| 1 | #define GGML_COMMON_IMPL_C |
| 2 | #include "ggml-common.h" |
| 3 | |
| 4 | #include "ggml-cpu-impl.h" |
| 5 | #include "simd-mappings.h" |
| 6 | #include "ggml-quants.h" |
| 7 | #include "quants.h" |
| 8 | |
| 9 | #include "arch-fallback.h" |
| 10 | |
| 11 | #include <string.h> |
| 12 | #include <assert.h> |
| 13 | #include <float.h> |
| 14 | #include <stdlib.h> // for qsort |
| 15 | #include <stdio.h> // for GGML_ASSERT |
| 16 | |
| 17 | #define GROUP_MAX_EPS 1e-15f |
| 18 | #define GROUP_MAX_EPS_IQ3_XXS 1e-8f |
| 19 | #define GROUP_MAX_EPS_IQ2_S 1e-8f |
| 20 | #define GROUP_MAX_EPS_IQ1_M 1e-7f |
| 21 | #define GROUP_MAX_EPS_IQ1_S 1e-12f |
| 22 | |
| 23 | #define UNUSED GGML_UNUSED |
| 24 | |
| 25 | void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 26 | quantize_row_q4_0_ref(x, y, k); |
| 27 | } |
| 28 | |
| 29 | void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 30 | quantize_row_q4_1_ref(x, y, k); |
| 31 | } |
| 32 | |
| 33 | void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 34 | quantize_row_q5_0_ref(x, y, k); |
| 35 | } |
| 36 | |
| 37 | void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 38 | quantize_row_q5_1_ref(x, y, k); |
| 39 | } |
| 40 | |
| 41 | void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 42 | quantize_row_q8_0_ref(x, y, k); |
| 43 | } |
| 44 | |
| 45 | void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 46 | quantize_row_q8_1_ref(x, y, k); |
| 47 | } |
| 48 | |
| 49 | void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 50 | quantize_row_mxfp4_ref(x, y, k); |
| 51 | } |
| 52 | |
| 53 | // |
| 54 | // 2-6 bit quantization in super-blocks |
| 55 | // |
| 56 | |
| 57 | //========================- 2-bit (de)-quantization |
| 58 | |
| 59 | void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 60 | quantize_row_q2_K_ref(x, y: vy, k); |
| 61 | } |
| 62 | |
| 63 | //========================= 3-bit (de)-quantization |
| 64 | |
| 65 | void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 66 | quantize_row_q3_K_ref(x, y: vy, k); |
| 67 | } |
| 68 | |
| 69 | // ====================== 4-bit (de)-quantization |
| 70 | |
| 71 | void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 72 | assert(k % QK_K == 0); |
| 73 | block_q4_K * GGML_RESTRICT y = vy; |
| 74 | quantize_row_q4_K_ref(x, y, k); |
| 75 | } |
| 76 | |
| 77 | // ====================== 5-bit (de)-quantization |
| 78 | |
| 79 | void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 80 | assert(k % QK_K == 0); |
| 81 | block_q5_K * GGML_RESTRICT y = vy; |
| 82 | quantize_row_q5_K_ref(x, y, k); |
| 83 | } |
| 84 | |
| 85 | // ====================== 6-bit (de)-quantization |
| 86 | |
| 87 | void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 88 | assert(k % QK_K == 0); |
| 89 | block_q6_K * GGML_RESTRICT y = vy; |
| 90 | quantize_row_q6_K_ref(x, y, k); |
| 91 | } |
| 92 | |
| 93 | // ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs) |
| 94 | |
| 95 | void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 96 | assert(k % QK_K == 0); |
| 97 | block_tq1_0 * GGML_RESTRICT y = vy; |
| 98 | quantize_row_tq1_0_ref(x, y, k); |
| 99 | } |
| 100 | |
| 101 | void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) { |
| 102 | assert(k % QK_K == 0); |
| 103 | block_tq2_0 * GGML_RESTRICT y = vy; |
| 104 | quantize_row_tq2_0_ref(x, y, k); |
| 105 | } |
| 106 | |
| 107 | //===================================== Q8_K ============================================== |
| 108 | |
| 109 | void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 110 | quantize_row_q8_K_ref(x, y, k); |
| 111 | } |
| 112 | |
| 113 | //===================================== Dot products ================================= |
| 114 | |
| 115 | void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 116 | const int qk = QK8_0; |
| 117 | const int nb = n / qk; |
| 118 | |
| 119 | assert(n % qk == 0); |
| 120 | assert(nrc == 1); |
| 121 | UNUSED(nrc); |
| 122 | UNUSED(bx); |
| 123 | UNUSED(by); |
| 124 | UNUSED(bs); |
| 125 | |
| 126 | const block_q4_0 * GGML_RESTRICT x = vx; |
| 127 | const block_q8_0 * GGML_RESTRICT y = vy; |
| 128 | |
| 129 | int ib = 0; |
| 130 | float sumf = 0; |
| 131 | |
| 132 | for (; ib < nb; ++ib) { |
| 133 | int sumi0 = 0; |
| 134 | int sumi1 = 0; |
| 135 | |
| 136 | for (int j = 0; j < qk/2; ++j) { |
| 137 | const int v0 = (x[ib].qs[j] & 0x0F) - 8; |
| 138 | const int v1 = (x[ib].qs[j] >> 4) - 8; |
| 139 | |
| 140 | sumi0 += (v0 * y[ib].qs[j]); |
| 141 | sumi1 += (v1 * y[ib].qs[j + qk/2]); |
| 142 | } |
| 143 | |
| 144 | int sumi = sumi0 + sumi1; |
| 145 | sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d); |
| 146 | } |
| 147 | |
| 148 | *s = sumf; |
| 149 | } |
| 150 | |
| 151 | // TODO: add WASM SIMD |
| 152 | void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 153 | const int qk = QK8_1; |
| 154 | const int nb = n / qk; |
| 155 | |
| 156 | assert(n % qk == 0); |
| 157 | assert(nrc == 1); |
| 158 | UNUSED(nrc); |
| 159 | UNUSED(bx); |
| 160 | UNUSED(by); |
| 161 | UNUSED(bs); |
| 162 | |
| 163 | const block_q4_1 * GGML_RESTRICT x = vx; |
| 164 | const block_q8_1 * GGML_RESTRICT y = vy; |
| 165 | |
| 166 | int ib = 0; |
| 167 | float sumf = 0; |
| 168 | |
| 169 | for (; ib < nb; ++ib) { |
| 170 | int sumi0 = 0; |
| 171 | int sumi1 = 0; |
| 172 | |
| 173 | for (int j = 0; j < qk/2; ++j) { |
| 174 | const int v0 = (x[ib].qs[j] & 0x0F); |
| 175 | const int v1 = (x[ib].qs[j] >> 4); |
| 176 | |
| 177 | sumi0 += (v0 * y[ib].qs[j]); |
| 178 | sumi1 += (v1 * y[ib].qs[j + qk/2]); |
| 179 | } |
| 180 | |
| 181 | int sumi = sumi0 + sumi1; |
| 182 | sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); |
| 183 | } |
| 184 | |
| 185 | *s = sumf; |
| 186 | } |
| 187 | |
| 188 | void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 189 | assert(nrc == 1); |
| 190 | UNUSED(nrc); |
| 191 | UNUSED(bx); |
| 192 | UNUSED(by); |
| 193 | UNUSED(bs); |
| 194 | assert(n % QK_MXFP4 == 0); |
| 195 | static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same" ); |
| 196 | |
| 197 | const block_mxfp4 * GGML_RESTRICT x = vx; |
| 198 | const block_q8_0 * GGML_RESTRICT y = vy; |
| 199 | |
| 200 | const int nb = n / QK_MXFP4; |
| 201 | |
| 202 | int ib = 0; |
| 203 | float sumf = 0; |
| 204 | |
| 205 | for (; ib < nb; ++ib) { |
| 206 | const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e); |
| 207 | |
| 208 | int sumi1 = 0; |
| 209 | int sumi2 = 0; |
| 210 | for (int j = 0; j < QK_MXFP4/2; ++j) { |
| 211 | sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf]; |
| 212 | sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4]; |
| 213 | } |
| 214 | sumf += d * (sumi1 + sumi2); |
| 215 | } |
| 216 | *s = sumf; |
| 217 | } |
| 218 | |
| 219 | void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 220 | const int qk = QK8_0; |
| 221 | const int nb = n / qk; |
| 222 | |
| 223 | int ib = 0; |
| 224 | float sumf = 0; |
| 225 | |
| 226 | assert(n % qk == 0); |
| 227 | assert(qk == QK5_0); |
| 228 | assert(nrc == 1); |
| 229 | UNUSED(nrc); |
| 230 | UNUSED(bx); |
| 231 | UNUSED(by); |
| 232 | UNUSED(bs); |
| 233 | |
| 234 | const block_q5_0 * GGML_RESTRICT x = vx; |
| 235 | const block_q8_0 * GGML_RESTRICT y = vy; |
| 236 | |
| 237 | for (; ib < nb; ++ib) { |
| 238 | uint32_t qh; |
| 239 | memcpy(dest: &qh, src: x[ib].qh, n: sizeof(qh)); |
| 240 | |
| 241 | int sumi0 = 0; |
| 242 | int sumi1 = 0; |
| 243 | |
| 244 | for (int j = 0; j < qk/2; ++j) { |
| 245 | const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4; |
| 246 | const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12)); |
| 247 | |
| 248 | const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16); |
| 249 | const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16); |
| 250 | |
| 251 | sumi0 += (x0 * y[ib].qs[j]); |
| 252 | sumi1 += (x1 * y[ib].qs[j + qk/2]); |
| 253 | } |
| 254 | |
| 255 | int sumi = sumi0 + sumi1; |
| 256 | sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi; |
| 257 | } |
| 258 | |
| 259 | *s = sumf; |
| 260 | } |
| 261 | |
| 262 | void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 263 | const int qk = QK8_1; |
| 264 | const int nb = n / qk; |
| 265 | |
| 266 | int ib = 0; |
| 267 | float sumf = 0; |
| 268 | |
| 269 | assert(n % qk == 0); |
| 270 | assert(qk == QK5_1); |
| 271 | assert(nrc == 1); |
| 272 | UNUSED(nrc); |
| 273 | UNUSED(bx); |
| 274 | UNUSED(by); |
| 275 | UNUSED(bs); |
| 276 | |
| 277 | const block_q5_1 * GGML_RESTRICT x = vx; |
| 278 | const block_q8_1 * GGML_RESTRICT y = vy; |
| 279 | |
| 280 | for (; ib < nb; ++ib) { |
| 281 | uint32_t qh; |
| 282 | memcpy(dest: &qh, src: x[ib].qh, n: sizeof(qh)); |
| 283 | |
| 284 | int sumi0 = 0; |
| 285 | int sumi1 = 0; |
| 286 | |
| 287 | for (int j = 0; j < qk/2; ++j) { |
| 288 | const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10; |
| 289 | const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10; |
| 290 | |
| 291 | const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0; |
| 292 | const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1; |
| 293 | |
| 294 | sumi0 += (x0 * y[ib].qs[j]); |
| 295 | sumi1 += (x1 * y[ib].qs[j + qk/2]); |
| 296 | } |
| 297 | |
| 298 | int sumi = sumi0 + sumi1; |
| 299 | sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s); |
| 300 | } |
| 301 | |
| 302 | *s = sumf; |
| 303 | } |
| 304 | |
| 305 | void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 306 | const int qk = QK8_0; |
| 307 | const int nb = n / qk; |
| 308 | |
| 309 | assert(n % qk == 0); |
| 310 | assert(nrc == 1); |
| 311 | UNUSED(nrc); |
| 312 | UNUSED(bx); |
| 313 | UNUSED(by); |
| 314 | UNUSED(bs); |
| 315 | |
| 316 | const block_q8_0 * GGML_RESTRICT x = vx; |
| 317 | const block_q8_0 * GGML_RESTRICT y = vy; |
| 318 | |
| 319 | int ib = 0; |
| 320 | float sumf = 0; |
| 321 | |
| 322 | for (; ib < nb; ++ib) { |
| 323 | int sumi = 0; |
| 324 | |
| 325 | for (int j = 0; j < qk; j++) { |
| 326 | sumi += x[ib].qs[j]*y[ib].qs[j]; |
| 327 | } |
| 328 | |
| 329 | sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)); |
| 330 | } |
| 331 | |
| 332 | *s = sumf; |
| 333 | } |
| 334 | |
| 335 | void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 336 | assert(nrc == 1); |
| 337 | UNUSED(nrc); |
| 338 | UNUSED(bx); |
| 339 | UNUSED(by); |
| 340 | UNUSED(bs); |
| 341 | |
| 342 | const block_tq1_0 * GGML_RESTRICT x = vx; |
| 343 | const block_q8_K * GGML_RESTRICT y = vy; |
| 344 | |
| 345 | const int nb = n / QK_K; |
| 346 | |
| 347 | const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243}; |
| 348 | |
| 349 | float sumf = 0.0f; |
| 350 | |
| 351 | for (int i = 0; i < nb; ++i) { |
| 352 | int sum = 0; |
| 353 | |
| 354 | for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) { |
| 355 | for (size_t l = 0; l < 5; ++l) { |
| 356 | for (size_t m = 0; m < 32; ++m) { |
| 357 | uint8_t q = x[i].qs[j + m] * pow3[l]; |
| 358 | uint16_t xi = ((uint16_t) q * 3) >> 8; |
| 359 | sum += (xi - 1) * y[i].qs[j*5 + l*32 + m]; |
| 360 | } |
| 361 | } |
| 362 | } |
| 363 | for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) { |
| 364 | for (size_t l = 0; l < 5; ++l) { |
| 365 | for (size_t m = 0; m < 16; ++m) { |
| 366 | uint8_t q = x[i].qs[j + m] * pow3[l]; |
| 367 | uint16_t xi = ((uint16_t) q * 3) >> 8; |
| 368 | sum += (xi - 1) * y[i].qs[j*5 + l*16 + m]; |
| 369 | } |
| 370 | } |
| 371 | } |
| 372 | |
| 373 | for (size_t l = 0; l < 4; ++l) { |
| 374 | for (size_t j = 0; j < sizeof(x->qh); ++j) { |
| 375 | uint8_t q = x[i].qh[j] * pow3[l]; |
| 376 | uint16_t xi = ((uint16_t) q * 3) >> 8; |
| 377 | sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j]; |
| 378 | } |
| 379 | } |
| 380 | |
| 381 | sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d); |
| 382 | } |
| 383 | |
| 384 | *s = sumf; |
| 385 | } |
| 386 | |
| 387 | void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 388 | assert(nrc == 1); |
| 389 | UNUSED(nrc); |
| 390 | UNUSED(bx); |
| 391 | UNUSED(by); |
| 392 | UNUSED(bs); |
| 393 | |
| 394 | const block_tq2_0 * GGML_RESTRICT x = vx; |
| 395 | const block_q8_K * GGML_RESTRICT y = vy; |
| 396 | |
| 397 | const int nb = n / QK_K; |
| 398 | float sumf = 0.0f; |
| 399 | |
| 400 | for (int i = 0; i < nb; ++i) { |
| 401 | int32_t sumi = 0; |
| 402 | |
| 403 | for (size_t j = 0; j < sizeof(x->qs); j += 32) { |
| 404 | for (size_t l = 0; l < 4; ++l) { |
| 405 | for (size_t k = 0; k < 32; ++k) { |
| 406 | sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1); |
| 407 | } |
| 408 | } |
| 409 | } |
| 410 | |
| 411 | const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); |
| 412 | |
| 413 | sumf += (float) sumi * d; |
| 414 | } |
| 415 | |
| 416 | *s = sumf; |
| 417 | } |
| 418 | |
| 419 | void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 420 | assert(nrc == 1); |
| 421 | UNUSED(nrc); |
| 422 | UNUSED(bx); |
| 423 | UNUSED(by); |
| 424 | UNUSED(bs); |
| 425 | |
| 426 | const block_q2_K * GGML_RESTRICT x = vx; |
| 427 | const block_q8_K * GGML_RESTRICT y = vy; |
| 428 | |
| 429 | const int nb = n / QK_K; |
| 430 | |
| 431 | float sumf = 0; |
| 432 | |
| 433 | for (int i = 0; i < nb; ++i) { |
| 434 | |
| 435 | const uint8_t * q2 = x[i].qs; |
| 436 | const int8_t * q8 = y[i].qs; |
| 437 | const uint8_t * sc = x[i].scales; |
| 438 | |
| 439 | int summs = 0; |
| 440 | for (int j = 0; j < 16; ++j) { |
| 441 | summs += y[i].bsums[j] * (sc[j] >> 4); |
| 442 | } |
| 443 | |
| 444 | const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d); |
| 445 | const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin); |
| 446 | |
| 447 | int isum = 0; |
| 448 | int is = 0; |
| 449 | int d; |
| 450 | for (int k = 0; k < QK_K/128; ++k) { |
| 451 | int shift = 0; |
| 452 | for (int j = 0; j < 4; ++j) { |
| 453 | d = sc[is++] & 0xF; |
| 454 | int isuml = 0; |
| 455 | for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); |
| 456 | isum += d * isuml; |
| 457 | d = sc[is++] & 0xF; |
| 458 | isuml = 0; |
| 459 | for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3); |
| 460 | isum += d * isuml; |
| 461 | shift += 2; |
| 462 | q8 += 32; |
| 463 | } |
| 464 | q2 += 32; |
| 465 | } |
| 466 | sumf += dall * isum - dmin * summs; |
| 467 | } |
| 468 | *s = sumf; |
| 469 | } |
| 470 | |
| 471 | void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 472 | assert(n % QK_K == 0); |
| 473 | assert(nrc == 1); |
| 474 | UNUSED(nrc); |
| 475 | UNUSED(bx); |
| 476 | UNUSED(by); |
| 477 | UNUSED(bs); |
| 478 | |
| 479 | const uint32_t kmask1 = 0x03030303; |
| 480 | const uint32_t kmask2 = 0x0f0f0f0f; |
| 481 | |
| 482 | const block_q3_K * GGML_RESTRICT x = vx; |
| 483 | const block_q8_K * GGML_RESTRICT y = vy; |
| 484 | |
| 485 | const int nb = n / QK_K; |
| 486 | |
| 487 | // scalar version |
| 488 | // This function is written like this so the compiler can manage to vectorize most of it |
| 489 | // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the |
| 490 | // manually vectorized version above. Every other version I tried would run at least 4 times slower. |
| 491 | // The ideal situation would be if we could just write the code once, and the compiler would |
| 492 | // automatically produce the best possible set of machine instructions, instead of us having to manually |
| 493 | // write vectorized versions for AVX, ARM_NEON, etc. |
| 494 | |
| 495 | int8_t aux8[QK_K]; |
| 496 | int16_t aux16[8]; |
| 497 | float sums [8]; |
| 498 | int32_t aux32[8]; |
| 499 | memset(s: sums, c: 0, n: 8*sizeof(float)); |
| 500 | |
| 501 | uint32_t auxs[4]; |
| 502 | const int8_t * scales = (const int8_t*)auxs; |
| 503 | |
| 504 | float sumf = 0; |
| 505 | for (int i = 0; i < nb; ++i) { |
| 506 | const uint8_t * GGML_RESTRICT q3 = x[i].qs; |
| 507 | const uint8_t * GGML_RESTRICT hm = x[i].hmask; |
| 508 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 509 | memset(s: aux32, c: 0, n: 8*sizeof(int32_t)); |
| 510 | int8_t * GGML_RESTRICT a = aux8; |
| 511 | uint8_t m = 1; |
| 512 | for (int j = 0; j < QK_K; j += 128) { |
| 513 | for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3; |
| 514 | for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); |
| 515 | a += 32; m <<= 1; |
| 516 | for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3; |
| 517 | for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); |
| 518 | a += 32; m <<= 1; |
| 519 | for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3; |
| 520 | for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); |
| 521 | a += 32; m <<= 1; |
| 522 | for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3; |
| 523 | for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4); |
| 524 | a += 32; m <<= 1; |
| 525 | q3 += 32; |
| 526 | } |
| 527 | a = aux8; |
| 528 | |
| 529 | memcpy(dest: auxs, src: x[i].scales, n: 12); |
| 530 | uint32_t tmp = auxs[2]; |
| 531 | auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4); |
| 532 | auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4); |
| 533 | auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4); |
| 534 | auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4); |
| 535 | for (int j = 0; j < QK_K/16; ++j) { |
| 536 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 537 | for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; |
| 538 | q8 += 8; a += 8; |
| 539 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 540 | for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l]; |
| 541 | q8 += 8; a += 8; |
| 542 | } |
| 543 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 544 | for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; |
| 545 | } |
| 546 | for (int l = 0; l < 8; ++l) sumf += sums[l]; |
| 547 | *s = sumf; |
| 548 | } |
| 549 | |
| 550 | void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 551 | assert(n % QK_K == 0); |
| 552 | assert(nrc == 1); |
| 553 | UNUSED(nrc); |
| 554 | UNUSED(bx); |
| 555 | UNUSED(by); |
| 556 | UNUSED(bs); |
| 557 | |
| 558 | const block_q4_K * GGML_RESTRICT x = vx; |
| 559 | const block_q8_K * GGML_RESTRICT y = vy; |
| 560 | |
| 561 | const int nb = n / QK_K; |
| 562 | |
| 563 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 564 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 565 | static const uint32_t kmask3 = 0x03030303; |
| 566 | |
| 567 | uint32_t utmp[4]; |
| 568 | |
| 569 | const uint8_t * scales = (const uint8_t*)&utmp[0]; |
| 570 | const uint8_t * mins = (const uint8_t*)&utmp[2]; |
| 571 | |
| 572 | int8_t aux8[QK_K]; |
| 573 | int16_t aux16[8]; |
| 574 | float sums [8]; |
| 575 | int32_t aux32[8]; |
| 576 | memset(s: sums, c: 0, n: 8*sizeof(float)); |
| 577 | |
| 578 | float sumf = 0; |
| 579 | for (int i = 0; i < nb; ++i) { |
| 580 | const uint8_t * GGML_RESTRICT q4 = x[i].qs; |
| 581 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 582 | memset(s: aux32, c: 0, n: 8*sizeof(int32_t)); |
| 583 | int8_t * GGML_RESTRICT a = aux8; |
| 584 | for (int j = 0; j < QK_K/64; ++j) { |
| 585 | for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); |
| 586 | a += 32; |
| 587 | for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); |
| 588 | a += 32; q4 += 32; |
| 589 | } |
| 590 | memcpy(dest: utmp, src: x[i].scales, n: 12); |
| 591 | utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); |
| 592 | const uint32_t uaux = utmp[1] & kmask1; |
| 593 | utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); |
| 594 | utmp[2] = uaux; |
| 595 | utmp[0] &= kmask1; |
| 596 | |
| 597 | int sumi = 0; |
| 598 | for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; |
| 599 | a = aux8; |
| 600 | int is = 0; |
| 601 | for (int j = 0; j < QK_K/32; ++j) { |
| 602 | int32_t scale = scales[is++]; |
| 603 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 604 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 605 | q8 += 8; a += 8; |
| 606 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 607 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 608 | q8 += 8; a += 8; |
| 609 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 610 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 611 | q8 += 8; a += 8; |
| 612 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 613 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 614 | q8 += 8; a += 8; |
| 615 | } |
| 616 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 617 | for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; |
| 618 | const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; |
| 619 | sumf -= dmin * sumi; |
| 620 | } |
| 621 | for (int l = 0; l < 8; ++l) sumf += sums[l]; |
| 622 | *s = sumf; |
| 623 | } |
| 624 | |
| 625 | void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 626 | assert(n % QK_K == 0); |
| 627 | assert(nrc == 1); |
| 628 | UNUSED(nrc); |
| 629 | UNUSED(bx); |
| 630 | UNUSED(by); |
| 631 | UNUSED(bs); |
| 632 | |
| 633 | const block_q5_K * GGML_RESTRICT x = vx; |
| 634 | const block_q8_K * GGML_RESTRICT y = vy; |
| 635 | |
| 636 | const int nb = n / QK_K; |
| 637 | |
| 638 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 639 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 640 | static const uint32_t kmask3 = 0x03030303; |
| 641 | |
| 642 | uint32_t utmp[4]; |
| 643 | |
| 644 | const uint8_t * scales = (const uint8_t*)&utmp[0]; |
| 645 | const uint8_t * mins = (const uint8_t*)&utmp[2]; |
| 646 | |
| 647 | int8_t aux8[QK_K]; |
| 648 | int16_t aux16[8]; |
| 649 | float sums [8]; |
| 650 | int32_t aux32[8]; |
| 651 | memset(s: sums, c: 0, n: 8*sizeof(float)); |
| 652 | |
| 653 | float sumf = 0; |
| 654 | for (int i = 0; i < nb; ++i) { |
| 655 | const uint8_t * GGML_RESTRICT q4 = x[i].qs; |
| 656 | const uint8_t * GGML_RESTRICT hm = x[i].qh; |
| 657 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 658 | memset(s: aux32, c: 0, n: 8*sizeof(int32_t)); |
| 659 | int8_t * GGML_RESTRICT a = aux8; |
| 660 | uint8_t m = 1; |
| 661 | for (int j = 0; j < QK_K/64; ++j) { |
| 662 | for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF); |
| 663 | for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); |
| 664 | a += 32; m <<= 1; |
| 665 | for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4); |
| 666 | for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0); |
| 667 | a += 32; m <<= 1; |
| 668 | q4 += 32; |
| 669 | } |
| 670 | memcpy(dest: utmp, src: x[i].scales, n: 12); |
| 671 | utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4); |
| 672 | const uint32_t uaux = utmp[1] & kmask1; |
| 673 | utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4); |
| 674 | utmp[2] = uaux; |
| 675 | utmp[0] &= kmask1; |
| 676 | |
| 677 | int sumi = 0; |
| 678 | for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2]; |
| 679 | a = aux8; |
| 680 | int is = 0; |
| 681 | for (int j = 0; j < QK_K/32; ++j) { |
| 682 | int32_t scale = scales[is++]; |
| 683 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 684 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 685 | q8 += 8; a += 8; |
| 686 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 687 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 688 | q8 += 8; a += 8; |
| 689 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 690 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 691 | q8 += 8; a += 8; |
| 692 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 693 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 694 | q8 += 8; a += 8; |
| 695 | } |
| 696 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 697 | for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; |
| 698 | const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d; |
| 699 | sumf -= dmin * sumi; |
| 700 | } |
| 701 | for (int l = 0; l < 8; ++l) sumf += sums[l]; |
| 702 | *s = sumf; |
| 703 | } |
| 704 | |
| 705 | void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 706 | assert(n % QK_K == 0); |
| 707 | assert(nrc == 1); |
| 708 | UNUSED(nrc); |
| 709 | UNUSED(bx); |
| 710 | UNUSED(by); |
| 711 | UNUSED(bs); |
| 712 | |
| 713 | const block_q6_K * GGML_RESTRICT x = vx; |
| 714 | const block_q8_K * GGML_RESTRICT y = vy; |
| 715 | |
| 716 | const int nb = n / QK_K; |
| 717 | |
| 718 | int8_t aux8[QK_K]; |
| 719 | int16_t aux16[8]; |
| 720 | float sums [8]; |
| 721 | int32_t aux32[8]; |
| 722 | memset(s: sums, c: 0, n: 8*sizeof(float)); |
| 723 | |
| 724 | float sumf = 0; |
| 725 | for (int i = 0; i < nb; ++i) { |
| 726 | const uint8_t * GGML_RESTRICT q4 = x[i].ql; |
| 727 | const uint8_t * GGML_RESTRICT qh = x[i].qh; |
| 728 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 729 | memset(s: aux32, c: 0, n: 8*sizeof(int32_t)); |
| 730 | int8_t * GGML_RESTRICT a = aux8; |
| 731 | for (int j = 0; j < QK_K; j += 128) { |
| 732 | for (int l = 0; l < 32; ++l) { |
| 733 | a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32; |
| 734 | a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32; |
| 735 | a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32; |
| 736 | a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32; |
| 737 | } |
| 738 | a += 128; |
| 739 | q4 += 64; |
| 740 | qh += 32; |
| 741 | } |
| 742 | a = aux8; |
| 743 | int is = 0; |
| 744 | for (int j = 0; j < QK_K/16; ++j) { |
| 745 | int scale = x[i].scales[is++]; |
| 746 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 747 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 748 | q8 += 8; a += 8; |
| 749 | for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l]; |
| 750 | for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l]; |
| 751 | q8 += 8; a += 8; |
| 752 | } |
| 753 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 754 | for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l]; |
| 755 | } |
| 756 | for (int l = 0; l < 8; ++l) sumf += sums[l]; |
| 757 | *s = sumf; |
| 758 | } |
| 759 | |
| 760 | void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 761 | assert(n % QK_K == 0); |
| 762 | assert(nrc == 1); |
| 763 | UNUSED(nrc); |
| 764 | UNUSED(bx); |
| 765 | UNUSED(by); |
| 766 | UNUSED(bs); |
| 767 | |
| 768 | const block_iq2_xxs * GGML_RESTRICT x = vx; |
| 769 | const block_q8_K * GGML_RESTRICT y = vy; |
| 770 | |
| 771 | const int nb = n / QK_K; |
| 772 | |
| 773 | uint32_t aux32[2]; |
| 774 | const uint8_t * aux8 = (const uint8_t *)aux32; |
| 775 | |
| 776 | float sumf = 0.f; |
| 777 | for (int i = 0; i < nb; ++i) { |
| 778 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 779 | const uint16_t * GGML_RESTRICT q2 = x[i].qs; |
| 780 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 781 | int32_t bsum = 0; |
| 782 | for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { |
| 783 | memcpy(dest: aux32, src: q2, n: 2*sizeof(uint32_t)); |
| 784 | q2 += 4; |
| 785 | const uint32_t ls = 2*(aux32[1] >> 28) + 1; |
| 786 | int32_t sumi = 0; |
| 787 | for (int l = 0; l < 4; ++l) { |
| 788 | const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]); |
| 789 | const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127]; |
| 790 | for (int j = 0; j < 8; ++j) { |
| 791 | sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); |
| 792 | } |
| 793 | q8 += 8; |
| 794 | } |
| 795 | bsum += sumi * ls; |
| 796 | } |
| 797 | sumf += d * bsum; |
| 798 | } |
| 799 | *s = 0.125f * sumf; |
| 800 | } |
| 801 | |
| 802 | void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 803 | assert(n % QK_K == 0); |
| 804 | assert(nrc == 1); |
| 805 | UNUSED(nrc); |
| 806 | UNUSED(bx); |
| 807 | UNUSED(by); |
| 808 | UNUSED(bs); |
| 809 | |
| 810 | const block_iq2_xs * GGML_RESTRICT x = vx; |
| 811 | const block_q8_K * GGML_RESTRICT y = vy; |
| 812 | |
| 813 | const int nb = n / QK_K; |
| 814 | |
| 815 | float sumf = 0.f; |
| 816 | for (int i = 0; i < nb; ++i) { |
| 817 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 818 | const uint16_t * GGML_RESTRICT q2 = x[i].qs; |
| 819 | const uint8_t * GGML_RESTRICT sc = x[i].scales; |
| 820 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 821 | int32_t bsum = 0; |
| 822 | for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { |
| 823 | const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1; |
| 824 | const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1; |
| 825 | int32_t sumi = 0; |
| 826 | for (int l = 0; l < 2; ++l) { |
| 827 | const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); |
| 828 | const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; |
| 829 | for (int j = 0; j < 8; ++j) { |
| 830 | sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); |
| 831 | } |
| 832 | q8 += 8; |
| 833 | } |
| 834 | bsum += sumi * ls1; |
| 835 | sumi = 0; |
| 836 | for (int l = 2; l < 4; ++l) { |
| 837 | const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511)); |
| 838 | const uint8_t signs = ksigns_iq2xs[q2[l] >> 9]; |
| 839 | for (int j = 0; j < 8; ++j) { |
| 840 | sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1); |
| 841 | } |
| 842 | q8 += 8; |
| 843 | } |
| 844 | bsum += sumi * ls2; |
| 845 | q2 += 4; |
| 846 | } |
| 847 | sumf += d * bsum; |
| 848 | } |
| 849 | *s = 0.125f * sumf; |
| 850 | } |
| 851 | |
| 852 | void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 853 | assert(n % QK_K == 0); |
| 854 | assert(nrc == 1); |
| 855 | UNUSED(nrc); |
| 856 | UNUSED(bx); |
| 857 | UNUSED(by); |
| 858 | UNUSED(bs); |
| 859 | |
| 860 | const block_iq2_s * GGML_RESTRICT x = vx; |
| 861 | const block_q8_K * GGML_RESTRICT y = vy; |
| 862 | |
| 863 | const int nb = n / QK_K; |
| 864 | |
| 865 | float sumf = 0; |
| 866 | for (int i = 0; i < nb; i++) { |
| 867 | |
| 868 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 869 | const int8_t * q8 = y[i].qs; |
| 870 | const uint8_t * qs = x[i].qs; |
| 871 | const uint8_t * qh = x[i].qh; |
| 872 | const uint8_t * signs = qs + QK_K/8; |
| 873 | |
| 874 | int bsum = 0; |
| 875 | for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { |
| 876 | int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf); |
| 877 | int ls2 = 1 + 2*(x[i].scales[ib32] >> 4); |
| 878 | int sumi1 = 0, sumi2 = 0; |
| 879 | for (int l = 0; l < 2; ++l) { |
| 880 | const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); |
| 881 | for (int j = 0; j < 8; ++j) { |
| 882 | sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); |
| 883 | } |
| 884 | q8 += 8; |
| 885 | } |
| 886 | for (int l = 2; l < 4; ++l) { |
| 887 | const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300))); |
| 888 | for (int j = 0; j < 8; ++j) { |
| 889 | sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1); |
| 890 | } |
| 891 | q8 += 8; |
| 892 | } |
| 893 | bsum += ls1 * sumi1 + ls2 * sumi2; |
| 894 | qs += 4; |
| 895 | signs += 4; |
| 896 | } |
| 897 | |
| 898 | sumf += d * bsum; |
| 899 | } |
| 900 | |
| 901 | *s = 0.125f * sumf; |
| 902 | } |
| 903 | |
| 904 | void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 905 | assert(n % QK_K == 0); |
| 906 | assert(nrc == 1); |
| 907 | UNUSED(nrc); |
| 908 | UNUSED(bx); |
| 909 | UNUSED(by); |
| 910 | UNUSED(bs); |
| 911 | |
| 912 | const block_iq3_xxs * GGML_RESTRICT x = vx; |
| 913 | const block_q8_K * GGML_RESTRICT y = vy; |
| 914 | |
| 915 | const int nb = n / QK_K; |
| 916 | |
| 917 | uint32_t aux32; |
| 918 | |
| 919 | float sumf = 0.f; |
| 920 | for (int i = 0; i < nb; ++i) { |
| 921 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 922 | const uint8_t * GGML_RESTRICT q3 = x[i].qs; |
| 923 | const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4; |
| 924 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 925 | int32_t bsum = 0; |
| 926 | for (int ib32 = 0; ib32 < QK_K/32; ++ib32) { |
| 927 | memcpy(dest: &aux32, src: gas, n: sizeof(uint32_t)); gas += sizeof(uint32_t); |
| 928 | const uint32_t ls = 2*(aux32 >> 28) + 1; |
| 929 | int32_t sumi = 0; |
| 930 | for (int l = 0; l < 4; ++l) { |
| 931 | const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]); |
| 932 | const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]); |
| 933 | const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127]; |
| 934 | for (int j = 0; j < 4; ++j) { |
| 935 | sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1); |
| 936 | sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1); |
| 937 | } |
| 938 | q8 += 8; |
| 939 | } |
| 940 | q3 += 8; |
| 941 | bsum += sumi * ls; |
| 942 | } |
| 943 | sumf += d * bsum; |
| 944 | } |
| 945 | *s = 0.25f * sumf; |
| 946 | } |
| 947 | |
| 948 | void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 949 | assert(n % QK_K == 0); |
| 950 | assert(nrc == 1); |
| 951 | UNUSED(nrc); |
| 952 | UNUSED(bx); |
| 953 | UNUSED(by); |
| 954 | UNUSED(bs); |
| 955 | |
| 956 | const block_iq3_s * GGML_RESTRICT x = vx; |
| 957 | const block_q8_K * GGML_RESTRICT y = vy; |
| 958 | |
| 959 | const int nb = n / QK_K; |
| 960 | |
| 961 | float sumf = 0.f; |
| 962 | for (int i = 0; i < nb; ++i) { |
| 963 | const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d; |
| 964 | const uint8_t * GGML_RESTRICT qs = x[i].qs; |
| 965 | const uint8_t * GGML_RESTRICT qh = x[i].qh; |
| 966 | const uint8_t * GGML_RESTRICT signs = x[i].signs; |
| 967 | const int8_t * GGML_RESTRICT q8 = y[i].qs; |
| 968 | int32_t bsum = 0; |
| 969 | for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) { |
| 970 | const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1; |
| 971 | const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1; |
| 972 | int32_t sumi = 0; |
| 973 | for (int l = 0; l < 4; ++l) { |
| 974 | const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256))); |
| 975 | const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256))); |
| 976 | for (int j = 0; j < 4; ++j) { |
| 977 | sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); |
| 978 | sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); |
| 979 | } |
| 980 | q8 += 8; |
| 981 | } |
| 982 | qs += 8; |
| 983 | signs += 4; |
| 984 | bsum += sumi * ls1; |
| 985 | sumi = 0; |
| 986 | for (int l = 0; l < 4; ++l) { |
| 987 | const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256))); |
| 988 | const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256))); |
| 989 | for (int j = 0; j < 4; ++j) { |
| 990 | sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1); |
| 991 | sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1); |
| 992 | } |
| 993 | q8 += 8; |
| 994 | } |
| 995 | qs += 8; |
| 996 | signs += 4; |
| 997 | bsum += sumi * ls2; |
| 998 | } |
| 999 | sumf += d * bsum; |
| 1000 | } |
| 1001 | *s = sumf; |
| 1002 | } |
| 1003 | |
| 1004 | void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 1005 | assert(n % QK_K == 0); |
| 1006 | assert(nrc == 1); |
| 1007 | UNUSED(nrc); |
| 1008 | UNUSED(bx); |
| 1009 | UNUSED(by); |
| 1010 | UNUSED(bs); |
| 1011 | |
| 1012 | const block_iq1_s * GGML_RESTRICT x = vx; |
| 1013 | const block_q8_K * GGML_RESTRICT y = vy; |
| 1014 | |
| 1015 | const int nb = n / QK_K; |
| 1016 | |
| 1017 | float sumf = 0; |
| 1018 | for (int i = 0; i < nb; i++) { |
| 1019 | |
| 1020 | const int8_t * q8 = y[i].qs; |
| 1021 | const uint8_t * qs = x[i].qs; |
| 1022 | const uint16_t * qh = x[i].qh; |
| 1023 | |
| 1024 | int sumi = 0, sumi1 = 0; |
| 1025 | for (int ib = 0; ib < QK_K/32; ++ib) { |
| 1026 | const int ls = 2*((qh[ib] >> 12) & 7) + 1; |
| 1027 | const int delta = qh[ib] & 0x8000 ? -1 : 1; |
| 1028 | int lsum = 0; |
| 1029 | for (int l = 0; l < 4; ++l) { |
| 1030 | const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8))); |
| 1031 | for (int j = 0; j < 8; ++j) { |
| 1032 | lsum += q8[j] * grid[j]; |
| 1033 | } |
| 1034 | q8 += 8; |
| 1035 | } |
| 1036 | sumi += ls * lsum; |
| 1037 | sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]); |
| 1038 | qs += 4; |
| 1039 | } |
| 1040 | |
| 1041 | sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1); |
| 1042 | } |
| 1043 | |
| 1044 | *s = sumf; |
| 1045 | } |
| 1046 | |
| 1047 | void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 1048 | assert(n % QK_K == 0); |
| 1049 | assert(nrc == 1); |
| 1050 | UNUSED(nrc); |
| 1051 | UNUSED(bx); |
| 1052 | UNUSED(by); |
| 1053 | UNUSED(bs); |
| 1054 | |
| 1055 | const block_iq1_m * GGML_RESTRICT x = vx; |
| 1056 | const block_q8_K * GGML_RESTRICT y = vy; |
| 1057 | |
| 1058 | const int nb = n / QK_K; |
| 1059 | |
| 1060 | iq1m_scale_t scale; |
| 1061 | |
| 1062 | int sum1[2], sum2[2], delta[4]; |
| 1063 | |
| 1064 | float sumf = 0; |
| 1065 | for (int i = 0; i < nb; i++) { |
| 1066 | |
| 1067 | const int8_t * q8 = y[i].qs; |
| 1068 | const uint8_t * qs = x[i].qs; |
| 1069 | const uint8_t * qh = x[i].qh; |
| 1070 | const uint16_t * sc = (const uint16_t *)x[i].scales; |
| 1071 | |
| 1072 | scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000); |
| 1073 | |
| 1074 | int sumi1 = 0, sumi2 = 0; |
| 1075 | for (int ib = 0; ib < QK_K/32; ++ib) { |
| 1076 | delta[0] = qh[0] & 0x08 ? -1 : 1; |
| 1077 | delta[1] = qh[0] & 0x80 ? -1 : 1; |
| 1078 | delta[2] = qh[1] & 0x08 ? -1 : 1; |
| 1079 | delta[3] = qh[1] & 0x80 ? -1 : 1; |
| 1080 | sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0; |
| 1081 | for (int l = 0; l < 4; ++l) { |
| 1082 | const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700))); |
| 1083 | int lsum1 = 0, lsum2 = 0; |
| 1084 | for (int j = 0; j < 8; ++j) { |
| 1085 | lsum1 += q8[j] * grid[j]; |
| 1086 | lsum2 += q8[j]; |
| 1087 | } |
| 1088 | q8 += 8; |
| 1089 | sum1[l/2] += lsum1; |
| 1090 | sum2[l/2] += lsum2*delta[l]; |
| 1091 | } |
| 1092 | |
| 1093 | const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1; |
| 1094 | const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1; |
| 1095 | |
| 1096 | sumi1 += sum1[0] * ls1 + sum1[1] * ls2; |
| 1097 | sumi2 += sum2[0] * ls1 + sum2[1] * ls2; |
| 1098 | qs += 4; |
| 1099 | qh += 2; |
| 1100 | } |
| 1101 | |
| 1102 | sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2); |
| 1103 | } |
| 1104 | |
| 1105 | *s = sumf; |
| 1106 | } |
| 1107 | |
| 1108 | void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 1109 | assert(nrc == 1); |
| 1110 | UNUSED(nrc); |
| 1111 | UNUSED(bx); |
| 1112 | UNUSED(by); |
| 1113 | UNUSED(bs); |
| 1114 | assert(n % QK4_NL == 0); |
| 1115 | static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same" ); |
| 1116 | |
| 1117 | const block_iq4_nl * GGML_RESTRICT x = vx; |
| 1118 | const block_q8_0 * GGML_RESTRICT y = vy; |
| 1119 | |
| 1120 | const int nb = n / QK4_NL; |
| 1121 | |
| 1122 | int ib = 0; |
| 1123 | float sumf = 0; |
| 1124 | |
| 1125 | for (; ib < nb; ++ib) { |
| 1126 | const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d); |
| 1127 | int sumi1 = 0, sumi2 = 0; |
| 1128 | for (int j = 0; j < QK4_NL/2; ++j) { |
| 1129 | sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf]; |
| 1130 | sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4]; |
| 1131 | } |
| 1132 | sumf += d * (sumi1 + sumi2); |
| 1133 | } |
| 1134 | *s = sumf; |
| 1135 | } |
| 1136 | |
| 1137 | void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) { |
| 1138 | assert(nrc == 1); |
| 1139 | UNUSED(nrc); |
| 1140 | UNUSED(bx); |
| 1141 | UNUSED(by); |
| 1142 | UNUSED(bs); |
| 1143 | assert(n % QK_K == 0); |
| 1144 | |
| 1145 | const block_iq4_xs * GGML_RESTRICT x = vx; |
| 1146 | const block_q8_K * GGML_RESTRICT y = vy; |
| 1147 | |
| 1148 | const int nb = n / QK_K; |
| 1149 | |
| 1150 | float sumf = 0; |
| 1151 | for (int ibl = 0; ibl < nb; ++ibl) { |
| 1152 | const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d; |
| 1153 | uint16_t h = x[ibl].scales_h; |
| 1154 | const uint8_t * qs = x[ibl].qs; |
| 1155 | const int8_t * q8 = y[ibl].qs; |
| 1156 | for (int ib = 0; ib < QK_K/32; ib += 2) { |
| 1157 | const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30); |
| 1158 | const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30); |
| 1159 | h >>= 4; |
| 1160 | const float d1 = d4d8*(ls1 - 32); |
| 1161 | const float d2 = d4d8*(ls2 - 32); |
| 1162 | int sumi1 = 0, sumi2 = 0; |
| 1163 | for (int j = 0; j < 16; ++j) { |
| 1164 | sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; |
| 1165 | sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; |
| 1166 | } |
| 1167 | sumf += d1 * (sumi1 + sumi2); |
| 1168 | qs += 16; |
| 1169 | q8 += 32; |
| 1170 | sumi1 = sumi2 = 0; |
| 1171 | for (int j = 0; j < 16; ++j) { |
| 1172 | sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf]; |
| 1173 | sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4]; |
| 1174 | } |
| 1175 | sumf += d2 * (sumi1 + sumi2); |
| 1176 | qs += 16; |
| 1177 | q8 += 32; |
| 1178 | } |
| 1179 | } |
| 1180 | *s = sumf; |
| 1181 | } |
| 1182 | |
| 1183 | // ============================ 4-bit non-linear quants |
| 1184 | |
| 1185 | void quantize_row_iq4_nl(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 1186 | assert(k % QK4_NL == 0); |
| 1187 | quantize_row_iq4_nl_ref(x, y, k); |
| 1188 | } |
| 1189 | |
| 1190 | void quantize_row_iq4_xs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) { |
| 1191 | assert(k % QK_K == 0); |
| 1192 | quantize_iq4_xs(src: x, dst: y, nrows: 1, n_per_row: k, NULL); |
| 1193 | } |
| 1194 | |