Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame^] | 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 | |
| 12 | #include <xnnpack/math.h> |
| 13 | #include <xnnpack/pack.h> |
| 14 | |
| 15 | |
| 16 | void xnn_pack_q8_gemm_goi_w( |
| 17 | size_t g, |
| 18 | size_t nc, |
| 19 | size_t kc, |
| 20 | uint32_t nr, |
| 21 | uint32_t kr, |
| 22 | uint8_t izp, |
| 23 | uint8_t kzp, |
| 24 | const uint8_t* k, |
| 25 | const int32_t* b, |
| 26 | void* packed_w) |
| 27 | { |
| 28 | const int32_t boff = (int32_t) kc * (int32_t) izp * (int32_t) kzp; |
| 29 | do { |
| 30 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 31 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 32 | int32_t* packed_b = (int32_t*) packed_w; |
| 33 | if XNN_LIKELY(b != NULL) { |
| 34 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 35 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 36 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 37 | } |
| 38 | } else { |
| 39 | size_t n = nr_block_size; |
| 40 | do { |
| 41 | *((int32_t*) packed_w) = boff; |
| 42 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 43 | } while (--n != 0); |
| 44 | } |
| 45 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 46 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 47 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 48 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 49 | int32_t ksum = 0; |
| 50 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 51 | const uint8_t kv = k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 52 | ksum += (int32_t) kv; |
| 53 | *((uint8_t*) packed_w) = kv; |
| 54 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 55 | } |
| 56 | packed_b[nr_block_offset] -= ksum * (int32_t) izp; |
| 57 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 58 | } |
| 59 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 60 | } |
| 61 | } |
| 62 | k += nc * kc; |
| 63 | if XNN_UNPREDICTABLE(b != NULL) { |
| 64 | b += nc; |
| 65 | } |
| 66 | } while (--g != 0); |
| 67 | } |
| 68 | |
| 69 | void xnn_pack_q8_gemm_io_w( |
| 70 | size_t nc, |
| 71 | size_t kc, |
| 72 | uint32_t nr, |
| 73 | uint32_t kr, |
| 74 | uint8_t izp, |
| 75 | uint8_t kzp, |
| 76 | const uint8_t* k, |
| 77 | const int32_t* b, |
| 78 | void* packed_w) |
| 79 | { |
| 80 | const int32_t boff = (int32_t) kc * (int32_t) izp * (int32_t) kzp; |
| 81 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 82 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 83 | int32_t* packed_b = (int32_t*) packed_w; |
| 84 | if XNN_LIKELY(b != NULL) { |
| 85 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 86 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 87 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 88 | } |
| 89 | } else { |
| 90 | size_t n = nr_block_size; |
| 91 | do { |
| 92 | *((int32_t*) packed_w) = boff; |
| 93 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 94 | } while (--n != 0); |
| 95 | } |
| 96 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 97 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 98 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 99 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 100 | int32_t ksum = 0; |
| 101 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 102 | const uint8_t kv = k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 103 | ksum += (int32_t) kv; |
| 104 | *((uint8_t*) packed_w) = kv; |
| 105 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 106 | } |
| 107 | packed_b[nr_block_offset] -= ksum * (int32_t) izp; |
| 108 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 109 | } |
| 110 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 111 | } |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | void xnn_pack_q8_conv_goki_w( |
| 116 | size_t g, |
| 117 | size_t nc, |
| 118 | size_t ks, |
| 119 | size_t kc, |
| 120 | uint32_t nr, |
| 121 | uint32_t kr, |
| 122 | uint8_t izp, |
| 123 | uint8_t kzp, |
| 124 | const uint8_t* k, |
| 125 | const int32_t* b, |
| 126 | void* packed_w) |
| 127 | { |
| 128 | const int32_t boff = (int32_t) ks * (int32_t) kc * (int32_t) izp * (int32_t) kzp; |
| 129 | do { |
| 130 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 131 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 132 | int32_t* packed_b = (int32_t*) packed_w; |
| 133 | if XNN_LIKELY(b != NULL) { |
| 134 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 135 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 136 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 137 | } |
| 138 | } else { |
| 139 | size_t n = nr_block_size; |
| 140 | do { |
| 141 | *((int32_t*) packed_w) = boff; |
| 142 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 143 | } while (--n != 0); |
| 144 | } |
| 145 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 146 | for (size_t ki = 0; ki < ks; ki++) { |
| 147 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 148 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 149 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 150 | int32_t ksum = 0; |
| 151 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 152 | const uint8_t kv = |
| 153 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 154 | ksum += (int32_t) kv; |
| 155 | *((uint8_t*) packed_w) = kv; |
| 156 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 157 | } |
| 158 | packed_b[nr_block_offset] -= ksum * (int32_t) izp; |
| 159 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 160 | } |
| 161 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 162 | } |
| 163 | } |
| 164 | } |
| 165 | k += ks * kc * nc; |
| 166 | if XNN_UNPREDICTABLE(b != NULL) { |
| 167 | b += nc; |
| 168 | } |
| 169 | } while (--g != 0); |
| 170 | } |
| 171 | |
| 172 | void xnn_pack_q8_conv_kgo_w( |
| 173 | size_t g, |
| 174 | size_t nc, |
| 175 | size_t ks, |
| 176 | uint32_t nr, |
| 177 | uint32_t kr, |
| 178 | uint8_t izp, |
| 179 | uint8_t kzp, |
| 180 | const uint8_t* k, |
| 181 | const int32_t* b, |
| 182 | void* packed_w) |
| 183 | { |
| 184 | const int32_t boff = (int32_t) ks * (int32_t) izp * (int32_t) kzp; |
| 185 | for (size_t i = 0; i < g; i++) { |
| 186 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 187 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 188 | int32_t* packed_b = (int32_t*) packed_w; |
| 189 | if XNN_LIKELY(b != NULL) { |
| 190 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 191 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 192 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 193 | } |
| 194 | } else { |
| 195 | size_t n = nr_block_size; |
| 196 | do { |
| 197 | *((int32_t*) packed_w) = boff; |
| 198 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 199 | } while (--n != 0); |
| 200 | } |
| 201 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 202 | for (size_t ki = 0; ki < ks; ki++) { |
| 203 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 204 | const uint8_t kv = |
| 205 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 206 | *((uint8_t*) packed_w) = kv; |
| 207 | packed_b[nr_block_offset] -= (int32_t) kv * (int32_t) izp; |
| 208 | packed_w = (void*) ((uintptr_t) packed_w + kr * sizeof(uint8_t)); |
| 209 | } |
| 210 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 211 | } |
| 212 | } |
| 213 | k += nc; |
| 214 | if XNN_UNPREDICTABLE(b != NULL) { |
| 215 | b += nc; |
| 216 | } |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | void xnn_pack_q8_deconv_goki_w( |
| 221 | size_t g, |
| 222 | size_t nc, |
| 223 | size_t kh, |
| 224 | size_t kw, |
| 225 | size_t kc, |
| 226 | size_t sh, |
| 227 | size_t sw, |
| 228 | size_t nr, |
| 229 | size_t kr, |
| 230 | uint8_t izp, |
| 231 | uint8_t kzp, |
| 232 | const uint8_t* k, |
| 233 | const int32_t* b, |
| 234 | void* packed_w, |
| 235 | struct subconvolution_params* params) |
| 236 | { |
| 237 | for (size_t i = 0; i < g; i++) { |
| 238 | for (size_t oy = 0; oy < sh; oy++) { |
| 239 | for (size_t ox = 0; ox < sw; ox++) { |
| 240 | if (i == 0) { |
| 241 | (*params++).weights = packed_w; |
| 242 | } |
| 243 | const int32_t boff = (int32_t) divide_round_up(kh - oy, sh) * (int32_t) divide_round_up(kw - ox, sw) * (int32_t) kc * (int32_t) izp * (int32_t) kzp; |
| 244 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 245 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 246 | int32_t* packed_b = (int32_t*) packed_w; |
| 247 | if XNN_LIKELY(b != 0) { |
| 248 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 249 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 250 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 251 | } |
| 252 | } else { |
| 253 | size_t n = nr_block_size; |
| 254 | do { |
| 255 | *((int32_t*) packed_w) = boff; |
| 256 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 257 | } while (--n != 0); |
| 258 | } |
| 259 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 260 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 261 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 262 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 263 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 264 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 265 | int32_t ksum = 0; |
| 266 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 267 | const uint8_t kv = |
| 268 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + (kr_block_start + kr_block_offset)]; |
| 269 | ksum += (int32_t) kv; |
| 270 | *((uint8_t*) packed_w) = kv; |
| 271 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 272 | } |
| 273 | packed_b[nr_block_offset] -= ksum * (int32_t) izp; |
| 274 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 275 | } |
| 276 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 277 | } |
| 278 | } |
| 279 | } |
| 280 | } |
| 281 | } |
| 282 | } |
| 283 | k += kh * kw * kc * nc; |
| 284 | if XNN_UNPREDICTABLE(b != NULL) { |
| 285 | b += nc; |
| 286 | } |
| 287 | } |
| 288 | } |
| 289 | |
| 290 | void xnn_pack_q8_dwconv_ghw_w( |
| 291 | size_t h, |
| 292 | size_t w, |
| 293 | size_t c, |
| 294 | size_t cr, |
| 295 | uint8_t izp, |
| 296 | uint8_t kzp, |
| 297 | const uint8_t* k, |
| 298 | const int32_t* b, |
| 299 | void* packed_w) |
| 300 | { |
| 301 | const int32_t boff = (int32_t) h * (int32_t) w * (int32_t) izp * (int32_t) kzp; |
| 302 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 303 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 304 | int32_t* packed_b = (int32_t*) packed_w; |
| 305 | if XNN_LIKELY(b != NULL) { |
| 306 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 307 | *((int32_t*) packed_w) = b[cr_block_start + cr_block_offset] + boff; |
| 308 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 309 | } |
| 310 | } else { |
| 311 | size_t n = cr_block_size; |
| 312 | do { |
| 313 | *((int32_t*) packed_w) = boff; |
| 314 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 315 | } while (--n != 0); |
| 316 | } |
| 317 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 318 | for (size_t x = 0; x < w; x++) { |
| 319 | for (size_t y = 0; y < h; y++) { |
| 320 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 321 | const uint8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 322 | packed_b[cr_block_offset] -= (int32_t) kv * (int32_t) izp; |
| 323 | *((uint8_t*) packed_w) = kv; |
| 324 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 325 | } |
| 326 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(uint8_t)); |
| 327 | } |
| 328 | } |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | void xnn_pack_q8_dwconv_hwg_w( |
| 333 | size_t h, |
| 334 | size_t w, |
| 335 | size_t c, |
| 336 | size_t cr, |
| 337 | uint8_t izp, |
| 338 | uint8_t kzp, |
| 339 | const uint8_t* k, |
| 340 | const int32_t* b, |
| 341 | void* packed_w) |
| 342 | { |
| 343 | const int32_t boff = (int32_t) h * (int32_t) w * (int32_t) izp * (int32_t) kzp; |
| 344 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 345 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 346 | int32_t* packed_b = (int32_t*) packed_w; |
| 347 | if XNN_LIKELY(b != NULL) { |
| 348 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 349 | *((int32_t*) packed_w) = b[cr_block_start + cr_block_offset] + boff; |
| 350 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 351 | } |
| 352 | } else { |
| 353 | size_t n = cr_block_size; |
| 354 | do { |
| 355 | *((int32_t*) packed_w) = boff; |
| 356 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 357 | } while (--n != 0); |
| 358 | } |
| 359 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 360 | for (size_t x = 0; x < w; x++) { |
| 361 | for (size_t y = 0; y < h; y++) { |
| 362 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 363 | const uint8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 364 | packed_b[cr_block_offset] -= (int32_t) kv * (int32_t) izp; |
| 365 | *((uint8_t*) packed_w) = kv; |
| 366 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 367 | } |
| 368 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(uint8_t)); |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | } |
| 373 | |
| 374 | void xnn_pack_f16_gemm_goi_w( |
| 375 | size_t g, |
| 376 | size_t nc, |
| 377 | size_t kc, |
| 378 | size_t nr, |
| 379 | size_t kr, |
| 380 | size_t sr, |
| 381 | const uint16_t* k, |
| 382 | const uint16_t* b, |
| 383 | uint16_t* packed_w) |
| 384 | { |
| 385 | const size_t skr = sr * kr; |
| 386 | const size_t skc = round_down_po2(kc, skr); |
| 387 | const size_t sr_mask = (sr - 1) * kr; |
| 388 | do { |
| 389 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 390 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 391 | if XNN_LIKELY(b != NULL) { |
| 392 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 393 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 394 | } |
| 395 | } |
| 396 | packed_w += nr; |
| 397 | |
| 398 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 399 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 400 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 401 | *packed_w++ = |
| 402 | k[(nr_block_start + nr_block_offset) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 403 | } |
| 404 | } |
| 405 | packed_w += (nr - nr_block_size) * kr; |
| 406 | } |
| 407 | |
| 408 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 409 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 410 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 411 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 412 | *packed_w++ = |
| 413 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 414 | } |
| 415 | packed_w += kr - kr_block_size; |
| 416 | } |
| 417 | packed_w += (nr - nr_block_size) * kr; |
| 418 | } |
| 419 | } |
| 420 | k += nc * kc; |
| 421 | if XNN_UNPREDICTABLE(b != NULL) { |
| 422 | b += nc; |
| 423 | } |
| 424 | } while (--g != 0); |
| 425 | } |
| 426 | |
| 427 | void xnn_pack_f16_gemm_io_w( |
| 428 | size_t nc, |
| 429 | size_t kc, |
| 430 | size_t nr, |
| 431 | size_t kr, |
| 432 | size_t sr, |
| 433 | const uint16_t* k, |
| 434 | const uint16_t* b, |
| 435 | uint16_t* packed_w) |
| 436 | { |
| 437 | const size_t skr = sr * kr; |
| 438 | const size_t skc = round_down_po2(kc, skr); |
| 439 | const size_t sr_mask = (sr - 1) * kr; |
| 440 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 441 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 442 | if XNN_LIKELY(b != NULL) { |
| 443 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 444 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 445 | } |
| 446 | } |
| 447 | packed_w += nr; |
| 448 | |
| 449 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 450 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 451 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 452 | *packed_w++ = |
| 453 | k[(round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 454 | } |
| 455 | } |
| 456 | packed_w += (nr - nr_block_size) * kr; |
| 457 | } |
| 458 | |
| 459 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 460 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 461 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 462 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 463 | *packed_w++ = |
| 464 | k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 465 | } |
| 466 | packed_w += kr - kr_block_size; |
| 467 | } |
| 468 | packed_w += (nr - nr_block_size) * kr; |
| 469 | } |
| 470 | } |
| 471 | } |
| 472 | |
| 473 | void xnn_pack_f16_gemminc_goi_w( |
| 474 | size_t g, |
| 475 | size_t nc, |
| 476 | size_t kc, |
| 477 | size_t nr, |
| 478 | size_t kr, |
| 479 | size_t sr, |
| 480 | const uint16_t* k, |
| 481 | uint16_t* packed_w) |
| 482 | { |
| 483 | const size_t skr = sr * kr; |
| 484 | const size_t skc = round_down_po2(kc, skr); |
| 485 | const size_t sr_mask = (sr - 1) * kr; |
| 486 | do { |
| 487 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 488 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 489 | |
| 490 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 491 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 492 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 493 | *packed_w++ = |
| 494 | k[(nr_block_start + nr_block_offset) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 495 | } |
| 496 | } |
| 497 | packed_w += (nr - nr_block_size) * kr; |
| 498 | } |
| 499 | |
| 500 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 501 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 502 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 503 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 504 | *packed_w++ = |
| 505 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 506 | } |
| 507 | packed_w += kr - kr_block_size; |
| 508 | } |
| 509 | packed_w += (nr - nr_block_size) * kr; |
| 510 | } |
| 511 | } |
| 512 | k += nc * kc; |
| 513 | } while (--g != 0); |
| 514 | } |
| 515 | |
| 516 | void xnn_pack_f16_conv_goki_w( |
| 517 | size_t g, |
| 518 | size_t nc, |
| 519 | size_t ks, |
| 520 | size_t kc, |
| 521 | size_t nr, |
| 522 | size_t kr, |
| 523 | size_t sr, |
| 524 | const uint16_t* k, |
| 525 | const uint16_t* b, |
| 526 | uint16_t* packed_w) |
| 527 | { |
| 528 | const size_t skr = sr * kr; |
| 529 | const size_t skc = round_down_po2(kc, skr); |
| 530 | const size_t sr_mask = (sr - 1) * kr; |
| 531 | do { |
| 532 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 533 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 534 | if XNN_LIKELY(b != NULL) { |
| 535 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 536 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 537 | } |
| 538 | } |
| 539 | packed_w += nr; |
| 540 | |
| 541 | for (size_t ki = 0; ki < ks; ki++) { |
| 542 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 543 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 544 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 545 | *packed_w++ = |
| 546 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 547 | } |
| 548 | } |
| 549 | packed_w += (nr - nr_block_size) * kr; |
| 550 | } |
| 551 | |
| 552 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 553 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 554 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 555 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 556 | *packed_w++ = |
| 557 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 558 | } |
| 559 | packed_w += kr - kr_block_size; |
| 560 | } |
| 561 | packed_w += (nr - nr_block_size) * kr; |
| 562 | } |
| 563 | } |
| 564 | } |
| 565 | k += ks * kc * nc; |
| 566 | if XNN_UNPREDICTABLE(b != NULL) { |
| 567 | b += nc; |
| 568 | } |
| 569 | } while (--g != 0); |
| 570 | } |
| 571 | |
| 572 | void xnn_pack_f16_conv_kgo_w( |
| 573 | size_t g, |
| 574 | size_t nc, |
| 575 | size_t ks, |
| 576 | size_t nr, |
| 577 | size_t kr, |
| 578 | const uint16_t* k, |
| 579 | const uint16_t* b, |
| 580 | uint16_t* packed_w) |
| 581 | { |
| 582 | for (size_t i = 0; i < g; i++) { |
| 583 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 584 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 585 | if XNN_LIKELY(b != NULL) { |
| 586 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 587 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 588 | } |
| 589 | } |
| 590 | packed_w += nr; |
| 591 | for (size_t ki = 0; ki < ks; ki++) { |
| 592 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 593 | *packed_w = |
| 594 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 595 | packed_w += kr; |
| 596 | } |
| 597 | packed_w += (nr - nr_block_size) * kr; |
| 598 | } |
| 599 | } |
| 600 | k += nc; |
| 601 | if XNN_UNPREDICTABLE(b != NULL) { |
| 602 | b += nc; |
| 603 | } |
| 604 | } |
| 605 | } |
| 606 | |
| 607 | void xnn_pack_f16_dconv_oki_w( |
| 608 | size_t nc, |
| 609 | size_t kc, |
| 610 | size_t nr, |
| 611 | size_t kh, |
| 612 | size_t kw, |
| 613 | const uint16_t* k, |
| 614 | const uint16_t* b, |
| 615 | uint16_t* packed_w) |
| 616 | { |
| 617 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 618 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 619 | if XNN_LIKELY(b != NULL) { |
| 620 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 621 | *packed_w++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 622 | } |
| 623 | } else { |
| 624 | size_t n = nr; |
| 625 | do { |
| 626 | *packed_w++ = 0; |
| 627 | } while (--n != 0); |
| 628 | } |
| 629 | |
| 630 | for (size_t kx = 0; kx < kw; kx++) { |
| 631 | for (size_t c = 0; c < kc; c++) { |
| 632 | for (size_t ky = 0; ky < kh; ky++) { |
| 633 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 634 | *packed_w++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 635 | } |
| 636 | } |
| 637 | } |
| 638 | } |
| 639 | if XNN_UNPREDICTABLE(b != NULL) { |
| 640 | b += nr; |
| 641 | } |
| 642 | } |
| 643 | } |
| 644 | |
| 645 | void xnn_pack_f16_deconv_goki_w( |
| 646 | size_t g, |
| 647 | size_t nc, |
| 648 | size_t kh, |
| 649 | size_t kw, |
| 650 | size_t kc, |
| 651 | size_t sh, |
| 652 | size_t sw, |
| 653 | size_t nr, |
| 654 | size_t kr, |
| 655 | size_t sr, |
| 656 | const uint16_t* k, |
| 657 | const uint16_t* b, |
| 658 | uint16_t* packed_w, |
| 659 | struct subconvolution_params* params) |
| 660 | { |
| 661 | const size_t skr = sr * kr; |
| 662 | const size_t skc = round_down_po2(kc, skr); |
| 663 | const size_t sr_mask = (sr - 1) * kr; |
| 664 | for (size_t i = 0; i < g; i++) { |
| 665 | for (size_t oy = 0; oy < sh; oy++) { |
| 666 | for (size_t ox = 0; ox < sw; ox++) { |
| 667 | if (i == 0) { |
| 668 | (*params++).weights = packed_w; |
| 669 | } |
| 670 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 671 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 672 | if XNN_LIKELY(b != NULL) { |
| 673 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 674 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 675 | } |
| 676 | } |
| 677 | packed_w += nr; |
| 678 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 679 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 680 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 681 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 682 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 683 | *packed_w++ = |
| 684 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 685 | } |
| 686 | } |
| 687 | packed_w += (nr - nr_block_size) * kr; |
| 688 | } |
| 689 | |
| 690 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 691 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 692 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 693 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 694 | *packed_w++ = |
| 695 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + (kr_block_start + kr_block_offset)]; |
| 696 | } |
| 697 | packed_w += kr - kr_block_size; |
| 698 | } |
| 699 | packed_w += (nr - nr_block_size) * kr; |
| 700 | } |
| 701 | } |
| 702 | } |
| 703 | } |
| 704 | } |
| 705 | } |
| 706 | k += kh * kw * kc * nc; |
| 707 | if XNN_UNPREDICTABLE(b != NULL) { |
| 708 | b += nc; |
| 709 | } |
| 710 | } |
| 711 | } |
| 712 | |
| 713 | void xnn_pack_f16_dwconv_ghw_w( |
| 714 | size_t h, |
| 715 | size_t w, |
| 716 | size_t c, |
| 717 | size_t cr, |
| 718 | const uint16_t* k, |
| 719 | const uint16_t* b, |
| 720 | uint16_t* packed_w) |
| 721 | { |
| 722 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 723 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 724 | if XNN_LIKELY(b != NULL) { |
| 725 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 726 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 727 | } |
| 728 | } else { |
| 729 | size_t n = cr_block_size; |
| 730 | do { |
| 731 | *packed_w++ = 0; |
| 732 | } while (--n != 0); |
| 733 | } |
| 734 | packed_w += cr - cr_block_size; |
| 735 | for (size_t x = 0; x < w; x++) { |
| 736 | for (size_t y = 0; y < h; y++) { |
| 737 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 738 | const uint16_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 739 | *packed_w++ = kv; |
| 740 | } |
| 741 | packed_w += cr - cr_block_size; |
| 742 | } |
| 743 | } |
| 744 | } |
| 745 | } |
| 746 | |
| 747 | void xnn_pack_f16_dwconv_hwg_w( |
| 748 | size_t h, |
| 749 | size_t w, |
| 750 | size_t c, |
| 751 | size_t cr, |
| 752 | const uint16_t* k, |
| 753 | const uint16_t* b, |
| 754 | uint16_t* packed_w) |
| 755 | { |
| 756 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 757 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 758 | if XNN_LIKELY(b != NULL) { |
| 759 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 760 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 761 | } |
| 762 | } else { |
| 763 | size_t n = cr_block_size; |
| 764 | do { |
| 765 | *packed_w++ = 0; |
| 766 | } while (--n != 0); |
| 767 | } |
| 768 | packed_w += cr - cr_block_size; |
| 769 | for (size_t x = 0; x < w; x++) { |
| 770 | for (size_t y = 0; y < h; y++) { |
| 771 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 772 | const uint16_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 773 | *packed_w++ = kv; |
| 774 | } |
| 775 | packed_w += cr - cr_block_size; |
| 776 | } |
| 777 | } |
| 778 | } |
| 779 | } |
| 780 | |
| 781 | void xnn_pack_f16_chw_dwconv_ghw_w( |
| 782 | size_t kernel_size, |
| 783 | size_t groups, |
| 784 | const uint16_t* kernel, |
| 785 | const uint16_t* bias, |
| 786 | uint16_t* packed_weights) |
| 787 | { |
| 788 | for (size_t g = 0; g < groups; g++) { |
| 789 | if XNN_LIKELY(bias != NULL) { |
| 790 | *packed_weights = *bias++; |
| 791 | } else { |
| 792 | *packed_weights = 0; |
| 793 | } |
| 794 | packed_weights += 1; |
| 795 | for (size_t i = 0; i < kernel_size; i++) { |
| 796 | *packed_weights++ = kernel[g * kernel_size + i]; |
| 797 | } |
| 798 | } |
| 799 | } |
| 800 | |
| 801 | void xnn_pack_f32_gemm_goi_w( |
| 802 | size_t g, |
| 803 | size_t nc, |
| 804 | size_t kc, |
| 805 | size_t nr, |
| 806 | size_t kr, |
| 807 | size_t sr, |
| 808 | const float* k, |
| 809 | const float* b, |
| 810 | float* packed_w) |
| 811 | { |
| 812 | const size_t skr = sr * kr; |
| 813 | const size_t skc = round_down_po2(kc, skr); |
| 814 | const size_t sr_mask = (sr - 1) * kr; |
| 815 | do { |
| 816 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 817 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 818 | if XNN_LIKELY(b != NULL) { |
| 819 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 820 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 821 | } |
| 822 | } |
| 823 | packed_w += nr; |
| 824 | |
| 825 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 826 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 827 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 828 | *packed_w++ = |
| 829 | k[(nr_block_start + nr_block_offset) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 830 | } |
| 831 | } |
| 832 | packed_w += (nr - nr_block_size) * kr; |
| 833 | } |
| 834 | |
| 835 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 836 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 837 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 838 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 839 | *packed_w++ = |
| 840 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 841 | } |
| 842 | packed_w += kr - kr_block_size; |
| 843 | } |
| 844 | packed_w += (nr - nr_block_size) * kr; |
| 845 | } |
| 846 | } |
| 847 | k += nc * kc; |
| 848 | if XNN_UNPREDICTABLE(b != NULL) { |
| 849 | b += nc; |
| 850 | } |
| 851 | } while (--g != 0); |
| 852 | } |
| 853 | |
| 854 | void xnn_pack_f32_gemm_io_w( |
| 855 | size_t nc, |
| 856 | size_t kc, |
| 857 | size_t nr, |
| 858 | size_t kr, |
| 859 | size_t sr, |
| 860 | const float* k, |
| 861 | const float* b, |
| 862 | float* packed_w) |
| 863 | { |
| 864 | const size_t skr = sr * kr; |
| 865 | const size_t skc = round_down_po2(kc, skr); |
| 866 | const size_t sr_mask = (sr - 1) * kr; |
| 867 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 868 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 869 | if XNN_LIKELY(b != NULL) { |
| 870 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 871 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 872 | } |
| 873 | } |
| 874 | packed_w += nr; |
| 875 | |
| 876 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 877 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 878 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 879 | *packed_w++ = |
| 880 | k[(round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 881 | } |
| 882 | } |
| 883 | packed_w += (nr - nr_block_size) * kr; |
| 884 | } |
| 885 | |
| 886 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 887 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 888 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 889 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 890 | *packed_w++ = |
| 891 | k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 892 | } |
| 893 | packed_w += kr - kr_block_size; |
| 894 | } |
| 895 | packed_w += (nr - nr_block_size) * kr; |
| 896 | } |
| 897 | } |
| 898 | } |
| 899 | |
| 900 | void xnn_pack_f32_gemminc_goi_w( |
| 901 | size_t g, |
| 902 | size_t nc, |
| 903 | size_t kc, |
| 904 | size_t nr, |
| 905 | size_t kr, |
| 906 | size_t sr, |
| 907 | const float* k, |
| 908 | float* packed_w) |
| 909 | { |
| 910 | const size_t skr = sr * kr; |
| 911 | const size_t skc = round_down_po2(kc, skr); |
| 912 | const size_t sr_mask = (sr - 1) * kr; |
| 913 | do { |
| 914 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 915 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 916 | |
| 917 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 918 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 919 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 920 | *packed_w++ = |
| 921 | k[(nr_block_start + nr_block_offset) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 922 | } |
| 923 | } |
| 924 | packed_w += (nr - nr_block_size) * kr; |
| 925 | } |
| 926 | |
| 927 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 928 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 929 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 930 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 931 | *packed_w++ = |
| 932 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 933 | } |
| 934 | packed_w += kr - kr_block_size; |
| 935 | } |
| 936 | packed_w += (nr - nr_block_size) * kr; |
| 937 | } |
| 938 | } |
| 939 | k += nc * kc; |
| 940 | } while (--g != 0); |
| 941 | } |
| 942 | |
| 943 | void xnn_pack_f32_conv_goki_w( |
| 944 | size_t g, |
| 945 | size_t nc, |
| 946 | size_t ks, |
| 947 | size_t kc, |
| 948 | size_t nr, |
| 949 | size_t kr, |
| 950 | size_t sr, |
| 951 | const float* k, |
| 952 | const float* b, |
| 953 | float* packed_w) |
| 954 | { |
| 955 | const size_t skr = sr * kr; |
| 956 | const size_t skc = round_down_po2(kc, skr); |
| 957 | const size_t sr_mask = (sr - 1) * kr; |
| 958 | do { |
| 959 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 960 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 961 | if XNN_LIKELY(b != NULL) { |
| 962 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 963 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 964 | } |
| 965 | } |
| 966 | packed_w += nr; |
| 967 | |
| 968 | for (size_t ki = 0; ki < ks; ki++) { |
| 969 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 970 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 971 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 972 | *packed_w++ = |
| 973 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 974 | } |
| 975 | } |
| 976 | packed_w += (nr - nr_block_size) * kr; |
| 977 | } |
| 978 | |
| 979 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 980 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 981 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 982 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 983 | *packed_w++ = |
| 984 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 985 | } |
| 986 | packed_w += kr - kr_block_size; |
| 987 | } |
| 988 | packed_w += (nr - nr_block_size) * kr; |
| 989 | } |
| 990 | } |
| 991 | } |
| 992 | k += ks * kc * nc; |
| 993 | if XNN_UNPREDICTABLE(b != NULL) { |
| 994 | b += nc; |
| 995 | } |
| 996 | } while (--g != 0); |
| 997 | } |
| 998 | |
| 999 | void xnn_pack_f32_conv_kgo_w( |
| 1000 | size_t g, |
| 1001 | size_t nc, |
| 1002 | size_t ks, |
| 1003 | size_t nr, |
| 1004 | size_t kr, |
| 1005 | const float* k, |
| 1006 | const float* b, |
| 1007 | float* packed_w) |
| 1008 | { |
| 1009 | for (size_t i = 0; i < g; i++) { |
| 1010 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1011 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1012 | if XNN_LIKELY(b != NULL) { |
| 1013 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1014 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 1015 | } |
| 1016 | } |
| 1017 | packed_w += nr; |
| 1018 | for (size_t ki = 0; ki < ks; ki++) { |
| 1019 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1020 | *packed_w = |
| 1021 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 1022 | packed_w += kr; |
| 1023 | } |
| 1024 | packed_w += (nr - nr_block_size) * kr; |
| 1025 | } |
| 1026 | } |
| 1027 | k += nc; |
| 1028 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1029 | b += nc; |
| 1030 | } |
| 1031 | } |
| 1032 | } |
| 1033 | |
| 1034 | void xnn_pack_f32_dconv_oki_w( |
| 1035 | size_t nc, |
| 1036 | size_t kc, |
| 1037 | size_t nr, |
| 1038 | size_t kh, |
| 1039 | size_t kw, |
| 1040 | const float* k, |
| 1041 | const float* b, |
| 1042 | float* packed_w) |
| 1043 | { |
| 1044 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1045 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1046 | if XNN_LIKELY(b != NULL) { |
| 1047 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1048 | *packed_w++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 1049 | } |
| 1050 | } else { |
| 1051 | size_t n = nr; |
| 1052 | do { |
| 1053 | *packed_w++ = 0.0f; |
| 1054 | } while (--n != 0); |
| 1055 | } |
| 1056 | |
| 1057 | for (size_t kx = 0; kx < kw; kx++) { |
| 1058 | for (size_t c = 0; c < kc; c++) { |
| 1059 | for (size_t ky = 0; ky < kh; ky++) { |
| 1060 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1061 | *packed_w++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 1062 | } |
| 1063 | } |
| 1064 | } |
| 1065 | } |
| 1066 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1067 | b += nr; |
| 1068 | } |
| 1069 | } |
| 1070 | } |
| 1071 | |
| 1072 | void xnn_pack_f32_deconv_goki_w( |
| 1073 | size_t g, |
| 1074 | size_t nc, |
| 1075 | size_t kh, |
| 1076 | size_t kw, |
| 1077 | size_t kc, |
| 1078 | size_t sh, |
| 1079 | size_t sw, |
| 1080 | size_t nr, |
| 1081 | size_t kr, |
| 1082 | size_t sr, |
| 1083 | const float* k, |
| 1084 | const float* b, |
| 1085 | float* packed_w, |
| 1086 | struct subconvolution_params* params) |
| 1087 | { |
| 1088 | const size_t skr = sr * kr; |
| 1089 | const size_t skc = round_down_po2(kc, skr); |
| 1090 | const size_t sr_mask = (sr - 1) * kr; |
| 1091 | for (size_t i = 0; i < g; i++) { |
| 1092 | for (size_t oy = 0; oy < sh; oy++) { |
| 1093 | for (size_t ox = 0; ox < sw; ox++) { |
| 1094 | if (i == 0) { |
| 1095 | (*params++).weights = packed_w; |
| 1096 | } |
| 1097 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1098 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1099 | if XNN_LIKELY(b != NULL) { |
| 1100 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1101 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 1102 | } |
| 1103 | } |
| 1104 | packed_w += nr; |
| 1105 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1106 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1107 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 1108 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1109 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1110 | *packed_w++ = |
| 1111 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + round_down_po2(kr_block_start, skr) + ((kr_block_start + nr_block_offset * kr) & sr_mask) + kr_block_offset]; |
| 1112 | } |
| 1113 | } |
| 1114 | packed_w += (nr - nr_block_size) * kr; |
| 1115 | } |
| 1116 | |
| 1117 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 1118 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 1119 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1120 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 1121 | *packed_w++ = |
| 1122 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + (kr_block_start + kr_block_offset)]; |
| 1123 | } |
| 1124 | packed_w += kr - kr_block_size; |
| 1125 | } |
| 1126 | packed_w += (nr - nr_block_size) * kr; |
| 1127 | } |
| 1128 | } |
| 1129 | } |
| 1130 | } |
| 1131 | } |
| 1132 | } |
| 1133 | k += kh * kw * kc * nc; |
| 1134 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1135 | b += nc; |
| 1136 | } |
| 1137 | } |
| 1138 | } |
| 1139 | |
| 1140 | void xnn_pack_f32_dwconv_ghw_w( |
| 1141 | size_t h, |
| 1142 | size_t w, |
| 1143 | size_t c, |
| 1144 | size_t cr, |
| 1145 | const float* k, |
| 1146 | const float* b, |
| 1147 | float* packed_w) |
| 1148 | { |
| 1149 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1150 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1151 | if XNN_LIKELY(b != NULL) { |
| 1152 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1153 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1154 | } |
| 1155 | } else { |
| 1156 | size_t n = cr_block_size; |
| 1157 | do { |
| 1158 | *packed_w++ = 0.0f; |
| 1159 | } while (--n != 0); |
| 1160 | } |
| 1161 | packed_w += cr - cr_block_size; |
| 1162 | for (size_t x = 0; x < w; x++) { |
| 1163 | for (size_t y = 0; y < h; y++) { |
| 1164 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1165 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1166 | *packed_w++ = kv; |
| 1167 | } |
| 1168 | packed_w += cr - cr_block_size; |
| 1169 | } |
| 1170 | } |
| 1171 | } |
| 1172 | } |
| 1173 | |
| 1174 | void xnn_pack_f32_dwconv_hwg_w( |
| 1175 | size_t h, |
| 1176 | size_t w, |
| 1177 | size_t c, |
| 1178 | size_t cr, |
| 1179 | const float* k, |
| 1180 | const float* b, |
| 1181 | float* packed_w) |
| 1182 | { |
| 1183 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1184 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1185 | if XNN_LIKELY(b != NULL) { |
| 1186 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1187 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1188 | } |
| 1189 | } else { |
| 1190 | size_t n = cr_block_size; |
| 1191 | do { |
| 1192 | *packed_w++ = 0.0f; |
| 1193 | } while (--n != 0); |
| 1194 | } |
| 1195 | packed_w += cr - cr_block_size; |
| 1196 | for (size_t x = 0; x < w; x++) { |
| 1197 | for (size_t y = 0; y < h; y++) { |
| 1198 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1199 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1200 | *packed_w++ = kv; |
| 1201 | } |
| 1202 | packed_w += cr - cr_block_size; |
| 1203 | } |
| 1204 | } |
| 1205 | } |
| 1206 | } |
| 1207 | |
| 1208 | void xnn_pack_f32_chw_dwconv_ghw_w( |
| 1209 | size_t kernel_size, |
| 1210 | size_t groups, |
| 1211 | const float* kernel, |
| 1212 | const float* bias, |
| 1213 | float* packed_weights) |
| 1214 | { |
| 1215 | for (size_t g = 0; g < groups; g++) { |
| 1216 | if XNN_LIKELY(bias != NULL) { |
| 1217 | *packed_weights = *bias++; |
| 1218 | } else { |
| 1219 | *packed_weights = 0.0f; |
| 1220 | } |
| 1221 | packed_weights += 1; |
| 1222 | for (size_t i = 0; i < kernel_size; i++) { |
| 1223 | *packed_weights++ = kernel[g * kernel_size + i]; |
| 1224 | } |
| 1225 | } |
| 1226 | } |
| 1227 | |
| 1228 | void xnn_pack_f32_chw_dwconv_hwg_w( |
| 1229 | size_t kernel_size, |
| 1230 | size_t groups, |
| 1231 | const float* kernel, |
| 1232 | const float* bias, |
| 1233 | float* packed_weights) |
| 1234 | { |
| 1235 | for (size_t g = 0; g < groups; g++) { |
| 1236 | if XNN_LIKELY(bias != NULL) { |
| 1237 | *packed_weights = *bias++; |
| 1238 | } else { |
| 1239 | *packed_weights = 0.0f; |
| 1240 | } |
| 1241 | packed_weights += 1; |
| 1242 | for (size_t i = 0; i < kernel_size; i++) { |
| 1243 | *packed_weights++ = kernel[i * groups + g]; |
| 1244 | } |
| 1245 | } |
| 1246 | } |
| 1247 | |
| 1248 | void xnn_pack_f32_vmulcaddc_w( |
| 1249 | size_t c, |
| 1250 | size_t cr, |
| 1251 | const float* s, |
| 1252 | const float* b, |
| 1253 | float* packed_w) |
| 1254 | { |
| 1255 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1256 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1257 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1258 | *packed_w++ = s[cr_block_start + cr_block_offset]; |
| 1259 | } |
| 1260 | packed_w += cr - cr_block_size; |
| 1261 | if XNN_LIKELY(b != NULL) { |
| 1262 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1263 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1264 | } |
| 1265 | } else { |
| 1266 | size_t n = cr_block_size; |
| 1267 | do { |
| 1268 | *packed_w++ = 0.0f; |
| 1269 | } while (--n != 0); |
| 1270 | } |
| 1271 | packed_w += cr - cr_block_size; |
| 1272 | } |
| 1273 | } |
| 1274 | |
| 1275 | void xnn_pack_f16_vmulcaddc_w( |
| 1276 | size_t c, |
| 1277 | size_t cr, |
| 1278 | const uint16_t* s, |
| 1279 | const uint16_t* b, |
| 1280 | uint16_t* packed_w) |
| 1281 | { |
| 1282 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1283 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1284 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1285 | *packed_w++ = s[cr_block_start + cr_block_offset]; |
| 1286 | } |
| 1287 | packed_w += cr - cr_block_size; |
| 1288 | if XNN_LIKELY(b != NULL) { |
| 1289 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1290 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1291 | } |
| 1292 | } else { |
| 1293 | size_t n = cr_block_size; |
| 1294 | do { |
| 1295 | *packed_w++ = 0; |
| 1296 | } while (--n != 0); |
| 1297 | } |
| 1298 | packed_w += cr - cr_block_size; |
| 1299 | } |
| 1300 | } |