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 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 16 | void xnn_pack_f32_gemm_goi_w( |
| 17 | size_t g, |
| 18 | size_t nc, |
| 19 | size_t kc, |
| 20 | size_t nr, |
| 21 | size_t kr, |
| 22 | size_t sr, |
| 23 | const float* k, |
| 24 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 25 | float* packed_w, |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 26 | size_t extra_bytes, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 27 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 28 | { |
| 29 | const size_t skr = sr * kr; |
| 30 | const size_t skc = round_down_po2(kc, skr); |
| 31 | const size_t sr_mask = (sr - 1) * kr; |
| 32 | do { |
| 33 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 34 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 35 | if XNN_LIKELY(b != NULL) { |
| 36 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 37 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 38 | } |
| 39 | } |
| 40 | packed_w += nr; |
| 41 | |
| 42 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 43 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 44 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 45 | *packed_w++ = |
| 46 | 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]; |
| 47 | } |
| 48 | } |
| 49 | packed_w += (nr - nr_block_size) * kr; |
| 50 | } |
| 51 | |
| 52 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 53 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 54 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 55 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 56 | *packed_w++ = |
| 57 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 58 | } |
| 59 | packed_w += kr - kr_block_size; |
| 60 | } |
| 61 | packed_w += (nr - nr_block_size) * kr; |
| 62 | } |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 63 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 64 | } |
| 65 | k += nc * kc; |
| 66 | if XNN_UNPREDICTABLE(b != NULL) { |
| 67 | b += nc; |
| 68 | } |
| 69 | } while (--g != 0); |
| 70 | } |
| 71 | |
| 72 | void xnn_pack_f16_gemm_goi_w( |
| 73 | size_t g, |
| 74 | size_t nc, |
| 75 | size_t kc, |
| 76 | size_t nr, |
| 77 | size_t kr, |
| 78 | size_t sr, |
| 79 | const uint16_t* k, |
| 80 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 81 | uint16_t* packed_w, |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 82 | size_t extra_bytes, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 83 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 84 | { |
| 85 | const size_t skr = sr * kr; |
| 86 | const size_t skc = round_down_po2(kc, skr); |
| 87 | const size_t sr_mask = (sr - 1) * kr; |
| 88 | do { |
| 89 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 90 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 91 | if XNN_LIKELY(b != NULL) { |
| 92 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 93 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 94 | } |
| 95 | } |
| 96 | packed_w += nr; |
| 97 | |
| 98 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 99 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 100 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 101 | *packed_w++ = |
| 102 | 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]; |
| 103 | } |
| 104 | } |
| 105 | packed_w += (nr - nr_block_size) * kr; |
| 106 | } |
| 107 | |
| 108 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 109 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 110 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 111 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 112 | *packed_w++ = |
| 113 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 114 | } |
| 115 | packed_w += kr - kr_block_size; |
| 116 | } |
| 117 | packed_w += (nr - nr_block_size) * kr; |
| 118 | } |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 119 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 120 | } |
| 121 | k += nc * kc; |
| 122 | if XNN_UNPREDICTABLE(b != NULL) { |
| 123 | b += nc; |
| 124 | } |
| 125 | } while (--g != 0); |
| 126 | } |
| 127 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 128 | void xnn_pack_qu8_gemm_goi_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 129 | size_t g, |
| 130 | size_t nc, |
| 131 | size_t kc, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 132 | size_t nr, |
| 133 | size_t kr, |
Marat Dukhan | bc08f31 | 2020-07-07 16:22:04 -0700 | [diff] [blame] | 134 | size_t sr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 135 | const uint8_t* k, |
| 136 | const int32_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 137 | void* packed_w, |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 138 | size_t extra_bytes, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 139 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 140 | { |
Marat Dukhan | bc08f31 | 2020-07-07 16:22:04 -0700 | [diff] [blame] | 141 | assert(sr == 1); |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 142 | const int32_t izp = (int32_t) params->input_zero_point; |
| 143 | const int32_t boff = (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 144 | do { |
| 145 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 146 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 147 | int32_t* packed_b = (int32_t*) packed_w; |
| 148 | if XNN_LIKELY(b != NULL) { |
| 149 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 150 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 151 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 152 | } |
| 153 | } else { |
| 154 | size_t n = nr_block_size; |
| 155 | do { |
| 156 | *((int32_t*) packed_w) = boff; |
| 157 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 158 | } while (--n != 0); |
| 159 | } |
| 160 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 161 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 162 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 163 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 164 | int32_t ksum = 0; |
| 165 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 166 | const uint8_t kv = k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 167 | ksum += (int32_t) kv; |
| 168 | *((uint8_t*) packed_w) = kv; |
| 169 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 170 | } |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 171 | packed_b[nr_block_offset] -= ksum * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 172 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 173 | } |
| 174 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 175 | } |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 176 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 177 | } |
| 178 | k += nc * kc; |
| 179 | if XNN_UNPREDICTABLE(b != NULL) { |
| 180 | b += nc; |
| 181 | } |
| 182 | } while (--g != 0); |
| 183 | } |
| 184 | |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 185 | void xnn_pack_qs8_gemm_goi_w( |
| 186 | size_t g, |
| 187 | size_t nc, |
| 188 | size_t kc, |
| 189 | size_t nr, |
| 190 | size_t kr, |
| 191 | size_t sr, |
| 192 | const int8_t* k, |
| 193 | const int32_t* b, |
| 194 | void* packed_w, |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 195 | size_t extra_bytes, |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 196 | const struct xnn_qs8_packing_params* params) |
| 197 | { |
| 198 | assert(sr == 1); |
| 199 | const int32_t izp = (int32_t) params->input_zero_point; |
| 200 | do { |
| 201 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 202 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 203 | int32_t* packed_b = (int32_t*) packed_w; |
| 204 | if XNN_LIKELY(b != NULL) { |
| 205 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 206 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset]; |
| 207 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 208 | } |
| 209 | } else { |
| 210 | size_t n = nr_block_size; |
| 211 | do { |
| 212 | *((int32_t*) packed_w) = 0; |
| 213 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 214 | } while (--n != 0); |
| 215 | } |
| 216 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 217 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 218 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 219 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 220 | int32_t ksum = 0; |
| 221 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 222 | const int8_t kv = k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 223 | ksum += (int32_t) kv; |
| 224 | *((int8_t*) packed_w) = kv; |
| 225 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 226 | } |
| 227 | packed_b[nr_block_offset] -= ksum * izp; |
| 228 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(int8_t)); |
| 229 | } |
| 230 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(int8_t)); |
| 231 | } |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 232 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | 595e170 | 2020-07-31 10:12:52 -0700 | [diff] [blame] | 233 | } |
| 234 | k += nc * kc; |
| 235 | if XNN_UNPREDICTABLE(b != NULL) { |
| 236 | b += nc; |
| 237 | } |
| 238 | } while (--g != 0); |
| 239 | } |
| 240 | |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 241 | void xnn_pack_qs8_gemm_xw_goi_w( |
| 242 | size_t g, |
| 243 | size_t nc, |
| 244 | size_t kc, |
| 245 | size_t nr, |
| 246 | size_t kr, |
| 247 | size_t sr, |
| 248 | const int8_t* k, |
| 249 | const int32_t* b, |
| 250 | void* packed_w, |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 251 | size_t extra_bytes, |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 252 | const struct xnn_qs8_packing_params* params) |
| 253 | { |
| 254 | assert(sr == 1); |
| 255 | const int32_t izp = (int32_t) params->input_zero_point; |
| 256 | do { |
| 257 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 258 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 259 | int32_t* packed_b = (int32_t*) packed_w; |
| 260 | if XNN_LIKELY(b != NULL) { |
| 261 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 262 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset]; |
| 263 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 264 | } |
| 265 | } else { |
| 266 | size_t n = nr_block_size; |
| 267 | do { |
| 268 | *((int32_t*) packed_w) = 0; |
| 269 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 270 | } while (--n != 0); |
| 271 | } |
| 272 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 273 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 274 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 275 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 276 | int32_t ksum = 0; |
| 277 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 278 | const int8_t kv = k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 279 | ksum += (int32_t) kv; |
| 280 | *((int16_t*) packed_w) = (int16_t) kv; |
| 281 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int16_t)); |
| 282 | } |
| 283 | packed_b[nr_block_offset] -= ksum * izp; |
| 284 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(int16_t)); |
| 285 | } |
| 286 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(int16_t)); |
| 287 | } |
Marat Dukhan | 0b04374 | 2021-06-02 18:29:11 -0700 | [diff] [blame] | 288 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | 683fab3 | 2020-08-03 19:42:52 -0700 | [diff] [blame] | 289 | } |
| 290 | k += nc * kc; |
| 291 | if XNN_UNPREDICTABLE(b != NULL) { |
| 292 | b += nc; |
| 293 | } |
| 294 | } while (--g != 0); |
| 295 | } |
| 296 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 297 | void xnn_pack_f32_gemm_io_w( |
| 298 | size_t nc, |
| 299 | size_t kc, |
| 300 | size_t nr, |
| 301 | size_t kr, |
| 302 | size_t sr, |
| 303 | const float* k, |
| 304 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 305 | float* packed_w, |
| 306 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 307 | { |
| 308 | const size_t skr = sr * kr; |
| 309 | const size_t skc = round_down_po2(kc, skr); |
| 310 | const size_t sr_mask = (sr - 1) * kr; |
| 311 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 312 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 313 | if XNN_LIKELY(b != NULL) { |
| 314 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 315 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 316 | } |
| 317 | } |
| 318 | packed_w += nr; |
| 319 | |
| 320 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 321 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 322 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 323 | *packed_w++ = |
| 324 | 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)]; |
| 325 | } |
| 326 | } |
| 327 | packed_w += (nr - nr_block_size) * kr; |
| 328 | } |
| 329 | |
| 330 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 331 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 332 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 333 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 334 | *packed_w++ = |
| 335 | k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 336 | } |
| 337 | packed_w += kr - kr_block_size; |
| 338 | } |
| 339 | packed_w += (nr - nr_block_size) * kr; |
| 340 | } |
| 341 | } |
| 342 | } |
| 343 | |
| 344 | void xnn_pack_f16_gemm_io_w( |
| 345 | size_t nc, |
| 346 | size_t kc, |
| 347 | size_t nr, |
| 348 | size_t kr, |
| 349 | size_t sr, |
| 350 | const uint16_t* k, |
| 351 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 352 | uint16_t* packed_w, |
| 353 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 354 | { |
| 355 | const size_t skr = sr * kr; |
| 356 | const size_t skc = round_down_po2(kc, skr); |
| 357 | const size_t sr_mask = (sr - 1) * kr; |
| 358 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 359 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 360 | if XNN_LIKELY(b != NULL) { |
| 361 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 362 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 363 | } |
| 364 | } |
| 365 | packed_w += nr; |
| 366 | |
| 367 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 368 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 369 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 370 | *packed_w++ = |
| 371 | 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)]; |
| 372 | } |
| 373 | } |
| 374 | packed_w += (nr - nr_block_size) * kr; |
| 375 | } |
| 376 | |
| 377 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 378 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 379 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 380 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 381 | *packed_w++ = |
| 382 | k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 383 | } |
| 384 | packed_w += kr - kr_block_size; |
| 385 | } |
| 386 | packed_w += (nr - nr_block_size) * kr; |
| 387 | } |
| 388 | } |
| 389 | } |
| 390 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 391 | void xnn_pack_qu8_gemm_io_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 392 | size_t nc, |
| 393 | size_t kc, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 394 | size_t nr, |
| 395 | size_t kr, |
Marat Dukhan | bc08f31 | 2020-07-07 16:22:04 -0700 | [diff] [blame] | 396 | size_t sr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 397 | const uint8_t* k, |
| 398 | const int32_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 399 | void* packed_w, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 400 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 401 | { |
Marat Dukhan | bc08f31 | 2020-07-07 16:22:04 -0700 | [diff] [blame] | 402 | assert(sr == 1); |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 403 | const int32_t izp = (int32_t) params->input_zero_point; |
| 404 | const int32_t boff = (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 405 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 406 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 407 | int32_t* packed_b = (int32_t*) packed_w; |
| 408 | if XNN_LIKELY(b != NULL) { |
| 409 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 410 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 411 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 412 | } |
| 413 | } else { |
| 414 | size_t n = nr_block_size; |
| 415 | do { |
| 416 | *((int32_t*) packed_w) = boff; |
| 417 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 418 | } while (--n != 0); |
| 419 | } |
| 420 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 421 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 422 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 423 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 424 | int32_t ksum = 0; |
| 425 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 426 | const uint8_t kv = k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 427 | ksum += (int32_t) kv; |
| 428 | *((uint8_t*) packed_w) = kv; |
| 429 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 430 | } |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 431 | packed_b[nr_block_offset] -= ksum * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 432 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 433 | } |
| 434 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 435 | } |
| 436 | } |
| 437 | } |
| 438 | |
Marat Dukhan | d23cb6e | 2021-04-01 01:18:58 -0700 | [diff] [blame] | 439 | void xnn_pack_qs8_gemm_io_w( |
| 440 | size_t nc, |
| 441 | size_t kc, |
| 442 | size_t nr, |
| 443 | size_t kr, |
| 444 | size_t sr, |
| 445 | const int8_t* k, |
| 446 | const int32_t* b, |
| 447 | void* packed_w, |
| 448 | const struct xnn_qs8_packing_params* params) |
| 449 | { |
| 450 | assert(sr == 1); |
| 451 | const int32_t izp = (int32_t) params->input_zero_point; |
| 452 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 453 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 454 | int32_t* packed_b = (int32_t*) packed_w; |
| 455 | if XNN_LIKELY(b != NULL) { |
| 456 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 457 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset]; |
| 458 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 459 | } |
| 460 | } else { |
| 461 | size_t n = nr_block_size; |
| 462 | do { |
| 463 | *((int32_t*) packed_w) = 0; |
| 464 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 465 | } while (--n != 0); |
| 466 | } |
| 467 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 468 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 469 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 470 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 471 | int32_t ksum = 0; |
| 472 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 473 | const int8_t kv = k[(kr_block_start + kr_block_offset) * nc + (nr_block_start + nr_block_offset)]; |
| 474 | ksum += (int32_t) kv; |
| 475 | *((int8_t*) packed_w) = kv; |
| 476 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 477 | } |
| 478 | packed_b[nr_block_offset] -= ksum * izp; |
| 479 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(int8_t)); |
| 480 | } |
| 481 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(int8_t)); |
| 482 | } |
| 483 | } |
| 484 | } |
| 485 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 486 | void xnn_pack_f32_conv_goki_w( |
| 487 | size_t g, |
| 488 | size_t nc, |
| 489 | size_t ks, |
| 490 | size_t kc, |
| 491 | size_t nr, |
| 492 | size_t kr, |
| 493 | size_t sr, |
| 494 | const float* k, |
| 495 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 496 | float* packed_w, |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 497 | size_t extra_bytes, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 498 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 499 | { |
| 500 | const size_t skr = sr * kr; |
| 501 | const size_t skc = round_down_po2(kc, skr); |
| 502 | const size_t sr_mask = (sr - 1) * kr; |
| 503 | do { |
| 504 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 505 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 506 | if XNN_LIKELY(b != NULL) { |
| 507 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 508 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 509 | } |
| 510 | } |
| 511 | packed_w += nr; |
| 512 | |
| 513 | for (size_t ki = 0; ki < ks; ki++) { |
| 514 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 515 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 516 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 517 | *packed_w++ = |
| 518 | 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]; |
| 519 | } |
| 520 | } |
| 521 | packed_w += (nr - nr_block_size) * kr; |
| 522 | } |
| 523 | |
| 524 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 525 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 526 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 527 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 528 | *packed_w++ = |
| 529 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 530 | } |
| 531 | packed_w += kr - kr_block_size; |
| 532 | } |
| 533 | packed_w += (nr - nr_block_size) * kr; |
| 534 | } |
| 535 | } |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 536 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 537 | } |
| 538 | k += ks * kc * nc; |
| 539 | if XNN_UNPREDICTABLE(b != NULL) { |
| 540 | b += nc; |
| 541 | } |
| 542 | } while (--g != 0); |
| 543 | } |
| 544 | |
| 545 | void xnn_pack_f16_conv_goki_w( |
| 546 | size_t g, |
| 547 | size_t nc, |
| 548 | size_t ks, |
| 549 | size_t kc, |
| 550 | size_t nr, |
| 551 | size_t kr, |
| 552 | size_t sr, |
| 553 | const uint16_t* k, |
| 554 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 555 | uint16_t* packed_w, |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 556 | size_t extra_bytes, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 557 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 558 | { |
| 559 | const size_t skr = sr * kr; |
| 560 | const size_t skc = round_down_po2(kc, skr); |
| 561 | const size_t sr_mask = (sr - 1) * kr; |
| 562 | do { |
| 563 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 564 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 565 | if XNN_LIKELY(b != NULL) { |
| 566 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 567 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 568 | } |
| 569 | } |
| 570 | packed_w += nr; |
| 571 | |
| 572 | for (size_t ki = 0; ki < ks; ki++) { |
| 573 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 574 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 575 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 576 | *packed_w++ = |
| 577 | 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]; |
| 578 | } |
| 579 | } |
| 580 | packed_w += (nr - nr_block_size) * kr; |
| 581 | } |
| 582 | |
| 583 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 584 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 585 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 586 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 587 | *packed_w++ = |
| 588 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 589 | } |
| 590 | packed_w += kr - kr_block_size; |
| 591 | } |
| 592 | packed_w += (nr - nr_block_size) * kr; |
| 593 | } |
| 594 | } |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 595 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 596 | } |
| 597 | k += ks * kc * nc; |
| 598 | if XNN_UNPREDICTABLE(b != NULL) { |
| 599 | b += nc; |
| 600 | } |
| 601 | } while (--g != 0); |
| 602 | } |
| 603 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 604 | void xnn_pack_qu8_conv_goki_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 605 | size_t g, |
| 606 | size_t nc, |
| 607 | size_t ks, |
| 608 | size_t kc, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 609 | size_t nr, |
| 610 | size_t kr, |
Marat Dukhan | 5a698bb | 2020-07-07 20:47:55 -0700 | [diff] [blame] | 611 | size_t sr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 612 | const uint8_t* k, |
| 613 | const int32_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 614 | void* packed_w, |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 615 | size_t extra_bytes, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 616 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 617 | { |
Marat Dukhan | 5a698bb | 2020-07-07 20:47:55 -0700 | [diff] [blame] | 618 | assert(sr == 1); |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 619 | const int32_t izp = (int32_t) params->input_zero_point; |
| 620 | const int32_t boff = (int32_t) ks * (int32_t) kc * izp * (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 621 | do { |
| 622 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 623 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 624 | int32_t* packed_b = (int32_t*) packed_w; |
| 625 | if XNN_LIKELY(b != NULL) { |
| 626 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 627 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 628 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 629 | } |
| 630 | } else { |
| 631 | size_t n = nr_block_size; |
| 632 | do { |
| 633 | *((int32_t*) packed_w) = boff; |
| 634 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 635 | } while (--n != 0); |
| 636 | } |
| 637 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 638 | for (size_t ki = 0; ki < ks; ki++) { |
| 639 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 640 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 641 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 642 | int32_t ksum = 0; |
| 643 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 644 | const uint8_t kv = |
| 645 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 646 | ksum += (int32_t) kv; |
| 647 | *((uint8_t*) packed_w) = kv; |
| 648 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 649 | } |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 650 | packed_b[nr_block_offset] -= ksum * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 651 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 652 | } |
| 653 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 654 | } |
| 655 | } |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 656 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 657 | } |
| 658 | k += ks * kc * nc; |
| 659 | if XNN_UNPREDICTABLE(b != NULL) { |
| 660 | b += nc; |
| 661 | } |
| 662 | } while (--g != 0); |
| 663 | } |
| 664 | |
Marat Dukhan | f948068 | 2020-07-31 14:50:24 -0700 | [diff] [blame] | 665 | void xnn_pack_qs8_conv_goki_w( |
| 666 | size_t g, |
| 667 | size_t nc, |
| 668 | size_t ks, |
| 669 | size_t kc, |
| 670 | size_t nr, |
| 671 | size_t kr, |
| 672 | size_t sr, |
| 673 | const int8_t* k, |
| 674 | const int32_t* b, |
| 675 | void* packed_w, |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 676 | size_t extra_bytes, |
Marat Dukhan | f948068 | 2020-07-31 14:50:24 -0700 | [diff] [blame] | 677 | const struct xnn_qs8_packing_params* params) |
| 678 | { |
| 679 | assert(sr == 1); |
| 680 | const int32_t izp = (int32_t) params->input_zero_point; |
| 681 | do { |
| 682 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 683 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 684 | int32_t* packed_b = (int32_t*) packed_w; |
| 685 | if XNN_LIKELY(b != NULL) { |
| 686 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 687 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset]; |
| 688 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 689 | } |
| 690 | } else { |
| 691 | size_t n = nr_block_size; |
| 692 | do { |
| 693 | *((int32_t*) packed_w) = 0; |
| 694 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 695 | } while (--n != 0); |
| 696 | } |
| 697 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 698 | for (size_t ki = 0; ki < ks; ki++) { |
| 699 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 700 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 701 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 702 | int32_t ksum = 0; |
| 703 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 704 | const int8_t kv = |
| 705 | k[((nr_block_start + nr_block_offset) * ks + ki) * kc + (kr_block_start + kr_block_offset)]; |
| 706 | ksum += (int32_t) kv; |
| 707 | *((int8_t*) packed_w) = kv; |
| 708 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 709 | } |
| 710 | packed_b[nr_block_offset] -= ksum * izp; |
| 711 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(int8_t)); |
| 712 | } |
| 713 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(int8_t)); |
| 714 | } |
| 715 | } |
Marat Dukhan | e06c813 | 2021-06-03 08:59:11 -0700 | [diff] [blame] | 716 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | f948068 | 2020-07-31 14:50:24 -0700 | [diff] [blame] | 717 | } |
| 718 | k += ks * kc * nc; |
| 719 | if XNN_UNPREDICTABLE(b != NULL) { |
| 720 | b += nc; |
| 721 | } |
| 722 | } while (--g != 0); |
| 723 | } |
| 724 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 725 | void xnn_pack_f32_conv_kgo_w( |
| 726 | size_t g, |
| 727 | size_t nc, |
| 728 | size_t ks, |
| 729 | size_t nr, |
| 730 | size_t kr, |
| 731 | const float* k, |
| 732 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 733 | float* packed_w, |
| 734 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 735 | { |
| 736 | for (size_t i = 0; i < g; i++) { |
| 737 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 738 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 739 | if XNN_LIKELY(b != NULL) { |
| 740 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 741 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 742 | } |
| 743 | } |
| 744 | packed_w += nr; |
| 745 | for (size_t ki = 0; ki < ks; ki++) { |
| 746 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 747 | *packed_w = |
| 748 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 749 | packed_w += kr; |
| 750 | } |
| 751 | packed_w += (nr - nr_block_size) * kr; |
| 752 | } |
| 753 | } |
| 754 | k += nc; |
| 755 | if XNN_UNPREDICTABLE(b != NULL) { |
| 756 | b += nc; |
| 757 | } |
| 758 | } |
| 759 | } |
| 760 | |
| 761 | void xnn_pack_f16_conv_kgo_w( |
| 762 | size_t g, |
| 763 | size_t nc, |
| 764 | size_t ks, |
| 765 | size_t nr, |
| 766 | size_t kr, |
| 767 | const uint16_t* k, |
| 768 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 769 | uint16_t* packed_w, |
| 770 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 771 | { |
| 772 | for (size_t i = 0; i < g; i++) { |
| 773 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 774 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 775 | if XNN_LIKELY(b != NULL) { |
| 776 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 777 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 778 | } |
| 779 | } |
| 780 | packed_w += nr; |
| 781 | for (size_t ki = 0; ki < ks; ki++) { |
| 782 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 783 | *packed_w = |
| 784 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 785 | packed_w += kr; |
| 786 | } |
| 787 | packed_w += (nr - nr_block_size) * kr; |
| 788 | } |
| 789 | } |
| 790 | k += nc; |
| 791 | if XNN_UNPREDICTABLE(b != NULL) { |
| 792 | b += nc; |
| 793 | } |
| 794 | } |
| 795 | } |
| 796 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 797 | void xnn_pack_qu8_conv_kgo_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 798 | size_t g, |
| 799 | size_t nc, |
| 800 | size_t ks, |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 801 | size_t nr, |
| 802 | size_t kr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 803 | const uint8_t* k, |
| 804 | const int32_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 805 | void* packed_w, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 806 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 807 | { |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 808 | const int32_t izp = (int32_t) params->input_zero_point; |
| 809 | const int32_t boff = (int32_t) ks * izp * (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 810 | for (size_t i = 0; i < g; i++) { |
| 811 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 812 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 813 | int32_t* packed_b = (int32_t*) packed_w; |
| 814 | if XNN_LIKELY(b != NULL) { |
| 815 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 816 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 817 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 818 | } |
| 819 | } else { |
| 820 | size_t n = nr_block_size; |
| 821 | do { |
| 822 | *((int32_t*) packed_w) = boff; |
| 823 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 824 | } while (--n != 0); |
| 825 | } |
| 826 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 827 | for (size_t ki = 0; ki < ks; ki++) { |
| 828 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 829 | const uint8_t kv = |
| 830 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 831 | *((uint8_t*) packed_w) = kv; |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 832 | packed_b[nr_block_offset] -= (int32_t) kv * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 833 | packed_w = (void*) ((uintptr_t) packed_w + kr * sizeof(uint8_t)); |
| 834 | } |
| 835 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 836 | } |
| 837 | } |
| 838 | k += nc; |
| 839 | if XNN_UNPREDICTABLE(b != NULL) { |
| 840 | b += nc; |
| 841 | } |
| 842 | } |
| 843 | } |
| 844 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 845 | void xnn_pack_qs8_conv_kgo_w( |
| 846 | size_t g, |
| 847 | size_t nc, |
| 848 | size_t ks, |
| 849 | size_t nr, |
| 850 | size_t kr, |
| 851 | const int8_t* k, |
| 852 | const int32_t* b, |
| 853 | void* packed_w, |
| 854 | const struct xnn_qs8_packing_params* params) |
| 855 | { |
| 856 | const int32_t izp = (int32_t) params->input_zero_point; |
| 857 | for (size_t i = 0; i < g; i++) { |
| 858 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 859 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 860 | int32_t* packed_b = (int32_t*) packed_w; |
| 861 | if XNN_LIKELY(b != NULL) { |
| 862 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 863 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset]; |
| 864 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 865 | } |
| 866 | } else { |
| 867 | size_t n = nr_block_size; |
| 868 | do { |
| 869 | *((int32_t*) packed_w) = 0; |
| 870 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 871 | } while (--n != 0); |
| 872 | } |
| 873 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 874 | for (size_t ki = 0; ki < ks; ki++) { |
| 875 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 876 | const int8_t kv = |
| 877 | k[ki * g * nc + (nr_block_start + nr_block_offset)]; |
| 878 | *((int8_t*) packed_w) = kv; |
| 879 | packed_b[nr_block_offset] -= (int32_t) kv * izp; |
| 880 | packed_w = (void*) ((uintptr_t) packed_w + kr * sizeof(int8_t)); |
| 881 | } |
| 882 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(int8_t)); |
| 883 | } |
| 884 | } |
| 885 | k += nc; |
| 886 | if XNN_UNPREDICTABLE(b != NULL) { |
| 887 | b += nc; |
| 888 | } |
| 889 | } |
| 890 | } |
| 891 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 892 | void xnn_pack_f32_deconv_goki_w( |
| 893 | size_t g, |
| 894 | size_t nc, |
| 895 | size_t kh, |
| 896 | size_t kw, |
| 897 | size_t kc, |
| 898 | size_t sh, |
| 899 | size_t sw, |
| 900 | size_t nr, |
| 901 | size_t kr, |
| 902 | size_t sr, |
| 903 | const float* k, |
| 904 | const float* b, |
| 905 | float* packed_w, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 906 | struct subconvolution_params* subconv_params, |
| 907 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 908 | { |
| 909 | const size_t skr = sr * kr; |
| 910 | const size_t skc = round_down_po2(kc, skr); |
| 911 | const size_t sr_mask = (sr - 1) * kr; |
| 912 | for (size_t i = 0; i < g; i++) { |
| 913 | for (size_t oy = 0; oy < sh; oy++) { |
| 914 | for (size_t ox = 0; ox < sw; ox++) { |
| 915 | if (i == 0) { |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 916 | (*subconv_params++).weights = packed_w; |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 917 | } |
| 918 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 919 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 920 | if XNN_LIKELY(b != NULL) { |
| 921 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 922 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 923 | } |
| 924 | } |
| 925 | packed_w += nr; |
| 926 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 927 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 928 | for (size_t kr_block_start = 0; kr_block_start < skc; 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; kr_block_offset++) { |
| 931 | *packed_w++ = |
| 932 | 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]; |
| 933 | } |
| 934 | } |
| 935 | packed_w += (nr - nr_block_size) * kr; |
| 936 | } |
| 937 | |
| 938 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 939 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 940 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 941 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 942 | *packed_w++ = |
| 943 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + (kr_block_start + kr_block_offset)]; |
| 944 | } |
| 945 | packed_w += kr - kr_block_size; |
| 946 | } |
| 947 | packed_w += (nr - nr_block_size) * kr; |
| 948 | } |
| 949 | } |
| 950 | } |
| 951 | } |
| 952 | } |
| 953 | } |
| 954 | k += kh * kw * kc * nc; |
| 955 | if XNN_UNPREDICTABLE(b != NULL) { |
| 956 | b += nc; |
| 957 | } |
| 958 | } |
| 959 | } |
| 960 | |
| 961 | void xnn_pack_f16_deconv_goki_w( |
| 962 | size_t g, |
| 963 | size_t nc, |
| 964 | size_t kh, |
| 965 | size_t kw, |
| 966 | size_t kc, |
| 967 | size_t sh, |
| 968 | size_t sw, |
| 969 | size_t nr, |
| 970 | size_t kr, |
| 971 | size_t sr, |
| 972 | const uint16_t* k, |
| 973 | const uint16_t* b, |
| 974 | uint16_t* packed_w, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 975 | struct subconvolution_params* subconv_params, |
| 976 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 977 | { |
| 978 | const size_t skr = sr * kr; |
| 979 | const size_t skc = round_down_po2(kc, skr); |
| 980 | const size_t sr_mask = (sr - 1) * kr; |
| 981 | for (size_t i = 0; i < g; i++) { |
| 982 | for (size_t oy = 0; oy < sh; oy++) { |
| 983 | for (size_t ox = 0; ox < sw; ox++) { |
| 984 | if (i == 0) { |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 985 | (*subconv_params++).weights = packed_w; |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 986 | } |
| 987 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 988 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 989 | if XNN_LIKELY(b != NULL) { |
| 990 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 991 | packed_w[nr_block_offset] = b[nr_block_start + nr_block_offset]; |
| 992 | } |
| 993 | } |
| 994 | packed_w += nr; |
| 995 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 996 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 997 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 998 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 999 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1000 | *packed_w++ = |
| 1001 | 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]; |
| 1002 | } |
| 1003 | } |
| 1004 | packed_w += (nr - nr_block_size) * kr; |
| 1005 | } |
| 1006 | |
| 1007 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 1008 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 1009 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1010 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 1011 | *packed_w++ = |
| 1012 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + (kr_block_start + kr_block_offset)]; |
| 1013 | } |
| 1014 | packed_w += kr - kr_block_size; |
| 1015 | } |
| 1016 | packed_w += (nr - nr_block_size) * kr; |
| 1017 | } |
| 1018 | } |
| 1019 | } |
| 1020 | } |
| 1021 | } |
| 1022 | } |
| 1023 | k += kh * kw * kc * nc; |
| 1024 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1025 | b += nc; |
| 1026 | } |
| 1027 | } |
| 1028 | } |
| 1029 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1030 | void xnn_pack_qu8_deconv_goki_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1031 | size_t g, |
| 1032 | size_t nc, |
| 1033 | size_t kh, |
| 1034 | size_t kw, |
| 1035 | size_t kc, |
| 1036 | size_t sh, |
| 1037 | size_t sw, |
| 1038 | size_t nr, |
| 1039 | size_t kr, |
Marat Dukhan | 5a698bb | 2020-07-07 20:47:55 -0700 | [diff] [blame] | 1040 | size_t sr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1041 | const uint8_t* k, |
| 1042 | const int32_t* b, |
| 1043 | void* packed_w, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1044 | struct subconvolution_params* subconv_params, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1045 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1046 | { |
Marat Dukhan | 5a698bb | 2020-07-07 20:47:55 -0700 | [diff] [blame] | 1047 | assert(sr == 1); |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1048 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1049 | const int32_t kzp = (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1050 | for (size_t i = 0; i < g; i++) { |
| 1051 | for (size_t oy = 0; oy < sh; oy++) { |
| 1052 | for (size_t ox = 0; ox < sw; ox++) { |
| 1053 | if (i == 0) { |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1054 | (*subconv_params++).weights = packed_w; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1055 | } |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1056 | const int32_t boff = (int32_t) divide_round_up(kh - oy, sh) * (int32_t) divide_round_up(kw - ox, sw) * (int32_t) kc * izp * kzp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1057 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1058 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1059 | int32_t* packed_b = (int32_t*) packed_w; |
| 1060 | if XNN_LIKELY(b != 0) { |
| 1061 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1062 | *((int32_t*) packed_w) = b[nr_block_start + nr_block_offset] + boff; |
| 1063 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1064 | } |
| 1065 | } else { |
| 1066 | size_t n = nr_block_size; |
| 1067 | do { |
| 1068 | *((int32_t*) packed_w) = boff; |
| 1069 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1070 | } while (--n != 0); |
| 1071 | } |
| 1072 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * sizeof(int32_t)); |
| 1073 | for (size_t ky = oy; ky < kh; ky += sh) { |
| 1074 | for (size_t kx = ox; kx < kw; kx += sw) { |
| 1075 | for (size_t kr_block_start = 0; kr_block_start < kc; kr_block_start += kr) { |
| 1076 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 1077 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1078 | int32_t ksum = 0; |
| 1079 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 1080 | const uint8_t kv = |
| 1081 | k[(((nr_block_start + nr_block_offset) * kh + ky) * kw + kx) * kc + (kr_block_start + kr_block_offset)]; |
| 1082 | ksum += (int32_t) kv; |
| 1083 | *((uint8_t*) packed_w) = kv; |
| 1084 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 1085 | } |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1086 | packed_b[nr_block_offset] -= ksum * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1087 | packed_w = (void*) ((uintptr_t) packed_w + (kr - kr_block_size) * sizeof(uint8_t)); |
| 1088 | } |
| 1089 | packed_w = (void*) ((uintptr_t) packed_w + (nr - nr_block_size) * kr * sizeof(uint8_t)); |
| 1090 | } |
| 1091 | } |
| 1092 | } |
| 1093 | } |
| 1094 | } |
| 1095 | } |
| 1096 | k += kh * kw * kc * nc; |
| 1097 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1098 | b += nc; |
| 1099 | } |
| 1100 | } |
| 1101 | } |
| 1102 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1103 | void xnn_pack_f32_dwconv_ghw_w( |
| 1104 | size_t h, |
| 1105 | size_t w, |
| 1106 | size_t c, |
| 1107 | size_t cr, |
| 1108 | const float* k, |
| 1109 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1110 | float* packed_w, |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1111 | size_t extra_bytes, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1112 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1113 | { |
| 1114 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1115 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1116 | if XNN_LIKELY(b != NULL) { |
| 1117 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1118 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1119 | } |
| 1120 | } else { |
| 1121 | size_t n = cr_block_size; |
| 1122 | do { |
| 1123 | *packed_w++ = 0.0f; |
| 1124 | } while (--n != 0); |
| 1125 | } |
| 1126 | packed_w += cr - cr_block_size; |
| 1127 | for (size_t x = 0; x < w; x++) { |
| 1128 | for (size_t y = 0; y < h; y++) { |
| 1129 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1130 | const float kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1131 | *packed_w++ = kv; |
| 1132 | } |
| 1133 | packed_w += cr - cr_block_size; |
| 1134 | } |
| 1135 | } |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1136 | packed_w = (float*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1137 | } |
| 1138 | } |
| 1139 | |
| 1140 | void xnn_pack_f16_dwconv_ghw_w( |
| 1141 | size_t h, |
| 1142 | size_t w, |
| 1143 | size_t c, |
| 1144 | size_t cr, |
| 1145 | const uint16_t* k, |
| 1146 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1147 | uint16_t* packed_w, |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1148 | size_t extra_bytes, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1149 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1150 | { |
| 1151 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1152 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1153 | if XNN_LIKELY(b != NULL) { |
| 1154 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1155 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1156 | } |
| 1157 | } else { |
| 1158 | size_t n = cr_block_size; |
| 1159 | do { |
| 1160 | *packed_w++ = 0; |
| 1161 | } while (--n != 0); |
| 1162 | } |
| 1163 | packed_w += cr - cr_block_size; |
| 1164 | for (size_t x = 0; x < w; x++) { |
| 1165 | for (size_t y = 0; y < h; y++) { |
| 1166 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1167 | const uint16_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1168 | *packed_w++ = kv; |
| 1169 | } |
| 1170 | packed_w += cr - cr_block_size; |
| 1171 | } |
| 1172 | } |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1173 | packed_w = (uint16_t*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1174 | } |
| 1175 | } |
| 1176 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1177 | void xnn_pack_qu8_dwconv_ghw_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1178 | size_t h, |
| 1179 | size_t w, |
| 1180 | size_t c, |
| 1181 | size_t cr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1182 | const uint8_t* k, |
| 1183 | const int32_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1184 | void* packed_w, |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1185 | size_t extra_bytes, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1186 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1187 | { |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1188 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1189 | const int32_t boff = (int32_t) h * (int32_t) w * izp * (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1190 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1191 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1192 | int32_t* packed_b = (int32_t*) packed_w; |
| 1193 | if XNN_LIKELY(b != NULL) { |
| 1194 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1195 | *((int32_t*) packed_w) = b[cr_block_start + cr_block_offset] + boff; |
| 1196 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1197 | } |
| 1198 | } else { |
| 1199 | size_t n = cr_block_size; |
| 1200 | do { |
| 1201 | *((int32_t*) packed_w) = boff; |
| 1202 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1203 | } while (--n != 0); |
| 1204 | } |
| 1205 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1206 | for (size_t x = 0; x < w; x++) { |
| 1207 | for (size_t y = 0; y < h; y++) { |
| 1208 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1209 | const uint8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1210 | packed_b[cr_block_offset] -= (int32_t) kv * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1211 | *((uint8_t*) packed_w) = kv; |
| 1212 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 1213 | } |
| 1214 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(uint8_t)); |
| 1215 | } |
| 1216 | } |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1217 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1218 | } |
| 1219 | } |
| 1220 | |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 1221 | void xnn_pack_qs8_dwconv_ghw_w( |
| 1222 | size_t h, |
| 1223 | size_t w, |
| 1224 | size_t c, |
| 1225 | size_t cr, |
| 1226 | const int8_t* k, |
| 1227 | const int32_t* b, |
| 1228 | void* packed_w, |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1229 | size_t extra_bytes, |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 1230 | const struct xnn_qs8_packing_params* params) |
| 1231 | { |
| 1232 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1233 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1234 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1235 | int32_t* packed_b = (int32_t*) packed_w; |
| 1236 | if XNN_LIKELY(b != NULL) { |
| 1237 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1238 | *((int32_t*) packed_w) = b[cr_block_start + cr_block_offset]; |
| 1239 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1240 | } |
| 1241 | } else { |
| 1242 | size_t n = cr_block_size; |
| 1243 | do { |
| 1244 | *((int32_t*) packed_w) = 0; |
| 1245 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1246 | } while (--n != 0); |
| 1247 | } |
| 1248 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1249 | for (size_t x = 0; x < w; x++) { |
| 1250 | for (size_t y = 0; y < h; y++) { |
| 1251 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1252 | const int8_t kv = k[((cr_block_start + cr_block_offset) * h + y) * w + x]; |
| 1253 | packed_b[cr_block_offset] -= (int32_t) kv * izp; |
| 1254 | *((int8_t*) packed_w) = kv; |
| 1255 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 1256 | } |
| 1257 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int8_t)); |
| 1258 | } |
| 1259 | } |
Marat Dukhan | 8228689 | 2021-06-04 17:27:27 -0700 | [diff] [blame] | 1260 | packed_w = (void*) ((uintptr_t) packed_w + extra_bytes); |
Marat Dukhan | f62bbdc | 2020-08-04 13:59:04 -0700 | [diff] [blame] | 1261 | } |
| 1262 | } |
| 1263 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1264 | void xnn_pack_f32_dwconv_hwg_w( |
| 1265 | size_t h, |
| 1266 | size_t w, |
| 1267 | size_t c, |
| 1268 | size_t cr, |
| 1269 | const float* k, |
| 1270 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1271 | float* packed_w, |
| 1272 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1273 | { |
| 1274 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1275 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1276 | if XNN_LIKELY(b != NULL) { |
| 1277 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1278 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1279 | } |
| 1280 | } else { |
| 1281 | size_t n = cr_block_size; |
| 1282 | do { |
| 1283 | *packed_w++ = 0.0f; |
| 1284 | } while (--n != 0); |
| 1285 | } |
| 1286 | packed_w += cr - cr_block_size; |
| 1287 | for (size_t x = 0; x < w; x++) { |
| 1288 | for (size_t y = 0; y < h; y++) { |
| 1289 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1290 | const float kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1291 | *packed_w++ = kv; |
| 1292 | } |
| 1293 | packed_w += cr - cr_block_size; |
| 1294 | } |
| 1295 | } |
| 1296 | } |
| 1297 | } |
| 1298 | |
| 1299 | void xnn_pack_f16_dwconv_hwg_w( |
| 1300 | size_t h, |
| 1301 | size_t w, |
| 1302 | size_t c, |
| 1303 | size_t cr, |
| 1304 | const uint16_t* k, |
| 1305 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1306 | uint16_t* packed_w, |
| 1307 | const void* params) |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1308 | { |
| 1309 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1310 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1311 | if XNN_LIKELY(b != NULL) { |
| 1312 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1313 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1314 | } |
| 1315 | } else { |
| 1316 | size_t n = cr_block_size; |
| 1317 | do { |
| 1318 | *packed_w++ = 0; |
| 1319 | } while (--n != 0); |
| 1320 | } |
| 1321 | packed_w += cr - cr_block_size; |
| 1322 | for (size_t x = 0; x < w; x++) { |
| 1323 | for (size_t y = 0; y < h; y++) { |
| 1324 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1325 | const uint16_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1326 | *packed_w++ = kv; |
| 1327 | } |
| 1328 | packed_w += cr - cr_block_size; |
| 1329 | } |
| 1330 | } |
| 1331 | } |
| 1332 | } |
| 1333 | |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1334 | void xnn_pack_qu8_dwconv_hwg_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1335 | size_t h, |
| 1336 | size_t w, |
| 1337 | size_t c, |
| 1338 | size_t cr, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1339 | const uint8_t* k, |
| 1340 | const int32_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1341 | void* packed_w, |
Marat Dukhan | 08b7a97 | 2020-07-14 18:17:29 -0700 | [diff] [blame] | 1342 | const struct xnn_qu8_packing_params* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1343 | { |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1344 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1345 | const int32_t boff = (int32_t) h * (int32_t) w * izp * (int32_t) params->kernel_zero_point; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1346 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1347 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1348 | int32_t* packed_b = (int32_t*) packed_w; |
| 1349 | if XNN_LIKELY(b != NULL) { |
| 1350 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1351 | *((int32_t*) packed_w) = b[cr_block_start + cr_block_offset] + boff; |
| 1352 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1353 | } |
| 1354 | } else { |
| 1355 | size_t n = cr_block_size; |
| 1356 | do { |
| 1357 | *((int32_t*) packed_w) = boff; |
| 1358 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1359 | } while (--n != 0); |
| 1360 | } |
| 1361 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1362 | for (size_t x = 0; x < w; x++) { |
| 1363 | for (size_t y = 0; y < h; y++) { |
| 1364 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1365 | const uint8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1366 | packed_b[cr_block_offset] -= (int32_t) kv * izp; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1367 | *((uint8_t*) packed_w) = kv; |
| 1368 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(uint8_t)); |
| 1369 | } |
| 1370 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(uint8_t)); |
| 1371 | } |
| 1372 | } |
| 1373 | } |
| 1374 | } |
| 1375 | |
Marat Dukhan | 16f1e1a | 2020-08-04 16:38:22 -0700 | [diff] [blame] | 1376 | void xnn_pack_qs8_dwconv_hwg_w( |
| 1377 | size_t h, |
| 1378 | size_t w, |
| 1379 | size_t c, |
| 1380 | size_t cr, |
| 1381 | const int8_t* k, |
| 1382 | const int32_t* b, |
| 1383 | void* packed_w, |
| 1384 | const struct xnn_qs8_packing_params* params) |
| 1385 | { |
| 1386 | const int32_t izp = (int32_t) params->input_zero_point; |
| 1387 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1388 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1389 | int32_t* packed_b = (int32_t*) packed_w; |
| 1390 | if XNN_LIKELY(b != NULL) { |
| 1391 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1392 | *((int32_t*) packed_w) = b[cr_block_start + cr_block_offset]; |
| 1393 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1394 | } |
| 1395 | } else { |
| 1396 | size_t n = cr_block_size; |
| 1397 | do { |
| 1398 | *((int32_t*) packed_w) = 0; |
| 1399 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int32_t)); |
| 1400 | } while (--n != 0); |
| 1401 | } |
| 1402 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int32_t)); |
| 1403 | for (size_t x = 0; x < w; x++) { |
| 1404 | for (size_t y = 0; y < h; y++) { |
| 1405 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1406 | const int8_t kv = k[(y * w + x) * c + (cr_block_start + cr_block_offset)]; |
| 1407 | packed_b[cr_block_offset] -= (int32_t) kv * izp; |
| 1408 | *((int8_t*) packed_w) = kv; |
| 1409 | packed_w = (void*) ((uintptr_t) packed_w + sizeof(int8_t)); |
| 1410 | } |
| 1411 | packed_w = (void*) ((uintptr_t) packed_w + (cr - cr_block_size) * sizeof(int8_t)); |
| 1412 | } |
| 1413 | } |
| 1414 | } |
| 1415 | } |
| 1416 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1417 | void xnn_pack_f32_gemminc_goi_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1418 | size_t g, |
| 1419 | size_t nc, |
| 1420 | size_t kc, |
| 1421 | size_t nr, |
| 1422 | size_t kr, |
| 1423 | size_t sr, |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1424 | const float* k, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1425 | float* packed_w, |
| 1426 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1427 | { |
| 1428 | const size_t skr = sr * kr; |
| 1429 | const size_t skc = round_down_po2(kc, skr); |
| 1430 | const size_t sr_mask = (sr - 1) * kr; |
| 1431 | do { |
| 1432 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1433 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1434 | |
| 1435 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 1436 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1437 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1438 | *packed_w++ = |
| 1439 | 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]; |
| 1440 | } |
| 1441 | } |
| 1442 | packed_w += (nr - nr_block_size) * kr; |
| 1443 | } |
| 1444 | |
| 1445 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 1446 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 1447 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1448 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 1449 | *packed_w++ = |
| 1450 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 1451 | } |
| 1452 | packed_w += kr - kr_block_size; |
| 1453 | } |
| 1454 | packed_w += (nr - nr_block_size) * kr; |
| 1455 | } |
| 1456 | } |
| 1457 | k += nc * kc; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1458 | } while (--g != 0); |
| 1459 | } |
| 1460 | |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1461 | void xnn_pack_f16_gemminc_goi_w( |
| 1462 | size_t g, |
| 1463 | size_t nc, |
| 1464 | size_t kc, |
| 1465 | size_t nr, |
| 1466 | size_t kr, |
| 1467 | size_t sr, |
| 1468 | const uint16_t* k, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1469 | uint16_t* packed_w, |
| 1470 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1471 | { |
| 1472 | const size_t skr = sr * kr; |
| 1473 | const size_t skc = round_down_po2(kc, skr); |
| 1474 | const size_t sr_mask = (sr - 1) * kr; |
| 1475 | do { |
| 1476 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1477 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1478 | |
| 1479 | for (size_t kr_block_start = 0; kr_block_start < skc; kr_block_start += kr) { |
| 1480 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1481 | for (size_t kr_block_offset = 0; kr_block_offset < kr; kr_block_offset++) { |
| 1482 | *packed_w++ = |
| 1483 | 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]; |
| 1484 | } |
| 1485 | } |
| 1486 | packed_w += (nr - nr_block_size) * kr; |
| 1487 | } |
| 1488 | |
| 1489 | for (size_t kr_block_start = skc; kr_block_start < kc; kr_block_start += kr) { |
| 1490 | const size_t kr_block_size = min(kc - kr_block_start, kr); |
| 1491 | for (size_t nr_block_offset = 0; nr_block_offset < nr_block_size; nr_block_offset++) { |
| 1492 | for (size_t kr_block_offset = 0; kr_block_offset < kr_block_size; kr_block_offset++) { |
| 1493 | *packed_w++ = |
| 1494 | k[(nr_block_start + nr_block_offset) * kc + (kr_block_start + kr_block_offset)]; |
| 1495 | } |
| 1496 | packed_w += kr - kr_block_size; |
| 1497 | } |
| 1498 | packed_w += (nr - nr_block_size) * kr; |
| 1499 | } |
| 1500 | } |
| 1501 | k += nc * kc; |
| 1502 | } while (--g != 0); |
| 1503 | } |
| 1504 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1505 | void xnn_pack_f32_dconv_oki_w( |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1506 | size_t nc, |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1507 | size_t kc, |
| 1508 | size_t nr, |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1509 | size_t kh, |
| 1510 | size_t kw, |
| 1511 | const float* k, |
| 1512 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1513 | float* packed_w, |
| 1514 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1515 | { |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1516 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1517 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1518 | if XNN_LIKELY(b != NULL) { |
| 1519 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1520 | *packed_w++ = b[min(nr_block_offset, nr_block_size - 1)]; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1521 | } |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1522 | } else { |
| 1523 | size_t n = nr; |
| 1524 | do { |
| 1525 | *packed_w++ = 0.0f; |
| 1526 | } while (--n != 0); |
| 1527 | } |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1528 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1529 | for (size_t kx = 0; kx < kw; kx++) { |
| 1530 | for (size_t c = 0; c < kc; c++) { |
| 1531 | for (size_t ky = 0; ky < kh; ky++) { |
| 1532 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1533 | *packed_w++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1534 | } |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1535 | } |
| 1536 | } |
| 1537 | } |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1538 | if XNN_UNPREDICTABLE(b != NULL) { |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1539 | b += nr; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1540 | } |
| 1541 | } |
| 1542 | } |
| 1543 | |
| 1544 | void xnn_pack_f16_dconv_oki_w( |
| 1545 | size_t nc, |
| 1546 | size_t kc, |
| 1547 | size_t nr, |
| 1548 | size_t kh, |
| 1549 | size_t kw, |
| 1550 | const uint16_t* k, |
| 1551 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1552 | uint16_t* packed_w, |
| 1553 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1554 | { |
| 1555 | for (size_t nr_block_start = 0; nr_block_start < nc; nr_block_start += nr) { |
| 1556 | const size_t nr_block_size = min(nc - nr_block_start, nr); |
| 1557 | if XNN_LIKELY(b != NULL) { |
| 1558 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1559 | *packed_w++ = b[min(nr_block_offset, nr_block_size - 1)]; |
| 1560 | } |
| 1561 | } else { |
| 1562 | size_t n = nr; |
| 1563 | do { |
| 1564 | *packed_w++ = 0; |
| 1565 | } while (--n != 0); |
| 1566 | } |
| 1567 | |
| 1568 | for (size_t kx = 0; kx < kw; kx++) { |
| 1569 | for (size_t c = 0; c < kc; c++) { |
| 1570 | for (size_t ky = 0; ky < kh; ky++) { |
| 1571 | for (size_t nr_block_offset = 0; nr_block_offset < nr; nr_block_offset++) { |
| 1572 | *packed_w++ = k[(((nr_block_start + min(nr_block_offset, nr_block_size - 1)) * kh + ky) * kw + kx) * kc + c]; |
| 1573 | } |
| 1574 | } |
| 1575 | } |
| 1576 | } |
| 1577 | if XNN_UNPREDICTABLE(b != NULL) { |
| 1578 | b += nr; |
| 1579 | } |
| 1580 | } |
| 1581 | } |
| 1582 | |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1583 | void xnn_pack_f32_chw_dwconv_ghw_w( |
| 1584 | size_t kernel_size, |
| 1585 | size_t groups, |
| 1586 | const float* kernel, |
| 1587 | const float* bias, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1588 | float* packed_weights, |
| 1589 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1590 | { |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1591 | for (size_t g = 0; g < groups; g++) { |
| 1592 | if XNN_LIKELY(bias != NULL) { |
| 1593 | *packed_weights = *bias++; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1594 | } else { |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1595 | *packed_weights = 0.0f; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1596 | } |
Marat Dukhan | a6879bd | 2020-07-06 14:25:08 -0700 | [diff] [blame] | 1597 | packed_weights += 1; |
| 1598 | for (size_t i = 0; i < kernel_size; i++) { |
| 1599 | *packed_weights++ = kernel[g * kernel_size + i]; |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1600 | } |
| 1601 | } |
| 1602 | } |
| 1603 | |
| 1604 | void xnn_pack_f16_chw_dwconv_ghw_w( |
| 1605 | size_t kernel_size, |
| 1606 | size_t groups, |
| 1607 | const uint16_t* kernel, |
| 1608 | const uint16_t* bias, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1609 | uint16_t* packed_weights, |
| 1610 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1611 | { |
| 1612 | for (size_t g = 0; g < groups; g++) { |
| 1613 | if XNN_LIKELY(bias != NULL) { |
| 1614 | *packed_weights = *bias++; |
| 1615 | } else { |
| 1616 | *packed_weights = 0; |
| 1617 | } |
| 1618 | packed_weights += 1; |
| 1619 | for (size_t i = 0; i < kernel_size; i++) { |
| 1620 | *packed_weights++ = kernel[g * kernel_size + i]; |
| 1621 | } |
| 1622 | } |
| 1623 | } |
| 1624 | |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1625 | void xnn_pack_f32_chw_dwconv_hwg_w( |
| 1626 | size_t kernel_size, |
| 1627 | size_t groups, |
| 1628 | const float* kernel, |
| 1629 | const float* bias, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1630 | float* packed_weights, |
| 1631 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1632 | { |
| 1633 | for (size_t g = 0; g < groups; g++) { |
| 1634 | if XNN_LIKELY(bias != NULL) { |
| 1635 | *packed_weights = *bias++; |
| 1636 | } else { |
| 1637 | *packed_weights = 0.0f; |
| 1638 | } |
| 1639 | packed_weights += 1; |
| 1640 | for (size_t i = 0; i < kernel_size; i++) { |
| 1641 | *packed_weights++ = kernel[i * groups + g]; |
| 1642 | } |
| 1643 | } |
| 1644 | } |
| 1645 | |
| 1646 | void xnn_pack_f32_vmulcaddc_w( |
| 1647 | size_t c, |
| 1648 | size_t cr, |
| 1649 | const float* s, |
| 1650 | const float* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1651 | float* packed_w, |
| 1652 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1653 | { |
| 1654 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1655 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1656 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1657 | *packed_w++ = s[cr_block_start + cr_block_offset]; |
| 1658 | } |
| 1659 | packed_w += cr - cr_block_size; |
| 1660 | if XNN_LIKELY(b != NULL) { |
| 1661 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1662 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1663 | } |
| 1664 | } else { |
| 1665 | size_t n = cr_block_size; |
| 1666 | do { |
| 1667 | *packed_w++ = 0.0f; |
| 1668 | } while (--n != 0); |
| 1669 | } |
| 1670 | packed_w += cr - cr_block_size; |
| 1671 | } |
| 1672 | } |
| 1673 | |
| 1674 | void xnn_pack_f16_vmulcaddc_w( |
| 1675 | size_t c, |
| 1676 | size_t cr, |
| 1677 | const uint16_t* s, |
| 1678 | const uint16_t* b, |
Marat Dukhan | b42f866 | 2020-07-06 20:46:13 -0700 | [diff] [blame] | 1679 | uint16_t* packed_w, |
| 1680 | const void* params) |
Marat Dukhan | ab58238 | 2020-07-06 13:32:08 -0700 | [diff] [blame] | 1681 | { |
| 1682 | for (size_t cr_block_start = 0; cr_block_start < c; cr_block_start += cr) { |
| 1683 | const size_t cr_block_size = min(c - cr_block_start, cr); |
| 1684 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1685 | *packed_w++ = s[cr_block_start + cr_block_offset]; |
| 1686 | } |
| 1687 | packed_w += cr - cr_block_size; |
| 1688 | if XNN_LIKELY(b != NULL) { |
| 1689 | for (size_t cr_block_offset = 0; cr_block_offset < cr_block_size; cr_block_offset++) { |
| 1690 | *packed_w++ = b[cr_block_start + cr_block_offset]; |
| 1691 | } |
| 1692 | } else { |
| 1693 | size_t n = cr_block_size; |
| 1694 | do { |
| 1695 | *packed_w++ = 0; |
| 1696 | } while (--n != 0); |
| 1697 | } |
| 1698 | packed_w += cr - cr_block_size; |
| 1699 | } |
| 1700 | } |