1R"( 2 3#ifndef ARM_COMPUTE_HELPERS_ASYMM_H 4#define ARM_COMPUTE_HELPERS_ASYMM_H 5 6 7#ifndef ARM_COMPUTE_HELPER_H 8#define ARM_COMPUTE_HELPER_H 9 10 11 12 13#define STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 14 VSTORE(N0) \ 15 (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0)); 16 17#define STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 18 STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 19 VSTORE(N0) \ 20 (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1)); 21 22#define STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 23 STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 24 VSTORE(N0) \ 25 (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2)); 26 27#define STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 28 STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 29 VSTORE(N0) \ 30 (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3)); 31 32#define STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 33 STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 34 VSTORE(N0) \ 35 (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4)); 36 37#define STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 38 STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 39 VSTORE(N0) \ 40 (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5)); 41 42#define STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 43 STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 44 VSTORE(N0) \ 45 (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6)); 46 47#define STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 48 STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 49 VSTORE(N0) \ 50 (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7)); 51 52#define STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 53 STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 54 VSTORE(N0) \ 55 (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8)); 56 57#define STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 58 STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 59 VSTORE(N0) \ 60 (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9)); 61 62#define STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 63 STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 64 VSTORE(N0) \ 65 (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A)); 66 67#define STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 68 STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 69 VSTORE(N0) \ 70 (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B)); 71 72#define STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 73 STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 74 VSTORE(N0) \ 75 (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C)); 76 77#define STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 78 STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 79 VSTORE(N0) \ 80 (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D)); 81 82#define STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 83 STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 84 VSTORE(N0) \ 85 (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E)); 86 87#define STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 88 STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 89 VSTORE(N0) \ 90 (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F)); 91 92 93 94#define CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 95 VSTORE(N0) \ 96 (CONVERT_SAT((BASENAME##0), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0)); 97 98#define CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 99 CONVERT_STORE_ROW_1(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 100 VSTORE(N0) \ 101 (CONVERT_SAT((BASENAME##1), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1)); 102 103#define CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 104 CONVERT_STORE_ROW_2(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 105 VSTORE(N0) \ 106 (CONVERT_SAT((BASENAME##2), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2)); 107 108#define CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 109 CONVERT_STORE_ROW_3(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 110 VSTORE(N0) \ 111 (CONVERT_SAT((BASENAME##3), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3)); 112 113#define CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 114 CONVERT_STORE_ROW_4(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 115 VSTORE(N0) \ 116 (CONVERT_SAT((BASENAME##4), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4)); 117 118#define CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 119 CONVERT_STORE_ROW_5(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 120 VSTORE(N0) \ 121 (CONVERT_SAT((BASENAME##5), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5)); 122 123#define CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 124 CONVERT_STORE_ROW_6(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 125 VSTORE(N0) \ 126 (CONVERT_SAT((BASENAME##6), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6)); 127 128#define CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 129 CONVERT_STORE_ROW_7(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 130 VSTORE(N0) \ 131 (CONVERT_SAT((BASENAME##7), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7)); 132 133#define CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 134 CONVERT_STORE_ROW_8(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 135 VSTORE(N0) \ 136 (CONVERT_SAT((BASENAME##8), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8)); 137 138#define CONVERT_STORE_ROW_10(N0, DATA, BASENAME, PTR, STRIDE_Y, Z) \ 139 CONVERT_STORE_ROW_9(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 140 VSTORE(N0) \ 141 (CONVERT_SAT((BASENAME##9), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9)); 142 143#define CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 144 CONVERT_STORE_ROW_10(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 145 VSTORE(N0) \ 146 (CONVERT_SAT((BASENAME##A), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A)); 147 148#define CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 149 CONVERT_STORE_ROW_11(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 150 VSTORE(N0) \ 151 (CONVERT_SAT((BASENAME##B), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B)); 152 153#define CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 154 CONVERT_STORE_ROW_12(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 155 VSTORE(N0) \ 156 (CONVERT_SAT((BASENAME##C), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C)); 157 158#define CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 159 CONVERT_STORE_ROW_13(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 160 VSTORE(N0) \ 161 (CONVERT_SAT((BASENAME##D), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D)); 162 163#define CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 164 CONVERT_STORE_ROW_14(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 165 VSTORE(N0) \ 166 (CONVERT_SAT((BASENAME##E), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E)); 167 168#define CONVERT_STORE_ROW_16(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 169 CONVERT_STORE_ROW_15(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 170 VSTORE(N0) \ 171 (CONVERT_SAT((BASENAME##F), VEC_DATA_TYPE(DATA_TYPE, N0)), 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F)); 172 173 174 175 176#define STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 177#define STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 178 179 180 181#define CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_ROW_##M0(N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 182#define CONVERT_STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) CONVERT_STORE_BLOCK_STR(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 183 184 185 186#define STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 187 VSTORE_PARTIAL(N0, STORE_N0) \ 188 (BASENAME##0, 0, (__global DATA_TYPE *)(PTR + 0 * STRIDE_Y + Z##0)); 189 190#define STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 191 STORE_ROW_PARTIAL_1(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 192 VSTORE_PARTIAL(N0, STORE_N0) \ 193 (BASENAME##1, 0, (__global DATA_TYPE *)(PTR + 1 * STRIDE_Y + Z##1)); 194 195#define STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 196 STORE_ROW_PARTIAL_2(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 197 VSTORE_PARTIAL(N0, STORE_N0) \ 198 (BASENAME##2, 0, (__global DATA_TYPE *)(PTR + 2 * STRIDE_Y + Z##2)); 199 200#define STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 201 STORE_ROW_PARTIAL_3(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 202 VSTORE_PARTIAL(N0, STORE_N0) \ 203 (BASENAME##3, 0, (__global DATA_TYPE *)(PTR + 3 * STRIDE_Y + Z##3)); 204 205#define STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 206 STORE_ROW_PARTIAL_4(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 207 VSTORE_PARTIAL(N0, STORE_N0) \ 208 (BASENAME##4, 0, (__global DATA_TYPE *)(PTR + 4 * STRIDE_Y + Z##4)); 209 210#define STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 211 STORE_ROW_PARTIAL_5(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 212 VSTORE_PARTIAL(N0, STORE_N0) \ 213 (BASENAME##5, 0, (__global DATA_TYPE *)(PTR + 5 * STRIDE_Y + Z##5)); 214 215#define STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 216 STORE_ROW_PARTIAL_6(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 217 VSTORE_PARTIAL(N0, STORE_N0) \ 218 (BASENAME##6, 0, (__global DATA_TYPE *)(PTR + 6 * STRIDE_Y + Z##6)); 219 220#define STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 221 STORE_ROW_PARTIAL_7(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 222 VSTORE_PARTIAL(N0, STORE_N0) \ 223 (BASENAME##7, 0, (__global DATA_TYPE *)(PTR + 7 * STRIDE_Y + Z##7)); 224 225#define STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 226 STORE_ROW_PARTIAL_8(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 227 VSTORE_PARTIAL(N0, STORE_N0) \ 228 (BASENAME##8, 0, (__global DATA_TYPE *)(PTR + 8 * STRIDE_Y + Z##8)); 229 230#define STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 231 STORE_ROW_PARTIAL_9(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 232 VSTORE_PARTIAL(N0, STORE_N0) \ 233 (BASENAME##9, 0, (__global DATA_TYPE *)(PTR + 9 * STRIDE_Y + Z##9)); 234 235#define STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 236 STORE_ROW_PARTIAL_10(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 237 VSTORE_PARTIAL(N0, STORE_N0) \ 238 (BASENAME##A, 0, (__global DATA_TYPE *)(PTR + 10 * STRIDE_Y + Z##A)); 239 240#define STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 241 STORE_ROW_PARTIAL_11(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 242 VSTORE_PARTIAL(N0, STORE_N0) \ 243 (BASENAME##B, 0, (__global DATA_TYPE *)(PTR + 11 * STRIDE_Y + Z##B)); 244 245#define STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 246 STORE_ROW_PARTIAL_12(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 247 VSTORE_PARTIAL(N0, STORE_N0) \ 248 (BASENAME##C, 0, (__global DATA_TYPE *)(PTR + 12 * STRIDE_Y + Z##C)); 249 250#define STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 251 STORE_ROW_PARTIAL_13(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 252 VSTORE_PARTIAL(N0, STORE_N0) \ 253 (BASENAME##D, 0, (__global DATA_TYPE *)(PTR + 13 * STRIDE_Y + Z##D)); 254 255#define STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 256 STORE_ROW_PARTIAL_14(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 257 VSTORE_PARTIAL(N0, STORE_N0) \ 258 (BASENAME##E, 0, (__global DATA_TYPE *)(PTR + 14 * STRIDE_Y + Z##E)); 259 260#define STORE_ROW_PARTIAL_16(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 261 STORE_ROW_PARTIAL_15(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) \ 262 VSTORE_PARTIAL(N0, STORE_N0) \ 263 (BASENAME##F, 0, (__global DATA_TYPE *)(PTR + 15 * STRIDE_Y + Z##F)); 264 265 266 267#define STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_ROW_PARTIAL_##STORE_M0(N0, STORE_N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 268#define STORE_BLOCK_PARTIAL(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) STORE_BLOCK_PARTIAL_STR(STORE_M0, STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 269 270#define STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ 271 if(!(PARTIAL_COND_X) && !(PARTIAL_COND_Y)) \ 272 { \ 273 STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 274 } \ 275 else if((PARTIAL_COND_Y) && !(PARTIAL_COND_X)) \ 276 { \ 277 STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 278 } \ 279 else if(!(PARTIAL_COND_Y) && (PARTIAL_COND_X)) \ 280 { \ 281 STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 282 } \ 283 else \ 284 { \ 285 STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 286 } 287 288#define STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) \ 289 if(!(PARTIAL_COND_X)) \ 290 { \ 291 STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 292 } \ 293 else \ 294 { \ 295 STORE_BLOCK_PARTIAL(M0, PARTIAL_STORE_N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 296 } 297 298#define STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) \ 299 if(!(PARTIAL_COND_Y)) \ 300 { \ 301 STORE_BLOCK_PARTIAL(M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 302 } \ 303 else \ 304 { \ 305 STORE_BLOCK_PARTIAL(PARTIAL_STORE_M0, N0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z); \ 306 } 307 308 309#if defined(PARTIAL_STORE_M0) && defined(PARTIAL_STORE_N0) 310 311 312#if PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 == 0 313 314#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ 315 STORE_BLOCK(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z) 316 317#elif PARTIAL_STORE_M0 > 0 && PARTIAL_STORE_N0 == 0 318 319#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ 320 STORE_BLOCK_PARTIAL_IN_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_COND_Y) 321 322#elif PARTIAL_STORE_M0 == 0 && PARTIAL_STORE_N0 > 0 323 324#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ 325 STORE_BLOCK_PARTIAL_IN_X(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_N0, PARTIAL_COND_X) 326 327#else 328 329#define STORE_BLOCK_BOUNDARY_AWARE(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) \ 330 STORE_BLOCK_PARTIAL_IN_X_AND_Y(M0, N0, DATA_TYPE, BASENAME, PTR, STRIDE_Y, Z, PARTIAL_STORE_M0, PARTIAL_STORE_N0, PARTIAL_COND_Y, PARTIAL_COND_X) 331 332#endif 333 334#endif 335 336 337#if defined(PARTIAL_STORE_M0) 338 339#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \ 340 ((uint)(max(0, (int)(y * M0) - (int)((M0 - PARTIAL_STORE_M0) % M0)))) 341#else 342#define COMPUTE_M0_START_ROW(y, M0, PARTIAL_STORE_M0) \ 343 ((uint)(y * M0)) 344#endif 345 346 347 348#define STORE_VECTOR_SELECT(basename, data_type, ptr, vec_size, leftover, cond) \ 349 STORE_BLOCK_PARTIAL_IN_X(1, vec_size, data_type, basename, ptr, 0, 0, leftover, cond) 350 351 352#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) 353#pragma OPENCL EXTENSION cl_khr_fp16 : enable 354#endif 355 356#if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) 357#pragma OPENCL EXTENSION cl_arm_integer_dot_product_int8 : enable 358#endif 359 360#if defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) 361#pragma OPENCL EXTENSION cl_arm_integer_dot_product_accumulate_int8 : enable 362#endif 363 364#if defined(ARM_COMPUTE_DEBUG_ENABLED) && defined(cl_arm_printf) 365#pragma OPENCL EXTENSION cl_arm_printf : enable 366#endif 367 368#define GPU_ARCH_MIDGARD 0x100 369#define GPU_ARCH_BIFROST 0x200 370#define GPU_ARCH_VALHALL 0x300 371 372 373#define CONCAT(a, b) a##b 374 375 376#define EXPAND(x) x 377 378 379#define CLAMP(x, min_val, max_val) min(max(x, min_val), max_val) 380 381 382#define REV1(x) ((x)) 383#define REV2(x) ((x).s10) 384#define REV3(x) ((x).s210) 385#define REV4(x) ((x).s3210) 386#define REV8(x) ((x).s76543210) 387#define REV16(x) ((x).sFEDCBA9876543210) 388 389 390 391#define REVERSE_STR(x, s) REV##s((x)) 392#define REVERSE(x, s) REVERSE_STR(x, s) 393 394 395 396#define ROT1_0(x) ((x)) 397#define ROT1_1(x) ((x)) 398 399#define ROT2_0(x) ((x)) 400#define ROT2_1(x) ((x).s10) 401#define ROT2_2(x) ((x)) 402 403#define ROT3_0(x) ((x)) 404#define ROT3_1(x) ((x).s201) 405#define ROT3_2(x) ((x).s120) 406#define ROT3_3(x) ((x)) 407 408#define ROT4_0(x) ((x)) 409#define ROT4_1(x) ((x).s3012) 410#define ROT4_2(x) ((x).s2301) 411#define ROT4_3(x) ((x).s1230) 412#define ROT4_4(x) ((x)) 413 414#define ROT8_0(x) ((x)) 415#define ROT8_1(x) ((x).s70123456) 416#define ROT8_2(x) ((x).s67012345) 417#define ROT8_3(x) ((x).s56701234) 418#define ROT8_4(x) ((x).s45670123) 419#define ROT8_5(x) ((x).s34567012) 420#define ROT8_6(x) ((x).s23456701) 421#define ROT8_7(x) ((x).s12345670) 422#define ROT8_8(x) ((x)) 423 424#define ROT16_0(x) ((x)) 425#define ROT16_1(x) ((x).sF0123456789ABCDE) 426#define ROT16_2(x) ((x).sEF0123456789ABCD) 427#define ROT16_3(x) ((x).sDEF0123456789ABC) 428#define ROT16_4(x) ((x).sCDEF0123456789AB) 429#define ROT16_5(x) ((x).sBCDEF0123456789A) 430#define ROT16_6(x) ((x).sABCDEF0123456789) 431#define ROT16_7(x) ((x).s9ABCDEF012345678) 432#define ROT16_8(x) ((x).s89ABCDEF01234567) 433#define ROT16_9(x) ((x).s789ABCDEF0123456) 434#define ROT16_10(x) ((x).s6789ABCDEF012345) 435#define ROT16_11(x) ((x).s56789ABCDEF01234) 436#define ROT16_12(x) ((x).s456789ABCDEF0123) 437#define ROT16_13(x) ((x).s3456789ABCDEF012) 438#define ROT16_14(x) ((x).s23456789ABCDEF01) 439#define ROT16_15(x) ((x).s123456789ABCDEF0) 440#define ROT16_16(x) ((x)) 441 442 443 444#define ROTATE_STR(x, s, n) ROT##s##_##n(x) 445#define ROTATE(x, s, n) ROTATE_STR(x, s, n) 446 447 448 449#define V_OFFS1(dt) (dt##1)(0) 450#define V_OFFS2(dt) (dt##2)(0, 1) 451#define V_OFFS3(dt) (dt##3)(0, 1, 2) 452#define V_OFFS4(dt) (dt##4)(0, 1, 2, 3) 453#define V_OFFS8(dt) (dt##8)(0, 1, 2, 3, 4, 5, 6, 7) 454#define V_OFFS16(dt) (dt##16)(0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15) 455 456 457 458#define VEC_OFFS_STR(dt, s) V_OFFS##s(dt) 459#define VEC_OFFS(dt, s) VEC_OFFS_STR(dt, s) 460 461 462#define VLOAD_STR(size) vload##size 463#define VLOAD(size) VLOAD_STR(size) 464 465 466#define VLOAD_PARTIAL_STR(size, load_size) vload_partial_##size##_##load_size 467#define VLOAD_PARTIAL(size, load_size) VLOAD_PARTIAL_STR(size, load_size) 468 469#define NO_LOAD(data, offs, ptr) \ 470 { \ 471 } 472 473 474#define vload_partial_1_0 NO_LOAD 475#define vload_partial_1_1 vload1 476#define vload_partial_1_2 NO_LOAD 477#define vload_partial_1_3 NO_LOAD 478#define vload_partial_1_4 NO_LOAD 479#define vload_partial_1_5 NO_LOAD 480#define vload_partial_1_6 NO_LOAD 481#define vload_partial_1_7 NO_LOAD 482#define vload_partial_1_8 NO_LOAD 483#define vload_partial_1_9 NO_LOAD 484#define vload_partial_1_10 NO_LOAD 485#define vload_partial_1_11 NO_LOAD 486#define vload_partial_1_12 NO_LOAD 487#define vload_partial_1_13 NO_LOAD 488#define vload_partial_1_14 NO_LOAD 489#define vload_partial_1_15 NO_LOAD 490#define vload_partial_1_16 NO_LOAD 491 492#define vload_partial_2_0 NO_LOAD 493#define vload_partial_2_1 vload_partial_1 494#define vload_partial_2_2 vload_partial_2 495#define vload_partial_2_3 NO_LOAD 496#define vload_partial_2_4 NO_LOAD 497#define vload_partial_2_5 NO_LOAD 498#define vload_partial_2_6 NO_LOAD 499#define vload_partial_2_7 NO_LOAD 500#define vload_partial_2_8 NO_LOAD 501#define vload_partial_2_9 NO_LOAD 502#define vload_partial_2_10 NO_LOAD 503#define vload_partial_2_11 NO_LOAD 504#define vload_partial_2_12 NO_LOAD 505#define vload_partial_2_13 NO_LOAD 506#define vload_partial_2_14 NO_LOAD 507#define vload_partial_2_15 NO_LOAD 508#define vload_partial_2_16 NO_LOAD 509 510#define vload_partial_3_0 NO_LOAD 511#define vload_partial_3_1 vload_partial_1 512#define vload_partial_3_2 vload_partial_2 513#define vload_partial_3_3 vload_partial_3 514#define vload_partial_3_4 NO_LOAD 515#define vload_partial_3_5 NO_LOAD 516#define vload_partial_3_6 NO_LOAD 517#define vload_partial_3_7 NO_LOAD 518#define vload_partial_3_8 NO_LOAD 519#define vload_partial_3_9 NO_LOAD 520#define vload_partial_3_10 NO_LOAD 521#define vload_partial_3_11 NO_LOAD 522#define vload_partial_3_12 NO_LOAD 523#define vload_partial_3_13 NO_LOAD 524#define vload_partial_3_14 NO_LOAD 525#define vload_partial_3_15 NO_LOAD 526#define vload_partial_3_16 NO_LOAD 527 528#define vload_partial_4_0 NO_LOAD 529#define vload_partial_4_1 vload_partial_1 530#define vload_partial_4_2 vload_partial_2 531#define vload_partial_4_3 vload_partial_3 532#define vload_partial_4_4 vload_partial_4 533#define vload_partial_4_5 NO_LOAD 534#define vload_partial_4_6 NO_LOAD 535#define vload_partial_4_7 NO_LOAD 536#define vload_partial_4_8 NO_LOAD 537#define vload_partial_4_9 NO_LOAD 538#define vload_partial_4_10 NO_LOAD 539#define vload_partial_4_11 NO_LOAD 540#define vload_partial_4_12 NO_LOAD 541#define vload_partial_4_13 NO_LOAD 542#define vload_partial_4_14 NO_LOAD 543#define vload_partial_4_15 NO_LOAD 544#define vload_partial_4_16 NO_LOAD 545 546#define vload_partial_8_0 NO_LOAD 547#define vload_partial_8_1 vload_partial_1 548#define vload_partial_8_2 vload_partial_2 549#define vload_partial_8_3 vload_partial_3 550#define vload_partial_8_4 vload_partial_4 551#define vload_partial_8_5 vload_partial_5 552#define vload_partial_8_6 vload_partial_6 553#define vload_partial_8_7 vload_partial_7 554#define vload_partial_8_8 vload_partial_8 555#define vload_partial_8_9 NO_LOAD 556#define vload_partial_8_10 NO_LOAD 557#define vload_partial_8_11 NO_LOAD 558#define vload_partial_8_12 NO_LOAD 559#define vload_partial_8_13 NO_LOAD 560#define vload_partial_8_14 NO_LOAD 561#define vload_partial_8_15 NO_LOAD 562#define vload_partial_8_16 NO_LOAD 563 564#define vload_partial_16_0 NO_LOAD 565#define vload_partial_16_1 vload_partial_1 566#define vload_partial_16_2 vload_partial_2 567#define vload_partial_16_3 vload_partial_3 568#define vload_partial_16_4 vload_partial_4 569#define vload_partial_16_5 vload_partial_5 570#define vload_partial_16_6 vload_partial_6 571#define vload_partial_16_7 vload_partial_7 572#define vload_partial_16_8 vload_partial_8 573#define vload_partial_16_9 vload_partial_9 574#define vload_partial_16_10 vload_partial_10 575#define vload_partial_16_11 vload_partial_11 576#define vload_partial_16_12 vload_partial_12 577#define vload_partial_16_13 vload_partial_13 578#define vload_partial_16_14 vload_partial_14 579#define vload_partial_16_15 vload_partial_15 580#define vload_partial_16_16 vload_partial_16 581 582 583#define vload_partial_1(DATA, OFFSET, PTR) \ 584 DATA.s0 = vload1(OFFSET, PTR); 585 586#define vload_partial_2(DATA, OFFSET, PTR) \ 587 DATA.s01 = vload2(OFFSET, PTR); 588 589#define vload_partial_3(DATA, OFFSET, PTR) \ 590 DATA.s012 = vload3(OFFSET, PTR); 591 592#define vload_partial_4(DATA, OFFSET, PTR) \ 593 DATA.s0123 = vload4(OFFSET, PTR); 594 595#define vload_partial_5(DATA, OFFSET, PTR) \ 596 vload_partial_4(DATA.s0123, OFFSET, PTR); \ 597 DATA.s4 = vload1(OFFSET, PTR + 4); 598 599#define vload_partial_6(DATA, OFFSET, PTR) \ 600 vload_partial_4(DATA.s0123, OFFSET, PTR); \ 601 vload_partial_2(DATA.s45, OFFSET, PTR + 4); 602 603#define vload_partial_7(DATA, OFFSET, PTR) \ 604 vload_partial_4(DATA.s0123, OFFSET, PTR); \ 605 vload_partial_3(DATA.s456, OFFSET, PTR + 4); 606 607#define vload_partial_8(DATA, OFFSET, PTR) \ 608 DATA.s01234567 = vload8(OFFSET, PTR); 609 610#define vload_partial_9(DATA, OFFSET, PTR) \ 611 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 612 DATA.s8 = vload1(OFFSET, PTR + 8); 613 614#define vload_partial_10(DATA, OFFSET, PTR) \ 615 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 616 vload_partial_2(DATA.s89, OFFSET, PTR + 8); 617 618#define vload_partial_11(DATA, OFFSET, PTR) \ 619 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 620 vload_partial_3(DATA.s89A, OFFSET, PTR + 8); 621 622#define vload_partial_12(DATA, OFFSET, PTR) \ 623 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 624 vload_partial_4(DATA.s89AB, OFFSET, PTR + 8); 625 626#define vload_partial_13(DATA, OFFSET, PTR) \ 627 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 628 vload_partial_5(DATA.s89ABCDEF, OFFSET, PTR + 8); 629 630#define vload_partial_14(DATA, OFFSET, PTR) \ 631 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 632 vload_partial_6(DATA.s89ABCDEF, OFFSET, PTR + 8); 633 634#define vload_partial_15(DATA, OFFSET, PTR) \ 635 vload_partial_8(DATA.s01234567, OFFSET, PTR); \ 636 vload_partial_7(DATA.s89ABCDEF, OFFSET, PTR + 8); 637 638#define vload_partial_16(DATA, OFFSET, PTR) \ 639 DATA = vload16(OFFSET, PTR); 640 641 642 643#define PIXEL_UNIT4 1 644#define PIXEL_UNIT8 2 645#define PIXEL_UNIT16 4 646 647 648#define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) PIXEL_UNIT##vec_size 649#define CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(vec_size) CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT_STR(vec_size) 650 651 652#define read_image2d_floatx1(img, x_coord, y_coord) (float4)(read_imagef(img, (int2)(x_coord, y_coord))); 653#define read_image2d_floatx2(img, x_coord, y_coord) (float8)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord))); 654#define read_image2d_floatx4(img, x_coord, y_coord) (float16)(read_imagef(img, (int2)(x_coord, y_coord)), read_imagef(img, (int2)(x_coord + 1, y_coord)), read_imagef(img, (int2)(x_coord + 2, y_coord)), read_imagef(img, (int2)(x_coord + 3, y_coord))); 655 656#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) 657#define read_image2d_halfx1(img, x_coord, y_coord) (half4)(read_imageh(img, (int2)(x_coord, y_coord))); 658#define read_image2d_halfx2(img, x_coord, y_coord) (half8)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord))); 659#define read_image2d_halfx4(img, x_coord, y_coord) (half16)(read_imageh(img, (int2)(x_coord, y_coord)), read_imageh(img, (int2)(x_coord + 1, y_coord)), read_imageh(img, (int2)(x_coord + 2, y_coord)), read_imageh(img, (int2)(x_coord + 3, y_coord))); 660#endif 661 662#define write_image2d_floatx1(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values)); 663#define write_image2d_floatx2(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values.s0123), write_imagef(img, (int2)(x_coord + 1, y_coord), values.s4567)); 664#define write_image2d_floatx4(img, x_coord, y_coord, values) (write_imagef(img, (int2)(x_coord, y_coord), values.s0123), write_imagef(img, (int2)(x_coord + 1, y_coord), values.s4567), write_imagef(img, (int2)(x_coord + 2, y_coord), values.s89AB), write_imagef(img, (int2)(x_coord + 3, y_coord), values.sCDEF)); 665 666#if defined(ARM_COMPUTE_OPENCL_FP16_ENABLED) && defined(cl_khr_fp16) 667#define write_image2d_halfx1(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values)); 668#define write_image2d_halfx2(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values.s0123), write_imageh(img, (int2)(x_coord + 1, y_coord), values.s4567)); 669#define write_image2d_halfx4(img, x_coord, y_coord, values) (write_imageh(img, (int2)(x_coord, y_coord), values.s0123), write_imageh(img, (int2)(x_coord + 1, y_coord), values.s4567), write_imageh(img, (int2)(x_coord + 2, y_coord), values.s89AB), write_imageh(img, (int2)(x_coord + 3, y_coord), values.sCDEF)); 670#endif 671 672 673#define READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) read_image2d_##data_type##x##n0(img, x_coord, y_coord) 674#define READ_IMAGE2D(data_type, n0, img, x_coord, y_coord) READ_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord) 675 676 677#define WRITE_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord, values) write_image2d_##data_type##x##n0(img, x_coord, y_coord, values) 678#define WRITE_IMAGE2D(data_type, n0, img, x_coord, y_coord, values) WRITE_IMAGE2D_STR(data_type, n0, img, x_coord, y_coord, values) 679 680#define VSTORE_STR(size) vstore##size 681#define VSTORE(size) VSTORE_STR(size) 682 683#define float1 float 684#define half1 half 685#define char1 char 686#define uchar1 uchar 687#define short1 short 688#define ushort1 ushort 689#define int1 int 690#define uint1 uint 691#define long1 long 692#define ulong1 ulong 693#define double1 double 694 695#define vload1(OFFSET, PTR) *(OFFSET + PTR) 696#define vstore1(DATA, OFFSET, PTR) *(OFFSET + PTR) = DATA 697 698 699#define VSTORE_PARTIAL_STR(size, store_size) vstore_partial_##size##_##store_size 700#define VSTORE_PARTIAL(size, store_size) VSTORE_PARTIAL_STR(size, store_size) 701 702#define NO_STORE(data, offs, ptr) \ 703 { \ 704 } 705 706 707#define vstore_partial_1_0 NO_STORE 708#define vstore_partial_1_1 vstore1 709#define vstore_partial_1_2 NO_STORE 710#define vstore_partial_1_3 NO_STORE 711#define vstore_partial_1_4 NO_STORE 712#define vstore_partial_1_5 NO_STORE 713#define vstore_partial_1_6 NO_STORE 714#define vstore_partial_1_7 NO_STORE 715#define vstore_partial_1_8 NO_STORE 716#define vstore_partial_1_9 NO_STORE 717#define vstore_partial_1_10 NO_STORE 718#define vstore_partial_1_11 NO_STORE 719#define vstore_partial_1_12 NO_STORE 720#define vstore_partial_1_13 NO_STORE 721#define vstore_partial_1_14 NO_STORE 722#define vstore_partial_1_15 NO_STORE 723#define vstore_partial_1_16 NO_STORE 724 725#define vstore_partial_2_0 NO_STORE 726#define vstore_partial_2_1 vstore_partial_1 727#define vstore_partial_2_2 vstore_partial_2 728#define vstore_partial_2_3 NO_STORE 729#define vstore_partial_2_4 NO_STORE 730#define vstore_partial_2_5 NO_STORE 731#define vstore_partial_2_6 NO_STORE 732#define vstore_partial_2_7 NO_STORE 733#define vstore_partial_2_8 NO_STORE 734#define vstore_partial_2_9 NO_STORE 735#define vstore_partial_2_10 NO_STORE 736#define vstore_partial_2_11 NO_STORE 737#define vstore_partial_2_12 NO_STORE 738#define vstore_partial_2_13 NO_STORE 739#define vstore_partial_2_14 NO_STORE 740#define vstore_partial_2_15 NO_STORE 741#define vstore_partial_2_16 NO_STORE 742 743#define vstore_partial_3_0 NO_STORE 744#define vstore_partial_3_1 vstore_partial_1 745#define vstore_partial_3_2 vstore_partial_2 746#define vstore_partial_3_3 vstore_partial_3 747#define vstore_partial_3_4 NO_STORE 748#define vstore_partial_3_5 NO_STORE 749#define vstore_partial_3_6 NO_STORE 750#define vstore_partial_3_7 NO_STORE 751#define vstore_partial_3_8 NO_STORE 752#define vstore_partial_3_9 NO_STORE 753#define vstore_partial_3_10 NO_STORE 754#define vstore_partial_3_11 NO_STORE 755#define vstore_partial_3_12 NO_STORE 756#define vstore_partial_3_13 NO_STORE 757#define vstore_partial_3_14 NO_STORE 758#define vstore_partial_3_15 NO_STORE 759#define vstore_partial_3_16 NO_STORE 760 761#define vstore_partial_4_0 NO_STORE 762#define vstore_partial_4_1 vstore_partial_1 763#define vstore_partial_4_2 vstore_partial_2 764#define vstore_partial_4_3 vstore_partial_3 765#define vstore_partial_4_4 vstore_partial_4 766#define vstore_partial_4_5 NO_STORE 767#define vstore_partial_4_6 NO_STORE 768#define vstore_partial_4_7 NO_STORE 769#define vstore_partial_4_8 NO_STORE 770#define vstore_partial_4_9 NO_STORE 771#define vstore_partial_4_10 NO_STORE 772#define vstore_partial_4_11 NO_STORE 773#define vstore_partial_4_12 NO_STORE 774#define vstore_partial_4_13 NO_STORE 775#define vstore_partial_4_14 NO_STORE 776#define vstore_partial_4_15 NO_STORE 777#define vstore_partial_4_16 NO_STORE 778 779#define vstore_partial_8_0 NO_STORE 780#define vstore_partial_8_1 vstore_partial_1 781#define vstore_partial_8_2 vstore_partial_2 782#define vstore_partial_8_3 vstore_partial_3 783#define vstore_partial_8_4 vstore_partial_4 784#define vstore_partial_8_5 vstore_partial_5 785#define vstore_partial_8_6 vstore_partial_6 786#define vstore_partial_8_7 vstore_partial_7 787#define vstore_partial_8_8 vstore_partial_8 788#define vstore_partial_8_9 NO_STORE 789#define vstore_partial_8_10 NO_STORE 790#define vstore_partial_8_11 NO_STORE 791#define vstore_partial_8_12 NO_STORE 792#define vstore_partial_8_13 NO_STORE 793#define vstore_partial_8_14 NO_STORE 794#define vstore_partial_8_15 NO_STORE 795#define vstore_partial_8_16 NO_STORE 796 797#define vstore_partial_16_0 NO_STORE 798#define vstore_partial_16_1 vstore_partial_1 799#define vstore_partial_16_2 vstore_partial_2 800#define vstore_partial_16_3 vstore_partial_3 801#define vstore_partial_16_4 vstore_partial_4 802#define vstore_partial_16_5 vstore_partial_5 803#define vstore_partial_16_6 vstore_partial_6 804#define vstore_partial_16_7 vstore_partial_7 805#define vstore_partial_16_8 vstore_partial_8 806#define vstore_partial_16_9 vstore_partial_9 807#define vstore_partial_16_10 vstore_partial_10 808#define vstore_partial_16_11 vstore_partial_11 809#define vstore_partial_16_12 vstore_partial_12 810#define vstore_partial_16_13 vstore_partial_13 811#define vstore_partial_16_14 vstore_partial_14 812#define vstore_partial_16_15 vstore_partial_15 813#define vstore_partial_16_16 vstore_partial_16 814 815 816#define vstore_partial_1(DATA, OFFSET, PTR) \ 817 vstore1(DATA.s0, OFFSET, PTR); 818 819#define vstore_partial_2(DATA, OFFSET, PTR) \ 820 vstore2(DATA.s01, OFFSET, PTR); 821 822#define vstore_partial_3(DATA, OFFSET, PTR) \ 823 vstore3(DATA.s012, OFFSET, PTR); 824 825#define vstore_partial_4(DATA, OFFSET, PTR) \ 826 vstore4(DATA.s0123, OFFSET, PTR); 827 828#define vstore_partial_5(DATA, OFFSET, PTR) \ 829 vstore_partial_4(DATA.s0123, OFFSET, PTR); \ 830 vstore1(DATA.s4, OFFSET, PTR + 4); 831 832#define vstore_partial_6(DATA, OFFSET, PTR) \ 833 vstore_partial_4(DATA.s0123, OFFSET, PTR); \ 834 vstore_partial_2(DATA.s45, OFFSET, PTR + 4); 835 836#define vstore_partial_7(DATA, OFFSET, PTR) \ 837 vstore_partial_4(DATA.s0123, OFFSET, PTR); \ 838 vstore_partial_3(DATA.s456, OFFSET, PTR + 4); 839 840#define vstore_partial_8(DATA, OFFSET, PTR) \ 841 vstore8(DATA.s01234567, OFFSET, PTR); 842 843#define vstore_partial_9(DATA, OFFSET, PTR) \ 844 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 845 vstore1(DATA.s8, OFFSET, PTR + 8); 846 847#define vstore_partial_10(DATA, OFFSET, PTR) \ 848 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 849 vstore_partial_2(DATA.s89, OFFSET, PTR + 8); 850 851#define vstore_partial_11(DATA, OFFSET, PTR) \ 852 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 853 vstore_partial_3(DATA.s89a, OFFSET, PTR + 8); 854 855#define vstore_partial_12(DATA, OFFSET, PTR) \ 856 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 857 vstore_partial_4(DATA.s89ab, OFFSET, PTR + 8); 858 859#define vstore_partial_13(DATA, OFFSET, PTR) \ 860 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 861 vstore_partial_5(DATA.s89abcdef, OFFSET, PTR + 8); 862 863#define vstore_partial_14(DATA, OFFSET, PTR) \ 864 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 865 vstore_partial_6(DATA.s89abcdef, OFFSET, PTR + 8); 866 867#define vstore_partial_15(DATA, OFFSET, PTR) \ 868 vstore_partial_8(DATA.s01234567, OFFSET, PTR); \ 869 vstore_partial_7(DATA.s89abcdef, OFFSET, PTR + 8); 870 871#define vstore_partial_16(DATA, OFFSET, PTR) \ 872 vstore16(DATA, OFFSET, PTR); 873 874 875 876 877 878#define convert_float_sat convert_float 879#define convert_float1_sat convert_float 880#define convert_float2_sat convert_float2 881#define convert_float3_sat convert_float3 882#define convert_float4_sat convert_float4 883#define convert_float8_sat convert_float8 884#define convert_float16_sat convert_float16 885#define convert_half_sat convert_float 886#define convert_half1_sat convert_half 887#define convert_half2_sat convert_half2 888#define convert_half3_sat convert_half3 889#define convert_half4_sat convert_half4 890#define convert_half8_sat convert_half8 891#define convert_half16_sat convert_half16 892 893#define convert_float1 convert_float 894#define convert_half1 convert_half 895#define convert_char1 convert_char 896#define convert_uchar1 convert_uchar 897#define convert_short1 convert_short 898#define convert_ushort1 convert_ushort 899#define convert_int1 convert_int 900#define convert_uint1 convert_uint 901#define convert_long1 convert_long 902#define convert_ulong1 convert_ulong 903#define convert_double1 convert_double 904 905#define convert_char1_sat convert_char_sat 906#define convert_uchar1_sat convert_uchar_sat 907#define convert_uchar2_sat convert_uchar2_sat 908#define convert_uchar3_sat convert_uchar3_sat 909#define convert_uchar4_sat convert_uchar4_sat 910#define convert_uchar8_sat convert_uchar8_sat 911#define convert_uchar16_sat convert_uchar16_sat 912#define convert_short1_sat convert_short_sat 913#define convert_ushort1_sat convert_ushort_sat 914#define convert_int1_sat convert_int_sat 915#define convert_uint1_sat convert_uint_sat 916#define convert_long1_sat convert_long_sat 917#define convert_ulong1_sat convert_ulong_sat 918#define convert_double1_sat convert_double_sat 919 920#define VEC_DATA_TYPE_STR(type, size) type##size 921#define VEC_DATA_TYPE(type, size) VEC_DATA_TYPE_STR(type, size) 922 923#define CONVERT_STR(x, type) (convert_##type((x))) 924#define CONVERT(x, type) CONVERT_STR(x, type) 925 926#define CONVERT_SAT_STR(x, type) (convert_##type##_sat((x))) 927#define CONVERT_SAT(x, type) CONVERT_SAT_STR(x, type) 928 929#define CONVERT_SAT_ROUND_STR(x, type, round) (convert_##type##_sat_##round((x))) 930#define CONVERT_SAT_ROUND(x, type, round) CONVERT_SAT_ROUND_STR(x, type, round) 931 932#define select_vec_dt_uchar(size) uchar##size 933#define select_vec_dt_char(size) char##size 934#define select_vec_dt_ushort(size) ushort##size 935#define select_vec_dt_short(size) short##size 936#define select_vec_dt_half(size) short##size 937#define select_vec_dt_uint(size) uint##size 938#define select_vec_dt_int(size) int##size 939#define select_vec_dt_float(size) int##size 940#define select_vec_dt_ulong(size) ulong##size 941#define select_vec_dt_long(size) long##size 942 943#define SELECT_VEC_DATA_TYPE_STR(type, size) select_vec_dt_##type(size) 944#define SELECT_VEC_DATA_TYPE(type, size) SELECT_VEC_DATA_TYPE_STR(type, size) 945#define SELECT_DATA_TYPE(type) SELECT_VEC_DATA_TYPE_STR(type, 1) 946 947#define signed_int_vec_dt_uchar(size) char##size 948#define signed_int_vec_dt_char(size) char##size 949#define signed_int_vec_dt_ushort(size) short##size 950#define signed_int_vec_dt_short(size) short##size 951#define signed_int_vec_dt_half(size) short##size 952#define signed_int_vec_dt_uint(size) int##size 953#define signed_int_vec_dt_int(size) int##size 954#define signed_int_vec_dt_float(size) int##size 955#define signed_int_vec_dt_ulong(size) long##size 956#define signed_int_vec_dt_long(size) long##size 957 958#define SIGNED_INT_VEC_DATA_TYPE_STR(type, size) signed_int_vec_dt_##type(size) 959#define SIGNED_INT_VEC_DATA_TYPE(type, size) SIGNED_INT_VEC_DATA_TYPE_STR(type, size) 960#define SIGNED_INT_DATA_TYPE(type) SIGNED_INT_VEC_DATA_TYPE_STR(type, 1) 961 962#define sum_reduce_1(x) (x) 963#define sum_reduce_2(x) ((x).s0) + ((x).s1) 964#define sum_reduce_3(x) sum_reduce_2((x).s01) + ((x).s2) 965#define sum_reduce_4(x) sum_reduce_2((x).s01) + sum_reduce_2((x).s23) 966#define sum_reduce_8(x) sum_reduce_4((x).s0123) + sum_reduce_4((x).s4567) 967#define sum_reduce_16(x) sum_reduce_8((x).s01234567) + sum_reduce_8((x).s89ABCDEF) 968 969#define SUM_REDUCE_STR(x, size) sum_reduce_##size(x) 970#define SUM_REDUCE(x, size) SUM_REDUCE_STR(x, size) 971 972#define prod_reduce_1(x) (x) 973#define prod_reduce_2(x) ((x).s0) * ((x).s1) 974#define prod_reduce_3(x) prod_reduce_2((x).s01) * ((x).s2) 975#define prod_reduce_4(x) prod_reduce_2((x).s01) * prod_reduce_2((x).s23) 976#define prod_reduce_8(x) prod_reduce_4((x).s0123) * prod_reduce_4((x).s4567) 977#define prod_reduce_16(x) prod_reduce_8((x).s01234567) * prod_reduce_8((x).s89ABCDEF) 978 979#define PROD_REDUCE_STR(x, size) prod_reduce_##size(x) 980#define PROD_REDUCE(x, size) PROD_REDUCE_STR(x, size) 981 982#define max_reduce_1(x) (x) 983#define max_reduce_2(x) max(((x).s0), ((x).s1)) 984#define max_reduce_3(x) max(max_reduce_2((x).s01), ((x).s2)) 985#define max_reduce_4(x) max(max_reduce_2((x).s01), max_reduce_2((x).s23)) 986#define max_reduce_8(x) max(max_reduce_4((x).s0123), max_reduce_4((x).s4567)) 987#define max_reduce_16(x) max(max_reduce_8((x).s01234567), max_reduce_8((x).s89ABCDEF)) 988 989#define MAX_REDUCE_STR(x, size) max_reduce_##size(x) 990#define MAX_REDUCE(x, size) MAX_REDUCE_STR(x, size) 991 992#define VECTOR_DECLARATION(name) \ 993 __global uchar *name##_ptr, \ 994 uint name##_stride_x, \ 995 uint name##_step_x, \ 996 uint name##_offset_first_element_in_bytes 997 998#define IMAGE_DECLARATION(name) \ 999 __global uchar *name##_ptr, \ 1000 uint name##_stride_x, \ 1001 uint name##_step_x, \ 1002 uint name##_stride_y, \ 1003 uint name##_step_y, \ 1004 uint name##_offset_first_element_in_bytes 1005 1006#define TENSOR3D_DECLARATION(name) \ 1007 __global uchar *name##_ptr, \ 1008 uint name##_stride_x, \ 1009 uint name##_step_x, \ 1010 uint name##_stride_y, \ 1011 uint name##_step_y, \ 1012 uint name##_stride_z, \ 1013 uint name##_step_z, \ 1014 uint name##_offset_first_element_in_bytes 1015 1016#define TENSOR4D_DECLARATION(name) \ 1017 __global uchar *name##_ptr, \ 1018 uint name##_stride_x, \ 1019 uint name##_step_x, \ 1020 uint name##_stride_y, \ 1021 uint name##_step_y, \ 1022 uint name##_stride_z, \ 1023 uint name##_step_z, \ 1024 uint name##_stride_w, \ 1025 uint name##_step_w, \ 1026 uint name##_offset_first_element_in_bytes 1027 1028#define TENSOR5D_DECLARATION(name) \ 1029 __global uchar *name##_ptr, \ 1030 uint name##_stride_x, \ 1031 uint name##_step_x, \ 1032 uint name##_stride_y, \ 1033 uint name##_step_y, \ 1034 uint name##_stride_z, \ 1035 uint name##_step_z, \ 1036 uint name##_stride_w, \ 1037 uint name##_step_w, \ 1038 uint name##_stride_v, \ 1039 uint name##_step_v, \ 1040 uint name##_offset_first_element_in_bytes 1041 1042#define CONVERT_TO_VECTOR_STRUCT(name) \ 1043 update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x) 1044 1045#define CONVERT_TO_VECTOR_STRUCT_NO_STEP(name) \ 1046 update_vector_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0) 1047 1048#define CONVERT_TO_IMAGE_STRUCT(name) \ 1049 update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y) 1050 1051#define CONVERT_TO_IMAGE_STRUCT_NO_STEP(name) \ 1052 update_image_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0) 1053 1054#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \ 1055 update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z) 1056 1057#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT_NO_STEP(name) \ 1058 update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, name##_step_z) 1059 1060#define CONVERT_TENSOR3D_TO_IMAGE_STRUCT(name) \ 1061 update_image_from_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, name##_stride_z, name##_step_z) 1062 1063#define CONVERT_TO_TENSOR3D_STRUCT(name) \ 1064 update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \ 1065 name##_stride_z, name##_step_z) 1066 1067#define CONVERT_TO_TENSOR3D_STRUCT_NO_STEP(name) \ 1068 update_tensor3D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0) 1069 1070#define CONVERT_TO_TENSOR4D_STRUCT(name, mod_size) \ 1071 update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \ 1072 name##_stride_z, name##_step_z, name##_stride_w, name##_step_w, mod_size) 1073 1074#define CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(name, mod_size) \ 1075 update_tensor4D_workitem_ptr(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, 0, name##_stride_y, 0, name##_stride_z, 0, name##_stride_w, 0, mod_size) 1076 1077#define CONVERT_TO_TENSOR3D_STRUCT_NO_UPDATE_PTR(name) \ 1078 tensor3D_ptr_no_update(name##_ptr, name##_offset_first_element_in_bytes, name##_stride_x, name##_step_x, name##_stride_y, name##_step_y, \ 1079 name##_stride_z, name##_step_z) 1080 1081 1082typedef struct Vector 1083{ 1084 __global uchar *ptr; 1085 int offset_first_element_in_bytes; 1086 int stride_x; 1087} Vector; 1088 1089 1090typedef struct Image 1091{ 1092 __global uchar *ptr; 1093 int offset_first_element_in_bytes; 1094 int stride_x; 1095 int stride_y; 1096} Image; 1097 1098 1099typedef struct Tensor3D 1100{ 1101 __global uchar *ptr; 1102 int offset_first_element_in_bytes; 1103 int stride_x; 1104 int stride_y; 1105 int stride_z; 1106} Tensor3D; 1107 1108 1109typedef struct Tensor4D 1110{ 1111 __global uchar *ptr; 1112 int offset_first_element_in_bytes; 1113 int stride_x; 1114 int stride_y; 1115 int stride_z; 1116 int stride_w; 1117} Tensor4D; 1118 1119 1120inline Vector update_vector_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x) 1121{ 1122 Vector vector = 1123 { 1124 .ptr = ptr, 1125 .offset_first_element_in_bytes = offset_first_element_in_bytes, 1126 .stride_x = stride_x, 1127 }; 1128 vector.ptr += vector.offset_first_element_in_bytes + get_global_id(0) * step_x; 1129 return vector; 1130} 1131 1132 1133inline Image update_image_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y) 1134{ 1135 Image img = 1136 { 1137 .ptr = ptr, 1138 .offset_first_element_in_bytes = offset_first_element_in_bytes, 1139 .stride_x = stride_x, 1140 .stride_y = stride_y 1141 }; 1142 img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y; 1143 return img; 1144} 1145 1146 1147inline Image update_image_from_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z) 1148{ 1149 Image img = 1150 { 1151 .ptr = ptr, 1152 .offset_first_element_in_bytes = offset_first_element_in_bytes, 1153 .stride_x = stride_x, 1154 .stride_y = stride_y 1155 }; 1156 img.ptr += img.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z; 1157 return img; 1158} 1159 1160 1161inline Tensor3D update_tensor3D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z) 1162{ 1163 Tensor3D tensor = 1164 { 1165 .ptr = ptr, 1166 .offset_first_element_in_bytes = offset_first_element_in_bytes, 1167 .stride_x = stride_x, 1168 .stride_y = stride_y, 1169 .stride_z = stride_z 1170 }; 1171 tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + get_global_id(2) * step_z; 1172 return tensor; 1173} 1174 1175 1176inline Tensor3D tensor3D_ptr_no_update(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z) 1177{ 1178 Tensor3D tensor = 1179 { 1180 .ptr = ptr, 1181 .offset_first_element_in_bytes = offset_first_element_in_bytes, 1182 .stride_x = stride_x, 1183 .stride_y = stride_y, 1184 .stride_z = stride_z 1185 }; 1186 return tensor; 1187} 1188 1189inline Tensor4D update_tensor4D_workitem_ptr(__global uchar *ptr, uint offset_first_element_in_bytes, uint stride_x, uint step_x, uint stride_y, uint step_y, uint stride_z, uint step_z, uint stride_w, 1190 uint step_w, 1191 uint mod_size) 1192{ 1193 Tensor4D tensor = 1194 { 1195 .ptr = ptr, 1196 .offset_first_element_in_bytes = offset_first_element_in_bytes, 1197 .stride_x = stride_x, 1198 .stride_y = stride_y, 1199 .stride_z = stride_z, 1200 .stride_w = stride_w 1201 }; 1202 1203 tensor.ptr += tensor.offset_first_element_in_bytes + get_global_id(0) * step_x + get_global_id(1) * step_y + (get_global_id(2) % mod_size) * step_z + (get_global_id(2) / mod_size) * step_w; 1204 return tensor; 1205} 1206 1207 1208inline __global const uchar *vector_offset(const Vector *vec, int x) 1209{ 1210 return vec->ptr + x * vec->stride_x; 1211} 1212 1213 1214inline __global uchar *offset(const Image *img, int x, int y) 1215{ 1216 return img->ptr + x * img->stride_x + y * img->stride_y; 1217} 1218 1219 1220inline __global const uchar *tensor3D_offset(const Tensor3D *tensor, int x, int y, int z) 1221{ 1222 return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z; 1223} 1224 1225 1226inline __global const uchar *tensor4D_offset(const Tensor4D *tensor, int x, int y, int z, int w) 1227{ 1228 return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + w * tensor->stride_w; 1229} 1230 1231 1232inline __global const uchar *tensor3D_index2ptr(const Tensor3D *tensor, uint width, uint height, uint depth, uint index) 1233{ 1234 uint num_elements = width * height; 1235 1236 const uint z = index / num_elements; 1237 1238 index %= num_elements; 1239 1240 const uint y = index / width; 1241 1242 index %= width; 1243 1244 const uint x = index; 1245 1246 return tensor->ptr + x * tensor->stride_x + y * tensor->stride_y + z * tensor->stride_z + tensor->offset_first_element_in_bytes; 1247} 1248 1249#endif 1250 1251 1252#define CONVERT_DOWN_RTE_STR(x, type) (convert_##type##_rte((x))) 1253#define CONVERT_DOWN_RTE(x, type) CONVERT_DOWN_RTE_STR(x, type) 1254 1255 1256inline uchar quantize_qasymm8(float input, float offset, float scale) 1257{ 1258 float out_f32 = input / scale + offset; 1259 uchar res_u8 = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, int), uchar); 1260 return res_u8; 1261} 1262 1263 1264inline float dequantize_qasymm8(uchar input, float offset, float scale) 1265{ 1266 return ((float)input - offset) * scale; 1267} 1268 1269 1270inline float dequantize_qasymm8_signed(char input, float offset, float scale) 1271{ 1272 return ((float)input - offset) * scale; 1273} 1274 1275 1276#define QUANTIZE_IMPL(type, size) \ 1277 inline VEC_DATA_TYPE(type, size) quantize_##type##size(VEC_DATA_TYPE(float, size) input, float offset, float scale) \ 1278 { \ 1279 VEC_DATA_TYPE(float, size) \ 1280 out_f32 = input / (VEC_DATA_TYPE(float, size))(scale) + (VEC_DATA_TYPE(float, size))(offset); \ 1281 VEC_DATA_TYPE(type, size) \ 1282 res = CONVERT_SAT(CONVERT_DOWN_RTE(out_f32, VEC_DATA_TYPE(int, size)), VEC_DATA_TYPE(type, size)); \ 1283 return res; \ 1284 } 1285 1286 1287#define DEQUANTIZE_IMPL(type, size) \ 1288 inline VEC_DATA_TYPE(float, size) dequantize_##type##size(VEC_DATA_TYPE(type, size) input, float offset, float scale) \ 1289 { \ 1290 return (CONVERT(input, VEC_DATA_TYPE(float, size)) - offset) * scale; \ 1291 } 1292 1293 1294#define ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(size) \ 1295 inline VEC_DATA_TYPE(int, size) asymm_rounding_divide_by_POW2_##size(VEC_DATA_TYPE(int, size) x, VEC_DATA_TYPE(int, size) exponent) \ 1296 { \ 1297 const VEC_DATA_TYPE(int, size) \ 1298 zero = (VEC_DATA_TYPE(int, size))0; \ 1299 const VEC_DATA_TYPE(int, size) \ 1300 one = (VEC_DATA_TYPE(int, size))1; \ 1301 VEC_DATA_TYPE(int, size) \ 1302 mask = (one << exponent) - one; \ 1303 VEC_DATA_TYPE(int, size) \ 1304 threshold = (mask >> 1) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))(x < 0)); \ 1305 return (x >> exponent) + select(zero, one, (SELECT_VEC_DATA_TYPE(int, size))((x & mask) > threshold)); \ 1306 } 1307 1308 1309#define ASYMM_MULT_IMPL(size) \ 1310 inline VEC_DATA_TYPE(int, size) asymm_mult##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \ 1311 { \ 1312 VEC_DATA_TYPE(int, size) \ 1313 overflow = a == b && a == INT_MIN; \ 1314 VEC_DATA_TYPE(long, size) \ 1315 a_64 = convert_long##size(a); \ 1316 VEC_DATA_TYPE(long, size) \ 1317 b_64 = convert_long##size(b); \ 1318 VEC_DATA_TYPE(long, size) \ 1319 ab_64 = a_64 * b_64; \ 1320 \ 1321 VEC_DATA_TYPE(long, size) \ 1322 mask1 = 1 << 30; \ 1323 VEC_DATA_TYPE(long, size) \ 1324 mask2 = 1 - (1 << 30); \ 1325 VEC_DATA_TYPE(long, size) \ 1326 is_positive_or_zero = ab_64 >= 0; \ 1327 VEC_DATA_TYPE(long, size) \ 1328 nudge = select(mask2, mask1, (SELECT_VEC_DATA_TYPE(long, size))(is_positive_or_zero)); \ 1329 VEC_DATA_TYPE(long, size) \ 1330 mask = 1ll << 31; \ 1331 VEC_DATA_TYPE(int, size) \ 1332 ab_x2_high32 = convert_int##size((ab_64 + nudge) / mask); \ 1333 return select(ab_x2_high32, INT_MAX, (SELECT_VEC_DATA_TYPE(int, size))(overflow)); \ 1334 } 1335 1336 1337#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(size) \ 1338 inline VEC_DATA_TYPE(int, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(VEC_DATA_TYPE(int, size) a) \ 1339 { \ 1340 const VEC_DATA_TYPE(int, size) constant_term = 1895147668; \ 1341 const VEC_DATA_TYPE(int, size) constant_1_over_3 = 715827883; \ 1342 const int k_fractional_bits = 31; \ 1343 VEC_DATA_TYPE(int, size) \ 1344 x = a + (1 << (k_fractional_bits - 3)); \ 1345 VEC_DATA_TYPE(int, size) \ 1346 x2 = ASYMM_MULT(x, x, size); \ 1347 VEC_DATA_TYPE(int, size) \ 1348 x3 = ASYMM_MULT(x2, x, size); \ 1349 VEC_DATA_TYPE(int, size) \ 1350 x4 = ASYMM_MULT(x2, x2, size); \ 1351 VEC_DATA_TYPE(int, size) \ 1352 x4_over_4 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4, 2, size); \ 1353 VEC_DATA_TYPE(int, size) \ 1354 x4_over_24_plus_x3_over_6_plus_x2 = ASYMM_MULT((x4_over_4 + x3), constant_1_over_3, size) + x2; \ 1355 VEC_DATA_TYPE(int, size) \ 1356 x4_over_24_plus_x3_over_6_plus_x2_over_2 = ASYMM_ROUNDING_DIVIDE_BY_POW2(x4_over_24_plus_x3_over_6_plus_x2, 1, size); \ 1357 return constant_term + ASYMM_MULT(constant_term, x + x4_over_24_plus_x3_over_6_plus_x2_over_2, size); \ 1358 } 1359 1360 1361#define ASYMM_SELECT_USING_MASK_IMPL(size) \ 1362 inline VEC_DATA_TYPE(int, size) asymm_select_using_mask##size(VEC_DATA_TYPE(int, size) if_mask, VEC_DATA_TYPE(int, size) then_val, VEC_DATA_TYPE(int, size) else_val) \ 1363 { \ 1364 return (if_mask & then_val) ^ (~if_mask & else_val); \ 1365 } 1366 1367 1368#define ASYMM_MASK_IF_ZERO_IMPL(size) \ 1369 inline VEC_DATA_TYPE(int, size) asymm_mask_if_zero##size(VEC_DATA_TYPE(int, size) a) \ 1370 { \ 1371 const VEC_DATA_TYPE(int, size) all_zeros = 0; \ 1372 const VEC_DATA_TYPE(int, size) all_ones = ~0; \ 1373 return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a == 0)); \ 1374 } 1375 1376 1377#define ASYMM_MASK_IF_NON_ZERO_IMPL(size) \ 1378 inline VEC_DATA_TYPE(int, size) asymm_mask_if_non_zero##size(VEC_DATA_TYPE(int, size) a) \ 1379 { \ 1380 const VEC_DATA_TYPE(int, size) all_zeros = 0; \ 1381 const VEC_DATA_TYPE(int, size) all_ones = ~0; \ 1382 return select(all_zeros, all_ones, (SELECT_VEC_DATA_TYPE(int, size))(a != 0)); \ 1383 } 1384 1385#define EXP_BARREL_SHIFTER_IMPL(size) \ 1386 inline VEC_DATA_TYPE(int, size) exp_barrel_shifter##size(VEC_DATA_TYPE(int, size) result, int exponent, int fp_multiplier, int k_integer_bits, int k_fractional_bits, VEC_DATA_TYPE(int, size) remainder) \ 1387 { \ 1388 if(k_integer_bits > exponent) \ 1389 { \ 1390 const int k_shift_amount = k_integer_bits > exponent ? k_fractional_bits + exponent : 0; \ 1391 return ASYMM_SELECT_USING_MASK( \ 1392 ASYMM_MASK_IF_NON_ZERO(remainder & (1 << k_shift_amount), size), \ 1393 ASYMM_MULT(result, fp_multiplier, size), result, size); \ 1394 } \ 1395 \ 1396 return result; \ 1397 } 1398 1399 1400#define ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(size) \ 1401 inline VEC_DATA_TYPE(int, size) asymm_exp_on_negative_values##size(VEC_DATA_TYPE(int, size) a, int k_integer_bits) \ 1402 { \ 1403 const int k_fractional_bits = 31 - k_integer_bits; \ 1404 VEC_DATA_TYPE(int, size) \ 1405 k_one_quarter = 1 << (k_fractional_bits - 2); \ 1406 VEC_DATA_TYPE(int, size) \ 1407 mask = k_one_quarter - 1; \ 1408 VEC_DATA_TYPE(int, size) \ 1409 a_mod_quarter_minus_one_quarter = (a & mask) - k_one_quarter; \ 1410 VEC_DATA_TYPE(int, size) \ 1411 a_mod_quarter_minus_one_quarter_scaled = a_mod_quarter_minus_one_quarter << k_integer_bits; \ 1412 VEC_DATA_TYPE(int, size) \ 1413 result = ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a_mod_quarter_minus_one_quarter_scaled, size); \ 1414 VEC_DATA_TYPE(int, size) \ 1415 remainder = a_mod_quarter_minus_one_quarter - a; \ 1416 \ 1417 result = EXP_BARREL_SHIFTER(result, -2, 1672461947, k_integer_bits, k_fractional_bits, remainder, size); \ 1418 result = EXP_BARREL_SHIFTER(result, -1, 1302514674, k_integer_bits, k_fractional_bits, remainder, size); \ 1419 result = EXP_BARREL_SHIFTER(result, +0, 790015084, k_integer_bits, k_fractional_bits, remainder, size); \ 1420 result = EXP_BARREL_SHIFTER(result, +1, 290630308, k_integer_bits, k_fractional_bits, remainder, size); \ 1421 result = EXP_BARREL_SHIFTER(result, +2, 39332535, k_integer_bits, k_fractional_bits, remainder, size); \ 1422 result = EXP_BARREL_SHIFTER(result, +3, 720401, k_integer_bits, k_fractional_bits, remainder, size); \ 1423 result = EXP_BARREL_SHIFTER(result, +4, 242, k_integer_bits, k_fractional_bits, remainder, size); \ 1424 \ 1425 if(k_integer_bits > 5) \ 1426 { \ 1427 const VEC_DATA_TYPE(int, size) clamp = -(1 << (k_fractional_bits + 5)); \ 1428 result = ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_NON_ZERO(a < clamp, size), 0, result, size); \ 1429 } \ 1430 \ 1431 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \ 1432 return ASYMM_SELECT_USING_MASK(ASYMM_MASK_IF_ZERO(a, size), Q0_one, result, size); \ 1433 } 1434 1435 1436#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(size) \ 1437 inline VEC_DATA_TYPE(int, size) asymm_saturating_rounding_mult_by_pow2##size(VEC_DATA_TYPE(int, size) x, int exponent) \ 1438 { \ 1439 if(exponent < 0) \ 1440 { \ 1441 return ASYMM_ROUNDING_DIVIDE_BY_POW2(x, -exponent, size); \ 1442 } \ 1443 \ 1444 const VEC_DATA_TYPE(int, size) min = INT_MIN; \ 1445 const VEC_DATA_TYPE(int, size) max = INT_MAX; \ 1446 int threshold = ((1 << (31 - exponent)) - 1); \ 1447 VEC_DATA_TYPE(int, size) \ 1448 positive_mask = ASYMM_MASK_IF_NON_ZERO(x > threshold, size); \ 1449 VEC_DATA_TYPE(int, size) \ 1450 negative_mask = ASYMM_MASK_IF_NON_ZERO(x < -threshold, size); \ 1451 VEC_DATA_TYPE(int, size) \ 1452 result = x << exponent; \ 1453 result = ASYMM_SELECT_USING_MASK(positive_mask, max, result, size); \ 1454 result = ASYMM_SELECT_USING_MASK(negative_mask, min, result, size); \ 1455 return result; \ 1456 } 1457 1458 1459#define ASYMM_ROUNDING_HALF_SUM_IMPL(size) \ 1460 inline VEC_DATA_TYPE(int, size) asymm_rounding_half_sum##size(VEC_DATA_TYPE(int, size) a, VEC_DATA_TYPE(int, size) b) \ 1461 { \ 1462 VEC_DATA_TYPE(long, size) \ 1463 a64 = convert_long##size(a); \ 1464 VEC_DATA_TYPE(long, size) \ 1465 b64 = convert_long##size(b); \ 1466 VEC_DATA_TYPE(long, size) \ 1467 sum = a64 + b64; \ 1468 const VEC_DATA_TYPE(long, size) one = 1; \ 1469 const VEC_DATA_TYPE(long, size) minus_one = -1; \ 1470 VEC_DATA_TYPE(long, size) \ 1471 sign = select(minus_one, one, (SELECT_VEC_DATA_TYPE(long, size))(sum >= 0)); \ 1472 return convert_int##size((sum + sign) / 2); \ 1473 } 1474 1475 1476#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(size) \ 1477 inline VEC_DATA_TYPE(int, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(VEC_DATA_TYPE(int, size) a) \ 1478 { \ 1479 const VEC_DATA_TYPE(int, size) Q0_one = INT_MAX; \ 1480 const VEC_DATA_TYPE(int, size) Q2_one = 1 << (31 - 2); \ 1481 VEC_DATA_TYPE(int, size) \ 1482 half_denominator = ASYMM_ROUNDING_HALF_SUM(a, Q0_one, size); \ 1483 const VEC_DATA_TYPE(int, size) Q2_48_over_17 = 1515870810; \ 1484 const VEC_DATA_TYPE(int, size) Q2_neg_32_over_17 = -1010580540; \ 1485 VEC_DATA_TYPE(int, size) \ 1486 x = Q2_48_over_17 + ASYMM_MULT(half_denominator, Q2_neg_32_over_17, size); \ 1487 for(int i = 0; i < 3; i++) \ 1488 { \ 1489 VEC_DATA_TYPE(int, size) \ 1490 half_denominator_times_x = ASYMM_MULT(half_denominator, x, size); \ 1491 VEC_DATA_TYPE(int, size) \ 1492 one_minus_half_denominator_times_x = Q2_one - half_denominator_times_x; \ 1493 VEC_DATA_TYPE(int, size) \ 1494 tmp = ASYMM_MULT(x, one_minus_half_denominator_times_x, size); \ 1495 x = x + ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(tmp, 2, size); \ 1496 } \ 1497 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, 1, size); \ 1498 } 1499 1500 1501#define ASYMM_RESCALE_IMPL(size) \ 1502 inline VEC_DATA_TYPE(int, size) asymm_rescale##size(VEC_DATA_TYPE(int, size) value, int src_integer_bits, int dst_integer_bits) \ 1503 { \ 1504 int exponent = src_integer_bits - dst_integer_bits; \ 1505 return ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(value, exponent, size); \ 1506 } 1507 1508#define QUANTIZE_STR(input, offset, scale, type, size) quantize_##type##size(input, offset, scale) 1509#define QUANTIZE(input, offset, scale, type, size) QUANTIZE_STR(input, offset, scale, type, size) 1510#define DEQUANTIZE_STR(input, offset, scale, type, size) dequantize_##type##size(input, offset, scale) 1511#define DEQUANTIZE(input, offset, scale, type, size) DEQUANTIZE_STR(input, offset, scale, type, size) 1512 1513#define ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size) asymm_rounding_divide_by_POW2_##size(x, exponent) 1514#define ASYMM_ROUNDING_DIVIDE_BY_POW2(x, exponent, size) ASYMM_ROUNDING_DIVIDE_BY_POW2_STR(x, exponent, size) 1515#define ASYMM_MULT_STR(a, b, size) asymm_mult##size(a, b) 1516#define ASYMM_MULT(a, b, size) ASYMM_MULT_STR(a, b, size) 1517#define ASYMM_MULT_BY_QUANT_MULTIPLIER_GREATER_THAN_ONE(x, quantized_multiplier, left_shift, size) \ 1518 ASYMM_MULT(x *((VEC_DATA_TYPE(int, size))(1) << (-left_shift)), quantized_multiplier, size) 1519#define ASYMM_MULT_BY_QUANT_MULTIPLIER_LESS_THAN_ONE(x, quantized_multiplier, right_shift, size) \ 1520 ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(x, quantized_multiplier, size), right_shift, size) 1521#define ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL(a, size) asymm_exp_on_interval_between_negative_one_quarter_and_0_excl##size(a) 1522#define ASYMM_SELECT_USING_MASK(if_mask, then_val, else_val, size) asymm_select_using_mask##size(if_mask, then_val, else_val) 1523#define ASYMM_MASK_IF_ZERO(a, size) asymm_mask_if_zero##size(a) 1524#define ASYMM_MASK_IF_NON_ZERO(a, size) asymm_mask_if_non_zero##size(a) 1525#define EXP_BARREL_SHIFTER(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder, size) exp_barrel_shifter##size(result, exponent, fp_multiplier, k_integer_bits, k_fractional_bits, remainder) 1526#define ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size) asymm_exp_on_negative_values##size(a, k_integer_bits) 1527#define ASYMM_EXP_ON_NEGATIVE_VALUES(a, k_integer_bits, size) ASYMM_EXP_ON_NEGATIVE_VALUES_STR(a, k_integer_bits, size) 1528#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_STR(a, size) asymm_one_over_one_plus_x_for_x_in_0_1##size(a) 1529#define ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1(a, size) ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_STR(a, size) 1530#define ASYMM_SATURATING_ROUNDING_MULT_BY_POW2(x, exponent, size) asymm_saturating_rounding_mult_by_pow2##size(x, exponent) 1531#define ASYMM_ROUNDING_HALF_SUM(a, b, size) asymm_rounding_half_sum##size(a, b) 1532#define ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size) asymm_rescale##size(value, src_integer_bits, dst_integer_bits) 1533#define ASYMM_RESCALE(value, src_integer_bits, dst_integer_bits, size) ASYMM_RESCALE_STR(value, src_integer_bits, dst_integer_bits, size) 1534 1535#define MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(size) \ 1536 inline VEC_DATA_TYPE(int, size) multiply_by_quantized_multiplier##size(VEC_DATA_TYPE(int, size) input, int qmul, int shift) \ 1537 { \ 1538 const int left_shift = shift > 0 ? shift : 0; \ 1539 const int right_shift = shift > 0 ? 0 : -shift; \ 1540 return ASYMM_ROUNDING_DIVIDE_BY_POW2(ASYMM_MULT(input * (1 << left_shift), qmul, size), right_shift, size); \ 1541 } 1542#define MULTIPLY_BY_QUANTIZED_MULTIPLIER(input, qmul, shift, size) multiply_by_quantized_multiplier##size(input, qmul, shift) 1543 1544QUANTIZE_IMPL(uchar, 1) 1545QUANTIZE_IMPL(char, 1) 1546QUANTIZE_IMPL(uint, 1) 1547QUANTIZE_IMPL(int, 1) 1548QUANTIZE_IMPL(uchar, 2) 1549QUANTIZE_IMPL(char, 2) 1550QUANTIZE_IMPL(uint, 2) 1551QUANTIZE_IMPL(int, 2) 1552QUANTIZE_IMPL(uchar, 3) 1553QUANTIZE_IMPL(char, 3) 1554QUANTIZE_IMPL(uint, 3) 1555QUANTIZE_IMPL(int, 3) 1556QUANTIZE_IMPL(uchar, 4) 1557QUANTIZE_IMPL(ushort, 4) 1558QUANTIZE_IMPL(short, 4) 1559QUANTIZE_IMPL(int, 4) 1560QUANTIZE_IMPL(uchar, 8) 1561QUANTIZE_IMPL(char, 8) 1562QUANTIZE_IMPL(uint, 8) 1563QUANTIZE_IMPL(int, 8) 1564QUANTIZE_IMPL(uchar, 16) 1565QUANTIZE_IMPL(char, 16) 1566QUANTIZE_IMPL(ushort, 16) 1567QUANTIZE_IMPL(short, 16) 1568QUANTIZE_IMPL(uint, 16) 1569QUANTIZE_IMPL(int, 16) 1570 1571DEQUANTIZE_IMPL(uchar, 1) 1572DEQUANTIZE_IMPL(char, 1) 1573DEQUANTIZE_IMPL(uint, 1) 1574DEQUANTIZE_IMPL(int, 1) 1575DEQUANTIZE_IMPL(uchar, 2) 1576DEQUANTIZE_IMPL(char, 2) 1577DEQUANTIZE_IMPL(uint, 2) 1578DEQUANTIZE_IMPL(int, 2) 1579DEQUANTIZE_IMPL(uchar, 3) 1580DEQUANTIZE_IMPL(char, 3) 1581DEQUANTIZE_IMPL(uint, 3) 1582DEQUANTIZE_IMPL(int, 3) 1583DEQUANTIZE_IMPL(uchar, 4) 1584DEQUANTIZE_IMPL(ushort, 4) 1585DEQUANTIZE_IMPL(short, 4) 1586DEQUANTIZE_IMPL(int, 4) 1587DEQUANTIZE_IMPL(uchar, 8) 1588DEQUANTIZE_IMPL(char, 8) 1589DEQUANTIZE_IMPL(uint, 8) 1590DEQUANTIZE_IMPL(int, 8) 1591DEQUANTIZE_IMPL(uchar, 16) 1592DEQUANTIZE_IMPL(char, 16) 1593DEQUANTIZE_IMPL(ushort, 16) 1594DEQUANTIZE_IMPL(short, 16) 1595DEQUANTIZE_IMPL(uint, 16) 1596DEQUANTIZE_IMPL(int, 16) 1597 1598ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(1) 1599ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(2) 1600ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(3) 1601ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(4) 1602ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(8) 1603ASYMM_ROUNDING_DIVIDE_BY_POW2_IMPL(16) 1604 1605ASYMM_MULT_IMPL(1) 1606ASYMM_MULT_IMPL(2) 1607ASYMM_MULT_IMPL(3) 1608ASYMM_MULT_IMPL(4) 1609ASYMM_MULT_IMPL(8) 1610ASYMM_MULT_IMPL(16) 1611 1612ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(1) 1613ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(2) 1614ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(3) 1615ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(4) 1616ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(8) 1617ASYMM_EXP_ON_INTERVAL_BETWEEN_NEGATIVE_ONE_QUARTER_AND_0_EXCL_IMPL(16) 1618 1619ASYMM_SELECT_USING_MASK_IMPL(1) 1620ASYMM_SELECT_USING_MASK_IMPL(2) 1621ASYMM_SELECT_USING_MASK_IMPL(3) 1622ASYMM_SELECT_USING_MASK_IMPL(4) 1623ASYMM_SELECT_USING_MASK_IMPL(8) 1624ASYMM_SELECT_USING_MASK_IMPL(16) 1625 1626ASYMM_MASK_IF_ZERO_IMPL(1) 1627ASYMM_MASK_IF_ZERO_IMPL(2) 1628ASYMM_MASK_IF_ZERO_IMPL(3) 1629ASYMM_MASK_IF_ZERO_IMPL(4) 1630ASYMM_MASK_IF_ZERO_IMPL(8) 1631ASYMM_MASK_IF_ZERO_IMPL(16) 1632 1633ASYMM_MASK_IF_NON_ZERO_IMPL(1) 1634ASYMM_MASK_IF_NON_ZERO_IMPL(2) 1635ASYMM_MASK_IF_NON_ZERO_IMPL(3) 1636ASYMM_MASK_IF_NON_ZERO_IMPL(4) 1637ASYMM_MASK_IF_NON_ZERO_IMPL(8) 1638ASYMM_MASK_IF_NON_ZERO_IMPL(16) 1639 1640EXP_BARREL_SHIFTER_IMPL(1) 1641EXP_BARREL_SHIFTER_IMPL(2) 1642EXP_BARREL_SHIFTER_IMPL(3) 1643EXP_BARREL_SHIFTER_IMPL(4) 1644EXP_BARREL_SHIFTER_IMPL(8) 1645EXP_BARREL_SHIFTER_IMPL(16) 1646 1647ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(1) 1648ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(2) 1649ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(3) 1650ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(4) 1651ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(8) 1652ASYMM_EXP_ON_NEGATIVE_VALUES_IMPL(16) 1653 1654ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(1) 1655ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(2) 1656ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(3) 1657ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(4) 1658ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(8) 1659ASYMM_SATURATING_ROUNDING_MULT_BY_POW2_IMPL(16) 1660 1661ASYMM_ROUNDING_HALF_SUM_IMPL(1) 1662ASYMM_ROUNDING_HALF_SUM_IMPL(2) 1663ASYMM_ROUNDING_HALF_SUM_IMPL(3) 1664ASYMM_ROUNDING_HALF_SUM_IMPL(4) 1665ASYMM_ROUNDING_HALF_SUM_IMPL(8) 1666ASYMM_ROUNDING_HALF_SUM_IMPL(16) 1667 1668ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(1) 1669ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(2) 1670ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(3) 1671ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(4) 1672ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(8) 1673ASYMM_ONE_OVER_ONE_PLUS_X_FOR_X_IN_0_1_IMPL(16) 1674 1675ASYMM_RESCALE_IMPL(1) 1676ASYMM_RESCALE_IMPL(2) 1677ASYMM_RESCALE_IMPL(3) 1678ASYMM_RESCALE_IMPL(4) 1679ASYMM_RESCALE_IMPL(8) 1680ASYMM_RESCALE_IMPL(16) 1681 1682MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(1) 1683MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(2) 1684MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(3) 1685MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(4) 1686MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(8) 1687MULTIPLY_BY_QUANTIZED_MULTIPLIER_IMPL(16) 1688 1689#endif 1690 1691#if defined(DATA_TYPE) && defined(DATA_TYPE_DELTAS) && defined(WEIGHT_X) && defined(WEIGHT_Y) && defined(WEIGHT_W) && defined(WEIGHT_H) && defined(IMG_WIDTH) && defined(IMG_HEIGHT) && defined(BOX_FIELDS) && defined(SCALE_BEFORE) && defined(OFFSET_BOXES) && defined(SCALE_BOXES) && defined(OFFSET_DELTAS) && defined(SCALE_DELTAS) && defined(OFFSET_PRED_BOXES) && defined(SCALE_PRED_BOXES) 1692 1693 1694__kernel void bounding_box_transform_quantized( 1695 VECTOR_DECLARATION(boxes), 1696 IMAGE_DECLARATION(pred_boxes), 1697 IMAGE_DECLARATION(deltas)) 1698{ 1699 1700 Vector boxes = CONVERT_TO_VECTOR_STRUCT_NO_STEP(boxes); 1701 Image pred_boxes = CONVERT_TO_IMAGE_STRUCT(pred_boxes); 1702 Image deltas = CONVERT_TO_IMAGE_STRUCT(deltas); 1703 1704 1705 const float one = 1.f; 1706 const float halfone = 0.5f; 1707 1708 const int py = get_global_id(1); 1709 float4 scale_before = (float4)SCALE_BEFORE; 1710 float4 delta = DEQUANTIZE(vload4(0, (__global DATA_TYPE_DELTAS *)deltas.ptr), OFFSET_DELTAS, SCALE_DELTAS, DATA_TYPE_DELTAS, 4); 1711 float4 box = DEQUANTIZE(vload4(0, (__global DATA_TYPE *)vector_offset(&boxes, BOX_FIELDS * py)), OFFSET_BOXES, SCALE_BOXES, DATA_TYPE, 4) / scale_before; 1712 1713 1714 float2 dims = box.s23 - box.s01 + one; 1715 float2 ctr = box.s01 + halfone * dims; 1716 float4 weights = (float4)(WEIGHT_X, WEIGHT_Y, WEIGHT_W, WEIGHT_H); 1717 delta /= weights; 1718 delta.s23 = min(delta.s23, (float)BBOX_XFORM_CLIP); 1719 1720 1721 float2 pred_ctr = delta.s01 * dims + ctr; 1722 float2 pred_dims = exp(delta.s23) * dims; 1723 1724 1725 float4 max_values = (float4)(IMG_WIDTH - 1, IMG_HEIGHT - 1, IMG_WIDTH - 1, IMG_HEIGHT - 1); 1726 float4 sign = (float4)(-1, -1, 1, 1); 1727 float4 min_values = 0; 1728 1729 1730 float4 pred_box = pred_ctr.s0101 + sign * halfone * pred_dims.s0101; 1731#ifdef OFFSET 1732 pred_box.s23 -= one; 1733#endif 1734 pred_box = CLAMP(pred_box, min_values, max_values); 1735#ifdef SCALE_AFTER 1736 pred_box *= (float4)SCALE_AFTER; 1737#endif 1738 1739 1740 vstore4(QUANTIZE(pred_box, OFFSET_PRED_BOXES, SCALE_PRED_BOXES, DATA_TYPE, 4), 0, (__global DATA_TYPE *)pred_boxes.ptr); 1741} 1742#endif )"