; ki++) { |
| 748 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 749 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 750 | int32_t ksum = 0; |
| 751 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 752 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 753 | if (kc_idx < kc) { |
| 754 | const uint8_t kv = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 755 | ksum += (int32_t) kv; |
| 756 | ((uint8_t*) packed_w)[kr_block_offset] = kv; |
| 757 | } |
| 758 | } |
| 759 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 760 | packed_w = (uint8_t*) packed_w + kr; |
| 761 | } |
| 762 | packed_w = (uint8_t*) packed_w + (nr - nr_block_size) * kr; |
| 763 | } |
| 764 | } |
| 765 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 766 | } |
| 767 | k += ks * kc * nc; |
| 768 | if XNN_UNPREDICTABLE(b != NULL) { |
| 769 | b += nc; |
| 770 | } |
| 771 | } while (--g != 0); |
| 772 | } |
| 773 | |
| 774 | void xnn_pack_qs8_conv_goki_w( |
| 775 | size_t g, |
| 776 | size_t nc, |
| 777 | size_t ks, |
| 778 | size_t kc, |
| 779 | size_t nr, |
| 780 | size_t kr, |
| 781 | size_t sr, |
| 782 | const int8_t* k, |
| 783 | const int32_t* b, |
| 784 | void* packed_w, |
| 785 | size_t extra_bytes, |
| 786 | const struct xnn_qs8_packing_params* params) |
| 787 | { |
| 788 | assert(nr >= sr); |
| 789 | |
| 790 | const size_t skr = sr * kr; |
| 791 | const uint32_t izp = (int32_t) params->input_zero_point; |
| 792 | do { |
| 793 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 794 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 795 | int32_t* packed_b = (int32_t*) packed_w; |
| 796 | if XNN_LIKELY(b != NULL) { |
| 797 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 798 | unaligned_store_s32(packed_w, b[nr_block_start + nr_block_offset]); |
| 799 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 800 | } |
| 801 | } else { |
| 802 | size_t n = nr_block_size; |
| 803 | do { |
| 804 | unaligned_store_s32(packed_w, 0); |
| 805 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 806 | } while (--n != 0); |
| 807 | } |
| 808 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 809 | |
| 810 | for (size_t ki = 0; ki < ks; ki++) { |
| 811 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 812 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 813 | uint32_t ksum = 0; |
| 814 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 815 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 816 | if (kc_idx < kc) { |
| 817 | const int8_t kv = k[((nr_block_start + nr_block_offset) * ks + ki) * kc + kc_idx]; |
| 818 | ksum += (uint32_t) kv; |
| 819 | ((int8_t*) packed_w)[kr_block_offset] = kv; |
| 820 | } |
| 821 | } |
| 822 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 823 | packed_w = (int8_t*) packed_w + kr; |
| 824 | } |
| 825 | packed_w = (int8_t*) packed_w + (nr - nr_block_size) * kr; |
| 826 | } |
| 827 | } |
| 828 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 829 | } |
| 830 | k += ks * kc * nc; |
| 831 | if XNN_UNPREDICTABLE(b != NULL) { |
| 832 | b += nc; |
| 833 | } |
| 834 | } while (--g != 0); |
| 835 | } |
| 836 | |
| 837 | void xnn_pack_f32_conv_kgo_w( |
| 838 | size_t g, |
| 839 | size_t nc, |
| 840 | size_t ks, |
| 841 | size_t nr, |
| 842 | size_t kr, |
| 843 | size_t sr, |
| 844 | const float* k, |
| 845 | const float* b, |
| 846 | float* packed_w, |
| 847 | size_t extra_bytes, |
| 848 | const void* params) |
| 849 | { |
| 850 | assert(nr >= sr); |
| 851 | |
| 852 | for (size_t i = 0; i < g; i++) { |
| 853 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 854 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 855 | if XNN_LIKELY(b != NULL) { |
| 856 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 857 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 858 | } |
| 859 | } |
| 860 | packed_w += nr; |
| 861 | |
| 862 | for (size_t ki = 0; ki < ks; ki++) { |
| 863 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 864 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 865 | packed_w[nr_block_offset * kr] = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 866 | } |
| 867 | packed_w += nr * kr; |
| 868 | } |
| 869 | } |
| 870 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 871 | } |
| 872 | k += nc; |
| 873 | if XNN_UNPREDICTABLE(b != NULL) { |
| 874 | b += nc; |
| 875 | } |
| 876 | } |
| 877 | } |
| 878 | |
| 879 | void xnn_pack_f16_conv_kgo_w( |
| 880 | size_t g, |
| 881 | size_t nc, |
| 882 | size_t ks, |
| 883 | size_t nr, |
| 884 | size_t kr, |
| 885 | size_t sr, |
| 886 | const uint16_t* k, |
| 887 | const uint16_t* b, |
| 888 | uint16_t* packed_w, |
| 889 | size_t extra_bytes, |
| 890 | const void* params) |
| 891 | { |
| 892 | assert(nr >= sr); |
| 893 | |
| 894 | for (size_t i = 0; i < g; i++) { |
| 895 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 896 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 897 | if XNN_LIKELY(b != NULL) { |
| 898 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 899 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 900 | } |
| 901 | } |
| 902 | packed_w += nr; |
| 903 | |
| 904 | for (size_t ki = 0; ki < ks; ki++) { |
| 905 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 906 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 907 | packed_w[nr_block_offset * kr] = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 908 | } |
| 909 | packed_w += nr * kr; |
| 910 | } |
| 911 | } |
| 912 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 913 | } |
| 914 | k += nc; |
| 915 | if XNN_UNPREDICTABLE(b != NULL) { |
| 916 | b += nc; |
| 917 | } |
| 918 | } |
| 919 | } |
| 920 | |
| 921 | void xnn_pack_f32_to_f16_conv_kgo_w( |
| 922 | size_t g, |
| 923 | size_t nc, |
| 924 | size_t ks, |
| 925 | size_t nr, |
| 926 | size_t kr, |
| 927 | size_t sr, |
| 928 | const float* k, |
| 929 | const float* b, |
| 930 | uint16_t* packed_w, |
| 931 | size_t extra_bytes, |
| 932 | const void* params) |
| 933 | { |
| 934 | assert(nr >= sr); |
| 935 | |
| 936 | for (size_t i = 0; i < g; i++) { |
| 937 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 938 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 939 | if XNN_LIKELY(b != NULL) { |
| 940 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 941 | packed_w[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 942 | } |
| 943 | } |
| 944 | packed_w += nr; |
| 945 | |
| 946 | for (size_t ki = 0; ki < ks; ki++) { |
| 947 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 948 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 949 | packed_w[nr_block_offset * kr] = fp16_ieee_from_fp32_value(k[ki * g * nc + (nr_block_start + nr_block_offset)]); |
| 950 | } |
| 951 | packed_w += nr * kr; |
| 952 | } |
| 953 | } |
| 954 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 955 | } |
| 956 | k += nc; |
| 957 | if XNN_UNPREDICTABLE(b != NULL) { |
| 958 | b += nc; |
| 959 | } |
| 960 | } |
| 961 | } |
| 962 | |
| 963 | void xnn_pack_qu8_conv_kgo_w( |
| 964 | size_t g, |
| 965 | size_t nc, |
| 966 | size_t ks, |
| 967 | size_t nr, |
| 968 | size_t kr, |
| 969 | size_t sr, |
| 970 | const uint8_t* k, |
| 971 | const int32_t* b, |
| 972 | void* packed_w, |
| 973 | size_t extra_bytes, |
| 974 | const struct xnn_qu8_packing_params* params) |
| 975 | { |
| 976 | assert(nr >= sr); |
| 977 | |
| 978 | const int32_t izp = (int32_t) params->input_zero_point; |
| 979 | const int32_t bzp = (int32_t) ks * izp * (int32_t) params->kernel_zero_point; |
| 980 | for (size_t i = 0; i < g; i++) { |
| 981 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 982 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 983 | int32_t* packed_b = (int32_t*) packed_w; |
| 984 | if XNN_LIKELY(b != NULL) { |
| 985 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 986 | unaligned_store_s32(packed_w, bzp + b[nr_block_start + nr_block_offset]); |
| 987 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 988 | } |
| 989 | } else { |
| 990 | size_t n = nr_block_size; |
| 991 | do { |
| 992 | unaligned_store_s32(packed_w, bzp); |
| 993 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 994 | } while (--n != 0); |
| 995 | } |
| 996 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 997 | |
| 998 | for (size_t ki = 0; ki < ks; ki++) { |
| 999 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 1000 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 1001 | const uint8_t kv = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 1002 | ((uint8_t*) packed_w)[nr_block_offset * kr] = kv; |
| 1003 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - (int32_t) kv * izp); |
| 1004 | } |
| 1005 | packed_w = (uint8_t*) packed_w + nr * kr; |
| 1006 | } |
| 1007 | } |
| 1008 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 1009 | } |
| 1010 | k += nc; |
| 1011 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1012 | b += nc; |
| 1013 | } |
| 1014 | } |
| 1015 | } |
| 1016 | |
| 1017 | void xnn_pack_qs8_conv_kgo_w( |
| 1018 | size_t g, |
| 1019 | size_t nc, |
| 1020 | size_t ks, |
| 1021 | size_t nr, |
| 1022 | size_t kr, |
| 1023 | size_t sr, |
| 1024 | const int8_t* k, |
| 1025 | const int32_t* b, |
| 1026 | void* packed_w, |
| 1027 | size_t extra_bytes, |
| 1028 | const struct xnn_qs8_packing_params* params) |
| 1029 | { |
| 1030 | assert(nr >= sr); |
| 1031 | |
| 1032 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 1033 | for (size_t i = 0; i < g; i++) { |
| 1034 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1035 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1036 | int32_t* packed_b = (int32_t*) packed_w; |
| 1037 | if XNN_LIKELY(b != NULL) { |
| 1038 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1039 | unaligned_store_s32(packed_w, b[nr_block_start + nr_block_offset]); |
| 1040 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1041 | } |
| 1042 | } else { |
| 1043 | size_t n = nr_block_size; |
| 1044 | do { |
| 1045 | unaligned_store_s32(packed_w, 0); |
| 1046 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1047 | } while (--n != 0); |
| 1048 | } |
| 1049 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 1050 | |
| 1051 | for (size_t ki = 0; ki < ks; ki++) { |
| 1052 | for (size_t sr_block_offset = 0; sr_block_offset < sr; sr_block_offset++) { |
| 1053 | for (size_t nr_block_offset = (-sr_block_offset) & (sr - 1); nr_block_offset < nr_block_size; nr_block_offset += sr) { |
| 1054 | const int8_t kv = k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 1055 | ((int8_t*) packed_w)[nr_block_offset * kr] = kv; |
| 1056 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - (uint32_t) kv * izp); |
| 1057 | } |
| 1058 | packed_w = (int8_t*) packed_w + nr * kr; |
| 1059 | } |
| 1060 | } |
| 1061 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 1062 | } |
| 1063 | k += nc; |
| 1064 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1065 | b += nc; |
| 1066 | } |
| 1067 | } |
| 1068 | } |
| 1069 | |
| 1070 | void xnn_pack_f32_deconv_goki_w( |
| 1071 | size_t g, |
| 1072 | size_t nc, |
| 1073 | size_t kh, |
| 1074 | size_t kw, |
| 1075 | size_t kc, |
| 1076 | size_t sh, |
| 1077 | size_t sw, |
| 1078 | size_t nr, |
| 1079 | size_t kr, |
| 1080 | size_t sr, |
| 1081 | const float* k, |
| 1082 | const float* b, |
| 1083 | float* packed_w, |
| 1084 | struct subconvolution_params* subconv_params, |
| 1085 | const void* params) |
| 1086 | { |
| 1087 | assert(nr >= sr); |
| 1088 | |
| 1089 | const size_t skr = sr * kr; |
| 1090 | for (size_t i = 0; i < g; i++) { |
| 1091 | for (size_t oy = 0; oy < sh; oy++) { |
| 1092 | for (size_t ox = 0; ox < sw; ox++) { |
| 1093 | if (i == 0) { |
| 1094 | (*subconv_params++).weights = packed_w; |
| 1095 | } |
| 1096 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1097 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1098 | if XNN_LIKELY(b != NULL) { |
| 1099 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1100 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 1101 | } |
| 1102 | } |
| 1103 | packed_w += nr; |
| 1104 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1105 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1106 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1107 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1108 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1109 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1110 | if (kc_idx < kc) { |
| 1111 | packed_w[kr_block_offset] = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1112 | } |
| 1113 | } |
| 1114 | packed_w += kr; |
| 1115 | } |
| 1116 | packed_w += (nr - nr_block_size) * kr; |
| 1117 | } |
| 1118 | } |
| 1119 | } |
| 1120 | } |
| 1121 | } |
| 1122 | } |
| 1123 | k += kh * kw * kc * nc; |
| 1124 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1125 | b += nc; |
| 1126 | } |
| 1127 | } |
| 1128 | } |
| 1129 | |
| 1130 | void xnn_pack_f16_deconv_goki_w( |
| 1131 | size_t g, |
| 1132 | size_t nc, |
| 1133 | size_t kh, |
| 1134 | size_t kw, |
| 1135 | size_t kc, |
| 1136 | size_t sh, |
| 1137 | size_t sw, |
| 1138 | size_t nr, |
| 1139 | size_t kr, |
| 1140 | size_t sr, |
| 1141 | const uint16_t* k, |
| 1142 | const uint16_t* b, |
| 1143 | uint16_t* packed_w, |
| 1144 | struct subconvolution_params* subconv_params, |
| 1145 | const void* params) |
| 1146 | { |
| 1147 | assert(nr >= sr); |
| 1148 | |
| 1149 | const size_t skr = sr * kr; |
| 1150 | for (size_t i = 0; i < g; i++) { |
| 1151 | for (size_t oy = 0; oy < sh; oy++) { |
| 1152 | for (size_t ox = 0; ox < sw; ox++) { |
| 1153 | if (i == 0) { |
| 1154 | (*subconv_params++).weights = packed_w; |
| 1155 | } |
| 1156 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1157 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1158 | if XNN_LIKELY(b != NULL) { |
| 1159 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1160 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 1161 | } |
| 1162 | } |
| 1163 | packed_w += nr; |
| 1164 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1165 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1166 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1167 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1168 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1169 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1170 | if (kc_idx < kc) { |
| 1171 | packed_w[kr_block_offset] = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1172 | } |
| 1173 | } |
| 1174 | packed_w += kr; |
| 1175 | } |
| 1176 | packed_w += (nr - nr_block_size) * kr; |
| 1177 | } |
| 1178 | } |
| 1179 | } |
| 1180 | } |
| 1181 | } |
| 1182 | } |
| 1183 | k += kh * kw * kc * nc; |
| 1184 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1185 | b += nc; |
| 1186 | } |
| 1187 | } |
| 1188 | } |
| 1189 | |
| 1190 | void xnn_pack_f32_to_f16_deconv_goki_w( |
| 1191 | size_t g, |
| 1192 | size_t nc, |
| 1193 | size_t kh, |
| 1194 | size_t kw, |
| 1195 | size_t kc, |
| 1196 | size_t sh, |
| 1197 | size_t sw, |
| 1198 | size_t nr, |
| 1199 | size_t kr, |
| 1200 | size_t sr, |
| 1201 | const float* k, |
| 1202 | const float* b, |
| 1203 | uint16_t* packed_w, |
| 1204 | struct subconvolution_params* subconv_params, |
| 1205 | const void* params) |
| 1206 | { |
| 1207 | assert(nr >= sr); |
| 1208 | |
| 1209 | const size_t skr = sr * kr; |
| 1210 | for (size_t i = 0; i < g; i++) { |
| 1211 | for (size_t oy = 0; oy < sh; oy++) { |
| 1212 | for (size_t ox = 0; ox < sw; ox++) { |
| 1213 | if (i == 0) { |
| 1214 | (*subconv_params++).weights = packed_w; |
| 1215 | } |
| 1216 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1217 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1218 | if XNN_LIKELY(b != NULL) { |
| 1219 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1220 | packed_w[nr_block_offset] = fp16_ieee_from_fp32_value(b[nr_block_start + nr_block_offset]); |
| 1221 | } |
| 1222 | } |
| 1223 | packed_w += nr; |
| 1224 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1225 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1226 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1227 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1228 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1229 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1230 | if (kc_idx < kc) { |
| 1231 | packed_w[kr_block_offset] = fp16_ieee_from_fp32_value(k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]); |
| 1232 | } |
| 1233 | } |
| 1234 | packed_w += kr; |
| 1235 | } |
| 1236 | packed_w += (nr - nr_block_size) * kr; |
| 1237 | } |
| 1238 | } |
| 1239 | } |
| 1240 | } |
| 1241 | } |
| 1242 | } |
| 1243 | k += kh * kw * kc * nc; |
| 1244 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1245 | b += nc; |
| 1246 | } |
| 1247 | } |
| 1248 | } |
| 1249 | |
| 1250 | void xnn_pack_qs8_deconv_goki_w( |
| 1251 | size_t g, |
| 1252 | size_t nc, |
| 1253 | size_t kh, |
| 1254 | size_t kw, |
| 1255 | size_t kc, |
| 1256 | size_t sh, |
| 1257 | size_t sw, |
| 1258 | size_t nr, |
| 1259 | size_t kr, |
| 1260 | size_t sr, |
| 1261 | const int8_t* k, |
| 1262 | const int32_t* b, |
| 1263 | void* packed_w, |
| 1264 | struct subconvolution_params* subconv_params, |
| 1265 | const struct xnn_qs8_packing_params* params) |
| 1266 | { |
| 1267 | assert(nr >= sr); |
| 1268 | |
| 1269 | const size_t skr = sr * kr; |
| 1270 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 1271 | for (size_t i = 0; i < g; i++) { |
| 1272 | for (size_t oy = 0; oy < sh; oy++) { |
| 1273 | for (size_t ox = 0; ox < sw; ox++) { |
| 1274 | if (i == 0) { |
| 1275 | (*subconv_params++).weights = packed_w; |
| 1276 | } |
| 1277 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1278 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1279 | int32_t* packed_b = (int32_t*) packed_w; |
| 1280 | if XNN_LIKELY(b != 0) { |
| 1281 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1282 | unaligned_store_s32(packed_w, b[nr_block_start + nr_block_offset]); |
| 1283 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1284 | } |
| 1285 | } else { |
| 1286 | size_t n = nr_block_size; |
| 1287 | do { |
| 1288 | unaligned_store_s32(packed_w, 0); |
| 1289 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1290 | } while (--n != 0); |
| 1291 | } |
| 1292 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 1293 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1294 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1295 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1296 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1297 | uint32_t ksum = 0; |
| 1298 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1299 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1300 | if (kc_idx < kc) { |
| 1301 | const int8_t kv = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1302 | ksum += (uint32_t) kv; |
| 1303 | ((int8_t*) packed_w)[kr_block_offset] = kv; |
| 1304 | } |
| 1305 | } |
| 1306 | unaligned_indexed_store_u32(packed_b, nr_block_offset, unaligned_indexed_load_u32(packed_b, nr_block_offset) - ksum * izp); |
| 1307 | packed_w = (int8_t*) packed_w + kr; |
| 1308 | } |
| 1309 | packed_w = (int8_t*) packed_w + (nr - nr_block_size) * kr; |
| 1310 | } |
| 1311 | } |
| 1312 | } |
| 1313 | } |
| 1314 | } |
| 1315 | } |
| 1316 | k += kh * kw * kc * nc; |
| 1317 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1318 | b += nc; |
| 1319 | } |
| 1320 | } |
| 1321 | } |
| 1322 | |
| 1323 | void xnn_pack_qu8_deconv_goki_w( |
| 1324 | size_t g, |
| 1325 | size_t nc, |
| 1326 | size_t kh, |
| 1327 | size_t kw, |
| 1328 | size_t kc, |
| 1329 | size_t sh, |
| 1330 | size_t sw, |
| 1331 | size_t nr, |
| 1332 | size_t kr, |
| 1333 | size_t sr, |
| 1334 | const uint8_t* k, |
| 1335 | const int32_t* b, |
| 1336 | void* packed_w, |
| 1337 | struct subconvolution_params* subconv_params, |
| 1338 | const struct xnn_qu8_packing_params* params) |
| 1339 | { |
| 1340 | assert(nr >= sr); |
| 1341 | |
| 1342 | const size_t skr = sr * kr; |
| 1343 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1344 | const int32_t kzp = (int32_t) params->kernel_zero_point; |
| 1345 | for (size_t i = 0; i < g; i++) { |
| 1346 | for (size_t oy = 0; oy < sh; oy++) { |
| 1347 | for (size_t ox = 0; ox < sw; ox++) { |
| 1348 | if (i == 0) { |
| 1349 | (*subconv_params++).weights = packed_w; |
| 1350 | } |
| 1351 | 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; |
| 1352 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1353 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1354 | int32_t* packed_b = (int32_t*) packed_w; |
| 1355 | if XNN_LIKELY(b != 0) { |
| 1356 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1357 | unaligned_store_s32(packed_w, bzp + b[nr_block_start + nr_block_offset]); |
| 1358 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1359 | } |
| 1360 | } else { |
| 1361 | size_t n = nr_block_size; |
| 1362 | do { |
| 1363 | unaligned_store_s32(packed_w, bzp); |
| 1364 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1365 | } while (--n != 0); |
| 1366 | } |
| 1367 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 1368 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1369 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1370 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1371 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1372 | int32_t ksum = 0; |
| 1373 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1374 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1375 | if (kc_idx < kc) { |
| 1376 | const uint8_t kv = k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + kc_idx]; |
| 1377 | ksum += (int32_t) kv; |
| 1378 | ((uint8_t*) packed_w)[kr_block_offset] = kv; |
| 1379 | } |
| 1380 | } |
| 1381 | unaligned_indexed_store_s32(packed_b, nr_block_offset, unaligned_indexed_load_s32(packed_b, nr_block_offset) - ksum * izp); |
| 1382 | packed_w = (uint8_t*) packed_w + kr; |
| 1383 | } |
| 1384 | packed_w = (uint8_t*) packed_w + (nr - nr_block_size) * kr; |
| 1385 | } |
| 1386 | } |
| 1387 | } |
| 1388 | } |
| 1389 | } |
| 1390 | } |
| 1391 | k += kh * kw * kc * nc; |
| 1392 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1393 | b += nc; |
| 1394 | } |
| 1395 | } |
| 1396 | } |
| 1397 | |
| 1398 | void xnn_pack_f32_dwconv_ghw_w( |
| 1399 | size_t primary_tile, |
| 1400 | size_t h, |
| 1401 | size_t w, |
| 1402 | size_t c, |
| 1403 | size_t cr, |
| 1404 | const float* k, |
| 1405 | const float* b, |
| 1406 | float* packed_w, |
| 1407 | size_t extra_bytes, |
| 1408 | const void* params) |
| 1409 | { |
| 1410 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1411 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1412 | if XNN_LIKELY(b != NULL) { |
| 1413 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1414 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1415 | } |
| 1416 | } else { |
| 1417 | size_t n = cr_block_size; |
| 1418 | do { |
| 1419 | *packed_w++ = 0.0f; |
| 1420 | } while (--n != 0); |
| 1421 | } |
| 1422 | packed_w += cr - cr_block_size; |
| 1423 | for (size_t x = 0; x < w; x++) { |
| 1424 | for (size_t y = 0; y < h; y++) { |
| 1425 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1426 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1427 | *packed_w++ = kv; |
| 1428 | } |
| 1429 | packed_w += cr - cr_block_size; |
| 1430 | } |
| 1431 | } |
| 1432 | packed_w += (primary_tile - (h * w)) * cr_block_size; |
| 1433 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 1434 | } |
| 1435 | } |
| 1436 | |
| 1437 | void xnn_pack_f16_dwconv_ghw_w( |
| 1438 | size_t primary_tile, |
| 1439 | size_t h, |
| 1440 | size_t w, |
| 1441 | size_t c, |
| 1442 | size_t cr, |
| 1443 | const uint16_t* k, |
| 1444 | const uint16_t* b, |
| 1445 | uint16_t* packed_w, |
| 1446 | size_t extra_bytes, |
| 1447 | const void* params) |
| 1448 | { |
| 1449 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1450 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1451 | if XNN_LIKELY(b != NULL) { |
| 1452 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1453 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1454 | } |
| 1455 | } else { |
| 1456 | size_t n = cr_block_size; |
| 1457 | do { |
| 1458 | *packed_w++ = 0; |
| 1459 | } while (--n != 0); |
| 1460 | } |
| 1461 | packed_w += cr - cr_block_size; |
| 1462 | for (size_t x = 0; x < w; x++) { |
| 1463 | for (size_t y = 0; y < h; y++) { |
| 1464 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1465 | const uint16_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1466 | *packed_w++ = kv; |
| 1467 | } |
| 1468 | packed_w += cr - cr_block_size; |
| 1469 | } |
| 1470 | } |
| 1471 | packed_w += (primary_tile - (h * w)) * cr_block_size; |
| 1472 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 1473 | } |
| 1474 | } |
| 1475 | |
| 1476 | void xnn_pack_f32_to_f16_dwconv_ghw_w( |
| 1477 | size_t primary_tile, |
| 1478 | size_t h, |
| 1479 | size_t w, |
| 1480 | size_t c, |
| 1481 | size_t cr, |
| 1482 | const float* k, |
| 1483 | const float* b, |
| 1484 | uint16_t* packed_w, |
| 1485 | size_t extra_bytes, |
| 1486 | const void* params) |
| 1487 | { |
| 1488 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1489 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1490 | if XNN_LIKELY(b != NULL) { |
| 1491 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1492 | *packed_w++ = fp16_ieee_from_fp32_value(b[cr_block_start + cr_block_offset]); |
| 1493 | } |
| 1494 | } else { |
| 1495 | size_t n = cr_block_size; |
| 1496 | do { |
| 1497 | *packed_w++ = 0; |
| 1498 | } while (--n != 0); |
| 1499 | } |
| 1500 | packed_w += cr - cr_block_size; |
| 1501 | for (size_t x = 0; x < w; x++) { |
| 1502 | for (size_t y = 0; y < h; y++) { |
| 1503 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1504 | const uint16_t kv = fp16_ieee_from_fp32_value(k[((cr_block_start + cr_block_offset) * h + y) * w + x]); |
| 1505 | *packed_w++ = kv; |
| 1506 | } |
| 1507 | packed_w += cr - cr_block_size; |
| 1508 | } |
| 1509 | } |
| 1510 | packed_w += (primary_tile - (h * w)) * cr_block_size; |
| 1511 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 1512 | } |
| 1513 | } |
| 1514 | |
| 1515 | void xnn_pack_qu8_dwconv_ghw_w( |
| 1516 | size_t primary_tile, |
| 1517 | size_t h, |
| 1518 | size_t w, |
| 1519 | size_t c, |
| 1520 | size_t cr, |
| 1521 | const uint8_t* k, |
| 1522 | const int32_t* b, |
| 1523 | void* packed_w, |
| 1524 | size_t extra_bytes, |
| 1525 | const struct xnn_qu8_packing_params* params) |
| 1526 | { |
| 1527 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1528 | const int32_t boff = (int32_t) h * (int32_t) w * izp * (int32_t) params->kernel_zero_point; |
| 1529 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1530 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1531 | int32_t* packed_b = (int32_t*) packed_w; |
| 1532 | if XNN_LIKELY(b != NULL) { |
| 1533 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1534 | unaligned_store_s32(packed_w, boff + b[cr_block_start + cr_block_offset]); |
| 1535 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1536 | } |
| 1537 | } else { |
| 1538 | size_t n = cr_block_size; |
| 1539 | do { |
| 1540 | unaligned_store_s32(packed_w, boff); |
| 1541 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1542 | } while (--n != 0); |
| 1543 | } |
| 1544 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1545 | for (size_t x = 0; x < w; x++) { |
| 1546 | for (size_t y = 0; y < h; y++) { |
| 1547 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1548 | const uint8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1549 | unaligned_indexed_store_s32(packed_b, cr_block_offset, unaligned_indexed_load_s32(packed_b, cr_block_offset) - (int32_t) kv * izp); |
| 1550 | *((uint8_t*) packed_w) = kv; |
| 1551 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 1552 | } |
| 1553 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(uint8_t)); |
| 1554 | } |
| 1555 | } |
| 1556 | packed_w = (void*) ((uintptr_t) packed_w + (primary_tile - (h * w)) * cr_block_size * sizeof(uint8_t)); |
| 1557 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 1558 | } |
| 1559 | } |
| 1560 | |
| 1561 | void xnn_pack_qs8_dwconv_ghw_w( |
| 1562 | size_t primary_tile, |
| 1563 | size_t h, |
| 1564 | size_t w, |
| 1565 | size_t c, |
| 1566 | size_t cr, |
| 1567 | const int8_t* k, |
| 1568 | const int32_t* b, |
| 1569 | void* packed_w, |
| 1570 | size_t extra_bytes, |
| 1571 | const struct xnn_qs8_packing_params* params) |
| 1572 | { |
| 1573 | const uint32_t izp = (uint32_t) params->input_zero_point; |
| 1574 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1575 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1576 | int32_t* packed_b = (int32_t*) packed_w; |
| 1577 | if XNN_LIKELY(b != NULL) { |
| 1578 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1579 | unaligned_store_s32(packed_w, b[cr_block_start + cr_block_offset]); |
| 1580 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1581 | } |
| 1582 | } else { |
| 1583 | size_t n = cr_block_size; |
| 1584 | do { |
| 1585 | unaligned_store_s32(packed_w, 0); |
| 1586 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1587 | } while (--n != 0); |
| 1588 | } |
| 1589 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1590 | for (size_t x = 0; x < w; x++) { |
| 1591 | for (size_t y = 0; y < h; y++) { |
| 1592 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1593 | const int8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1594 | unaligned_indexed_store_u32(packed_b, cr_block_offset, unaligned_indexed_load_u32(packed_b, cr_block_offset) - (uint32_t) kv * izp); |
| 1595 | *((int8_t*) packed_w) = kv; |
| 1596 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 1597 | } |
| 1598 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int8_t)); |
| 1599 | } |
| 1600 | } |
| 1601 | packed_w = (void*) ((uintptr_t) packed_w + (primary_tile - (h * w)) * cr_block_size * sizeof(int8_t)); |
| 1602 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 1603 | } |
| 1604 | } |
| 1605 | |
| 1606 | void xnn_pack_f32_dwconv_hwg_w( |
| 1607 | size_t primary_tile, |
| 1608 | size_t h, |
| 1609 | size_t w, |
| 1610 | size_t c, |
| 1611 | size_t cr, |
| 1612 | const float* k, |
| 1613 | const float* b, |
| 1614 | float* packed_w, |
| 1615 | size_t extra_bytes, |
| 1616 | const void* params) |
| 1617 | { |
| 1618 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1619 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1620 | if XNN_LIKELY(b != NULL) { |
| 1621 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1622 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1623 | } |
| 1624 | } else { |
| 1625 | size_t n = cr_block_size; |
| 1626 | do { |
| 1627 | *packed_w++ = 0.0f; |
| 1628 | } while (--n != 0); |
| 1629 | } |
| 1630 | packed_w += cr - cr_block_size; |
| 1631 | for (size_t x = 0; x < w; x++) { |
| 1632 | for (size_t y = 0; y < h; y++) { |
| 1633 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1634 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1635 | *packed_w++ = kv; |
| 1636 | } |
| 1637 | packed_w += cr - cr_block_size; |
| 1638 | } |
| 1639 | } |
| 1640 | packed_w += (primary_tile - (h * w)) * cr_block_size; |
| 1641 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
| 1642 | } |
| 1643 | } |
| 1644 | |
| 1645 | void xnn_pack_f16_dwconv_hwg_w( |
| 1646 | size_t primary_tile, |
| 1647 | size_t h, |
| 1648 | size_t w, |
| 1649 | size_t c, |
| 1650 | size_t cr, |
| 1651 | const uint16_t* k, |
| 1652 | const uint16_t* b, |
| 1653 | uint16_t* packed_w, |
| 1654 | size_t extra_bytes, |
| 1655 | const void* params) |
| 1656 | { |
| 1657 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1658 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1659 | if XNN_LIKELY(b != NULL) { |
| 1660 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1661 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1662 | } |
| 1663 | } else { |
| 1664 | size_t n = cr_block_size; |
| 1665 | do { |
| 1666 | *packed_w++ = 0; |
| 1667 | } while (--n != 0); |
| 1668 | } |
| 1669 | packed_w += cr - cr_block_size; |
| 1670 | for (size_t x = 0; x < w; x++) { |
| 1671 | for (size_t y = 0; y < h; y++) { |
| 1672 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1673 | const uint16_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1674 | *packed_w++ = kv; |
| 1675 | } |
| 1676 | packed_w += cr - cr_block_size; |
| 1677 | } |
| 1678 | } |
| 1679 | packed_w += (primary_tile - (h * w)) * cr_block_size; |
| 1680 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 1681 | } |
| 1682 | } |
| 1683 | |
| 1684 | void xnn_pack_f32_to_f16_dwconv_hwg_w( |
| 1685 | size_t primary_tile, |
| 1686 | size_t h, |
| 1687 | size_t w, |
| 1688 | size_t c, |
| 1689 | size_t cr, |
| 1690 | const float* k, |
| 1691 | const float* b, |
| 1692 | uint16_t* packed_w, |
| 1693 | size_t extra_bytes, |
| 1694 | const void* params) |
| 1695 | { |
| 1696 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1697 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1698 | if XNN_LIKELY(b != NULL) { |
| 1699 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1700 | *packed_w++ = fp16_ieee_from_fp32_value(b[cr_block_start + cr_block_offset]); |
| 1701 | } |
| 1702 | } else { |
| 1703 | size_t n = cr_block_size; |
| 1704 | do { |
| 1705 | *packed_w++ = 0; |
| 1706 | } while (--n != 0); |
| 1707 | } |
| 1708 | packed_w += cr - cr_block_size; |
| 1709 | for (size_t x = 0; x < w; x++) { |
| 1710 | for (size_t y = 0; y < h; y++) { |
| 1711 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1712 | const uint16_t kv = fp16_ieee_from_fp32_value(k[(y * w + x) * c + (cr_block_start + cr_block_offset)]); |
| 1713 | *packed_w++ = kv; |
| 1714 | } |
| 1715 | packed_w += cr - cr_block_size; |
| 1716 | } |
| 1717 | } |
| 1718 | packed_w += (primary_tile - (h * w)) * cr_block_size; |
| 1719 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
| 1720 | } |
| 1721 | } |
| 1722 | |
| 1723 | void xnn_pack_qu8_dwconv_hwg_w( |
| 1724 | size_t primary_tile, |
| 1725 | size_t h, |
| 1726 | size_t w, |
| 1727 | size_t c, |
| 1728 | size_t cr, |
| 1729 | const uint8_t* k, |
| 1730 | const int32_t* b, |
| 1731 | void* packed_w, |
| 1732 | size_t extra_bytes, |
| 1733 | const struct xnn_qu8_packing_params* params) |
| 1734 | { |
| 1735 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1736 | const int32_t boff = (int32_t) h * (int32_t) w * izp * (int32_t) params->kernel_zero_point; |
| 1737 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1738 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1739 | int32_t* packed_b = (int32_t*) packed_w; |
| 1740 | if XNN_LIKELY(b != NULL) { |
| 1741 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1742 | unaligned_store_s32(packed_w, boff + b[cr_block_start + cr_block_offset]); |
| 1743 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1744 | } |
| 1745 | } else { |
| 1746 | size_t n = cr_block_size; |
| 1747 | do { |
| 1748 | unaligned_store_s32(packed_w, boff); |
| 1749 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1750 | } while (--n != 0); |
| 1751 | } |
| 1752 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1753 | for (size_t x = 0; x < w; x++) { |
| 1754 | for (size_t y = 0; y < h; y++) { |
| 1755 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1756 | const uint8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1757 | unaligned_indexed_store_s32(packed_b, cr_block_offset, unaligned_indexed_load_s32(packed_b, cr_block_offset) - (int32_t) kv * izp); |
| 1758 | *((uint8_t*) packed_w) = kv; |
| 1759 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 1760 | } |
| 1761 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(uint8_t)); |
| 1762 | } |
| 1763 | } |
| 1764 | packed_w = (void*) ((uintptr_t) packed_w + (primary_tile - (h * w)) * cr_block_size * sizeof(uint8_t)); |
| 1765 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 1766 | } |
| 1767 | } |
| 1768 | |
| 1769 | void xnn_pack_qs8_dwconv_hwg_w( |
| 1770 | size_t primary_tile, |
| 1771 | size_t h, |
| 1772 | size_t w, |
| 1773 | size_t c, |
| 1774 | size_t cr, |
| 1775 | const int8_t* k, |
| 1776 | const int32_t* b, |
| 1777 | void* packed_w, |
| 1778 | size_t extra_bytes, |
| 1779 | const struct xnn_qs8_packing_params* params) |
| 1780 | { |
| 1781 | const uint32_t izp = (int32_t) params->input_zero_point; |
| 1782 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1783 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1784 | int32_t* packed_b = (int32_t*) packed_w; |
| 1785 | if XNN_LIKELY(b != NULL) { |
| 1786 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1787 | unaligned_store_s32(packed_w, b[cr_block_start + cr_block_offset]); |
| 1788 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1789 | } |
| 1790 | } else { |
| 1791 | size_t n = cr_block_size; |
| 1792 | do { |
| 1793 | unaligned_store_s32(packed_w, 0); |
| 1794 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1795 | } while (--n != 0); |
| 1796 | } |
| 1797 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1798 | for (size_t x = 0; x < w; x++) { |
| 1799 | for (size_t y = 0; y < h; y++) { |
| 1800 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1801 | const int8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1802 | unaligned_indexed_store_u32(packed_b, cr_block_offset, unaligned_indexed_load_u32(packed_b, cr_block_offset) - (uint32_t) kv * izp); |
| 1803 | *((int8_t*) packed_w) = kv; |
| 1804 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 1805 | } |
| 1806 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int8_t)); |
| 1807 | } |
| 1808 | } |
| 1809 | packed_w = (void*) ((uintptr_t) packed_w + (primary_tile - (h * w)) * cr_block_size * sizeof(int8_t)); |
| 1810 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
| 1811 | } |
| 1812 | } |
| 1813 | |
| 1814 | void xnn_pack_f32_gemminc_goi_w( |
| 1815 | size_t g, |
| 1816 | size_t nc, |
| 1817 | size_t kc, |
| 1818 | size_t nr, |
| 1819 | size_t kr, |
| 1820 | size_t sr, |
| 1821 | const float* k, |
| 1822 | float* packed_w, |
| 1823 | const void* params) |
| 1824 | { |
| 1825 | const size_t skr = sr * kr; |
| 1826 | do { |
| 1827 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1828 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1829 | |
| 1830 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1831 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1832 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1833 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1834 | if (kc_idx < kc) { |
| 1835 | packed_w[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 1836 | } |
| 1837 | } |
| 1838 | packed_w += kr; |
| 1839 | } |
| 1840 | packed_w += (nr - nr_block_size) * kr; |
| 1841 | } |
| 1842 | } |
| 1843 | k += nc * kc; |
| 1844 | } while (--g != 0); |
| 1845 | } |
| 1846 | |
| 1847 | void xnn_pack_f16_gemminc_goi_w( |
| 1848 | size_t g, |
| 1849 | size_t nc, |
| 1850 | size_t kc, |
| 1851 | size_t nr, |
| 1852 | size_t kr, |
| 1853 | size_t sr, |
| 1854 | const uint16_t* k, |
| 1855 | uint16_t* packed_w, |
| 1856 | const void* params) |
| 1857 | { |
| 1858 | const size_t skr = sr * kr; |
| 1859 | do { |
| 1860 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1861 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1862 | |
| 1863 | for (size_t kr_block_start = 0; kr_block_start < round_up_po2(kc, skr); kr_block_start += kr) { |
| 1864 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1865 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1866 | const size_t kc_idx = round_down_po2(kr_block_start, skr) + ((kr_block_start + kr_block_offset + nr_block_offset * kr) & (skr - 1)); |
| 1867 | if (kc_idx < kc) { |
| 1868 | packed_w[kr_block_offset] = k[(nr_block_start + nr_block_offset) * kc + kc_idx]; |
| 1869 | } |
| 1870 | } |
| 1871 | packed_w += kr; |
| 1872 | } |
| 1873 | packed_w += (nr - nr_block_size) * kr; |
| 1874 | } |
| 1875 | } |
| 1876 | k += nc * kc; |
| 1877 | } while (--g != 0); |
| 1878 | } |
| 1879 | |
| 1880 | void xnn_pack_f32_dconv_oki_w( |
| 1881 | size_t nc, |
| 1882 | size_t kc, |
| 1883 | size_t nr, |
| 1884 | size_t kh, |
| 1885 | size_t kw, |
| 1886 | const float* k, |
| 1887 | const float* b, |
| 1888 | float* packed_w, |
| 1889 | const void* params) |
| 1890 | { |
| 1891 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1892 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1893 | if XNN_LIKELY(b != NULL) { |
| 1894 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1895 | *packed_w++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 1896 | } |
| 1897 | } else { |
| 1898 | size_t n = nr; |
| 1899 | do { |
| 1900 | *packed_w++ = 0.0f; |
| 1901 | } while (--n != 0); |
| 1902 | } |
| 1903 | |
| 1904 | for (size_t kx = 0; kx < kw; kx++) { |
| 1905 | for (size_t c = 0; c < kc; c++) { |
| 1906 | for (size_t ky = 0; ky < kh; ky++) { |
| 1907 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1908 | *packed_w++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 1909 | } |
| 1910 | } |
| 1911 | } |
| 1912 | } |
| 1913 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1914 | b += nr; |
| 1915 | } |
| 1916 | } |
| 1917 | } |
| 1918 | |
| 1919 | void xnn_pack_f16_dconv_oki_w( |
| 1920 | size_t nc, |
| 1921 | size_t kc, |
| 1922 | size_t nr, |
| 1923 | size_t kh, |
| 1924 | size_t kw, |
| 1925 | const uint16_t* k, |
| 1926 | const uint16_t* b, |
| 1927 | uint16_t* packed_w, |
| 1928 | const void* params) |
| 1929 | { |
| 1930 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1931 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1932 | if XNN_LIKELY(b != NULL) { |
| 1933 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1934 | *packed_w++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 1935 | } |
| 1936 | } else { |
| 1937 | size_t n = nr; |
| 1938 | do { |
| 1939 | *packed_w++ = 0; |
| 1940 | } while (--n != 0); |
| 1941 | } |
| 1942 | |
| 1943 | for (size_t kx = 0; kx < kw; kx++) { |
| 1944 | for (size_t c = 0; c < kc; c++) { |
| 1945 | for (size_t ky = 0; ky < kh; ky++) { |
| 1946 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1947 | *packed_w++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 1948 | } |
| 1949 | } |
| 1950 | } |
| 1951 | } |
| 1952 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1953 | b += nr; |
| 1954 | } |
| 1955 | } |
| 1956 | } |
| 1957 | |
| 1958 | void xnn_pack_f32_chw_dwconv_ghw_w( |
| 1959 | size_t kernel_size, |
| 1960 | size_t groups, |
| 1961 | const float* kernel, |
| 1962 | const float* bias, |
| 1963 | float* packed_weights, |
| 1964 | const void* params) |
| 1965 | { |
| 1966 | for (size_t g = 0; g < groups; g++) { |
| 1967 | if XNN_LIKELY(bias != NULL) { |
| 1968 | *packed_weights = *bias++; |
| 1969 | } else { |
| 1970 | *packed_weights = 0.0f; |
| 1971 | } |
| 1972 | packed_weights += 1; |
| 1973 | for (size_t i = 0; i < kernel_size; i++) { |
| 1974 | *packed_weights++ = kernel[g * kernel_size + i]; |
| 1975 | } |
| 1976 | } |
| 1977 | } |
| 1978 | |
| 1979 | void xnn_pack_f16_chw_dwconv_ghw_w( |
| 1980 | size_t kernel_size, |
| 1981 | size_t groups, |
| 1982 | const uint16_t* kernel, |
| 1983 | const uint16_t* bias, |
| 1984 | uint16_t* packed_weights, |
| 1985 | const void* params) |
| 1986 | { |
| 1987 | for (size_t g = 0; g < groups; g++) { |
| 1988 | if XNN_LIKELY(bias != NULL) { |
| 1989 | *packed_weights = *bias++; |
| 1990 | } else { |
| 1991 | *packed_weights = 0; |
| 1992 | } |
| 1993 | packed_weights += 1; |
| 1994 | for (size_t i = 0; i < kernel_size; i++) { |
| 1995 | *packed_weights++ = kernel[g * kernel_size + i]; |
| 1996 | } |
| 1997 | } |
| 1998 | } |
| 1999 | |
| 2000 | void xnn_pack_f32_chw_dwconv_hwg_w( |
| 2001 | size_t kernel_size, |
| 2002 | size_t groups, |
| 2003 | const float* kernel, |
| 2004 | const float* bias, |
| 2005 | float* packed_weights, |
| 2006 | const void* params) |
| 2007 | { |
| 2008 | for (size_t g = 0; g < groups; g++) { |
| 2009 | if XNN_LIKELY(bias != NULL) { |
| 2010 | *packed_weights = *bias++; |
| 2011 | } else { |
| 2012 | *packed_weights = 0.0f; |
| 2013 | } |
| 2014 | packed_weights += 1; |
| 2015 | for (size_t i = 0; i < kernel_size; i++) { |
| 2016 | *packed_weights++ = kernel[i * groups + g]; |
| 2017 | } |
| 2018 | } |
| 2019 | } |
| 2020 | |
| 2021 | void xnn_pack_f32_vmulcaddc_w( |
| 2022 | size_t c, |
| 2023 | size_t cr, |
| 2024 | const float* s, |
| 2025 | const float* b, |
| 2026 | float* packed_w, |
| 2027 | const void* params) |
| 2028 | { |
| 2029 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2030 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2031 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2032 | *packed_w++ = s[cr_block_start + cr_block_offset]; |
| 2033 | } |
| 2034 | packed_w += cr - cr_block_size; |
| 2035 | if XNN_LIKELY(b != NULL) { |
| 2036 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2037 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 2038 | } |
| 2039 | } else { |
| 2040 | size_t n = cr_block_size; |
| 2041 | do { |
| 2042 | *packed_w++ = 0.0f; |
| 2043 | } while (--n != 0); |
| 2044 | } |
| 2045 | packed_w += cr - cr_block_size; |
| 2046 | } |
| 2047 | } |
| 2048 | |
| 2049 | void xnn_pack_f16_vmulcaddc_w( |
| 2050 | size_t c, |
| 2051 | size_t cr, |
| 2052 | const uint16_t* s, |
| 2053 | const uint16_t* b, |
| 2054 | uint16_t* packed_w, |
| 2055 | const void* params) |
| 2056 | { |
| 2057 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2058 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2059 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2060 | *packed_w++ = s[cr_block_start + cr_block_offset]; |
| 2061 | } |
| 2062 | packed_w += cr - cr_block_size; |
| 2063 | if XNN_LIKELY(b != NULL) { |
| 2064 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2065 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 2066 | } |
| 2067 | } else { |
| 2068 | size_t n = cr_block_size; |
| 2069 | do { |
| 2070 | *packed_w++ = 0; |
| 2071 | } while (--n != 0); |
| 2072 | } |
| 2073 | packed_w += cr - cr_block_size; |
| 2074 | } |
| 2075 | } |
| 2076 | |
| 2077 | void xnn_pack_f32_to_f16_vmulcaddc_w( |
| 2078 | size_t c, |
| 2079 | size_t cr, |
| 2080 | const float* s, |
| 2081 | const float* b, |
| 2082 | uint16_t* packed_w, |
| 2083 | const void* params) |
| 2084 | { |
| 2085 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 2086 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 2087 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2088 | *packed_w++ = fp16_ieee_from_fp32_value(s[cr_block_start + cr_block_offset]); |
| 2089 | } |
| 2090 | packed_w += cr - cr_block_size; |
| 2091 | if XNN_LIKELY(b != NULL) { |
| 2092 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 2093 | *packed_w++ = fp16_ieee_from_fp32_value(b[cr_block_start + cr_block_offset]); |
| 2094 | } |
| 2095 | } else { |
| 2096 | size_t n = cr_block_size; |
| 2097 | do { |
| 2098 | *packed_w++ = 0; |
| 2099 | } while (--n != 0); |
| 2100 | } |
| 2101 | packed_w += cr - cr_block_size; |
| 2102 | } |
| 2103 | } |
| 2104 | |
| 2105 | void xnn_pack_f32_prelu_w( |
| 2106 | size_t c, |
| 2107 | const float* s, |
| 2108 | float* packed_w) |
| 2109 | { |
| 2110 | memcpy(packed_w, s, c * sizeof(float)); |
| 2111 | } |
| 2112 | |
| 2113 | void xnn_pack_f16_prelu_w( |
| 2114 | size_t c, |
| 2115 | const uint16_t* s, |
| 2116 | uint16_t* packed_w) |
| 2117 | { |
| 2118 | memcpy(packed_w, s, c * sizeof(uint16_t)); |
| 2119 | } |
| 2120 | |
| 2121 | void xnn_pack_f32_to_f16_prelu_w( |
| 2122 | size_t c, |
| 2123 | const float* s, |
| 2124 | uint16_t* packed_w) |
| 2125 | { |
| 2126 | do { |
| 2127 | *packed_w++ = fp16_ieee_from_fp32_value(*s++); |
| 2128 | } while (--c != 0); |
| 2129 | } |
| 2130 |
Generated on 2022-Sep-26 from project tensorflow revision f2e850b6cce
Powered by 
Code Browser 2.1
Generator usage only permitted with license.