| File: | root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp |
| Warning: | line 931, column 21 Value stored to 'sumi2' is never read |
Press '?' to see keyboard shortcuts
Keyboard shortcuts:
| 1 | #define GGML_COMMON_IMPL_CPP |
| 2 | #define GGML_COMMON_DECL_CPP |
| 3 | #include "ggml-common.h" |
| 4 | #include "ggml-backend-impl.h" |
| 5 | |
| 6 | #include "ggml-impl.h" |
| 7 | #include "ggml-cpu.h" |
| 8 | #include "ggml-cpu-impl.h" |
| 9 | #include "simd-mappings.h" |
| 10 | #include "traits.h" |
| 11 | |
| 12 | #include "arch-fallback.h" |
| 13 | |
| 14 | #include <cmath> |
| 15 | #include <cstring> |
| 16 | #include <cassert> |
| 17 | #include <cstdio> // for GGML_ASSERT |
| 18 | |
| 19 | #include "repack.h" |
| 20 | |
| 21 | #if defined(__GNUC__4) |
| 22 | #pragma GCC diagnostic ignored "-Woverlength-strings" |
| 23 | #endif |
| 24 | |
| 25 | #define UNUSEDGGML_UNUSED GGML_UNUSED |
| 26 | |
| 27 | static inline int nearest_int(float fval) { |
| 28 | assert(fabsf(fval) <= 4194303.f)(static_cast <bool> (fabsf(fval) <= 4194303.f) ? void (0) : __assert_fail ("fabsf(fval) <= 4194303.f", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 29 | float val = fval + 12582912.f; |
| 30 | int i; memcpy(&i, &val, sizeof(int)); |
| 31 | return (i & 0x007fffff) - 0x00400000; |
| 32 | } |
| 33 | |
| 34 | // Functions to create the interleaved data layout formats |
| 35 | |
| 36 | // interleave 4 block_q4_0s in blocks of blck_size_interleave |
| 37 | // returns an interleaved block_q4_0x4 |
| 38 | // in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks |
| 39 | // first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave |
| 40 | // |
| 41 | // - in : an array of block_q4_0 pointers |
| 42 | // - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of |
| 43 | // blck_size_interleave bytes |
| 44 | // - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes |
| 45 | // from bias offset form to pure sign form (this saves subtract |
| 46 | // operations durin unpacking) |
| 47 | // |
| 48 | |
| 49 | extern "C" { |
| 50 | |
| 51 | #if defined __riscv_zvfh |
| 52 | void ggml_quantize_mat_q8_0_4x1_generic(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t k) { |
| 53 | assert(QK8_0 == 32)(static_cast <bool> (32 == 32) ? void (0) : __assert_fail ("QK8_0 == 32", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 54 | assert(k % QK8_0 == 0)(static_cast <bool> (k % 32 == 0) ? void (0) : __assert_fail ("k % QK8_0 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 55 | const int nb = k / QK8_032; |
| 56 | |
| 57 | block_q8_0x4 * GGML_RESTRICT__restrict__ y = (block_q8_0x4 *) vy; |
| 58 | |
| 59 | // scalar |
| 60 | const int blck_size_interleave = 1; |
| 61 | float srcv[4][QK8_032]; |
| 62 | float id[4]; |
| 63 | |
| 64 | for (int i = 0; i < nb; i++) { |
| 65 | for (int row_iter = 0; row_iter < 4; row_iter++) { |
| 66 | float amax = 0.0f; // absolute max |
| 67 | |
| 68 | for (int j = 0; j < QK8_032; j++) { |
| 69 | srcv[row_iter][j] = x[row_iter * k + i * QK8_032 + j]; |
| 70 | amax = MAX(amax, fabsf(srcv[row_iter][j]))((amax) > (fabsf(srcv[row_iter][j])) ? (amax) : (fabsf(srcv [row_iter][j]))); |
| 71 | } |
| 72 | |
| 73 | const float d = amax / ((1 << 7) - 1); |
| 74 | id[row_iter] = d ? 1.0f / d : 0.0f; |
| 75 | |
| 76 | y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d)ggml_compute_fp32_to_fp16(d); |
| 77 | } |
| 78 | |
| 79 | for (int j = 0; j < QK8_032 * 4; j++) { |
| 80 | int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; |
| 81 | int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; |
| 82 | src_offset += (j % blck_size_interleave); |
| 83 | |
| 84 | float x0 = srcv[src_id][src_offset] * id[src_id]; |
| 85 | y[i].qs[j] = roundf(x0); |
| 86 | } |
| 87 | } |
| 88 | } |
| 89 | |
| 90 | void ggml_quantize_mat_q8_K_4x1_generic(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t k) { |
| 91 | assert(QK_K == 256)(static_cast <bool> (256 == 256) ? void (0) : __assert_fail ("QK_K == 256", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 92 | assert(k % QK_K == 0)(static_cast <bool> (k % 256 == 0) ? void (0) : __assert_fail ("k % QK_K == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 93 | const int nb = k / QK_K256; |
| 94 | |
| 95 | block_q8_Kx4 * GGML_RESTRICT__restrict__ y = (block_q8_Kx4 *) vy; |
| 96 | |
| 97 | const int blck_size_interleave = 1; |
| 98 | float srcv[4][QK_K256]; |
| 99 | float iscale[4]; |
| 100 | |
| 101 | for (int i = 0; i < nb; i++) { |
| 102 | for (int row_iter = 0; row_iter < 4; row_iter++) { |
| 103 | float amax = 0.0f; // absolute max |
| 104 | float max = 0; |
| 105 | |
| 106 | for (int j = 0; j < QK_K256; j++) { |
| 107 | srcv[row_iter][j] = x[row_iter * k + i * QK_K256 + j]; |
| 108 | // Update the maximum value of the corresponding super block |
| 109 | if(amax < fabsf(srcv[row_iter][j])) { |
| 110 | amax = fabsf(srcv[row_iter][j]); |
| 111 | max = srcv[row_iter][j]; |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | iscale[row_iter] = amax ? -127.f/max : 0; |
| 116 | y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; |
| 117 | } |
| 118 | |
| 119 | for (int j = 0; j < QK_K256 / 4; j++) { |
| 120 | y[i].bsums[j] = 0; |
| 121 | } |
| 122 | for (int j = 0; j < QK_K256 * 4; j++) { |
| 123 | int src_id = j % 4; |
| 124 | int src_offset = j / 4; |
| 125 | int index = ((j >> 6) << 2) + (j & 3); |
| 126 | |
| 127 | float x0 = srcv[src_id][src_offset] * iscale[src_id]; |
| 128 | y[i].qs[j] = nearest_int(x0); |
| 129 | y[i].bsums[index] += y[i].qs[j]; |
| 130 | } |
| 131 | } |
| 132 | } |
| 133 | #endif |
| 134 | |
| 135 | void ggml_quantize_mat_q8_0_4x4_genericggml_quantize_mat_q8_0_4x4(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t k) { |
| 136 | assert(QK8_0 == 32)(static_cast <bool> (32 == 32) ? void (0) : __assert_fail ("QK8_0 == 32", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 137 | assert(k % QK8_0 == 0)(static_cast <bool> (k % 32 == 0) ? void (0) : __assert_fail ("k % QK8_0 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 138 | const int nb = k / QK8_032; |
| 139 | |
| 140 | block_q8_0x4 * GGML_RESTRICT__restrict__ y = (block_q8_0x4 *) vy; |
| 141 | |
| 142 | // scalar |
| 143 | const int blck_size_interleave = 4; |
| 144 | float srcv[4][QK8_032]; |
| 145 | float id[4]; |
| 146 | |
| 147 | for (int i = 0; i < nb; i++) { |
| 148 | for (int row_iter = 0; row_iter < 4; row_iter++) { |
| 149 | float amax = 0.0f; // absolute max |
| 150 | |
| 151 | for (int j = 0; j < QK8_032; j++) { |
| 152 | srcv[row_iter][j] = x[row_iter * k + i * QK8_032 + j]; |
| 153 | amax = MAX(amax, fabsf(srcv[row_iter][j]))((amax) > (fabsf(srcv[row_iter][j])) ? (amax) : (fabsf(srcv [row_iter][j]))); |
| 154 | } |
| 155 | |
| 156 | const float d = amax / ((1 << 7) - 1); |
| 157 | id[row_iter] = d ? 1.0f / d : 0.0f; |
| 158 | |
| 159 | y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d)ggml_compute_fp32_to_fp16(d); |
| 160 | } |
| 161 | |
| 162 | for (int j = 0; j < QK8_032 * 4; j++) { |
| 163 | int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; |
| 164 | int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; |
| 165 | src_offset += (j % blck_size_interleave); |
| 166 | |
| 167 | float x0 = srcv[src_id][src_offset] * id[src_id]; |
| 168 | y[i].qs[j] = roundf(x0); |
| 169 | } |
| 170 | } |
| 171 | } |
| 172 | |
| 173 | void ggml_quantize_mat_q8_0_4x8_generic(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t k) { |
| 174 | assert(QK8_0 == 32)(static_cast <bool> (32 == 32) ? void (0) : __assert_fail ("QK8_0 == 32", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 175 | assert(k % QK8_0 == 0)(static_cast <bool> (k % 32 == 0) ? void (0) : __assert_fail ("k % QK8_0 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 176 | const int nb = k / QK8_032; |
| 177 | |
| 178 | block_q8_0x4 * GGML_RESTRICT__restrict__ y = (block_q8_0x4 *) vy; |
| 179 | |
| 180 | // scalar |
| 181 | const int blck_size_interleave = 8; |
| 182 | float srcv[4][QK8_032]; |
| 183 | float id[4]; |
| 184 | |
| 185 | for (int i = 0; i < nb; i++) { |
| 186 | for (int row_iter = 0; row_iter < 4; row_iter++) { |
| 187 | float amax = 0.0f; // absolute max |
| 188 | |
| 189 | for (int j = 0; j < QK8_032; j++) { |
| 190 | srcv[row_iter][j] = x[row_iter * k + i * QK8_032 + j]; |
| 191 | amax = MAX(amax, fabsf(srcv[row_iter][j]))((amax) > (fabsf(srcv[row_iter][j])) ? (amax) : (fabsf(srcv [row_iter][j]))); |
| 192 | } |
| 193 | |
| 194 | const float d = amax / ((1 << 7) - 1); |
| 195 | id[row_iter] = d ? 1.0f / d : 0.0f; |
| 196 | |
| 197 | y[i].d[row_iter] = GGML_CPU_FP32_TO_FP16(d)ggml_compute_fp32_to_fp16(d); |
| 198 | } |
| 199 | |
| 200 | for (int j = 0; j < QK8_032 * 4; j++) { |
| 201 | int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; |
| 202 | int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; |
| 203 | src_offset += (j % blck_size_interleave); |
| 204 | |
| 205 | float x0 = srcv[src_id][src_offset] * id[src_id]; |
| 206 | y[i].qs[j] = roundf(x0); |
| 207 | } |
| 208 | } |
| 209 | } |
| 210 | |
| 211 | void ggml_quantize_mat_q8_K_4x4_genericggml_quantize_mat_q8_K_4x4(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t k) { |
| 212 | assert(QK_K == 256)(static_cast <bool> (256 == 256) ? void (0) : __assert_fail ("QK_K == 256", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 213 | assert(k % QK_K == 0)(static_cast <bool> (k % 256 == 0) ? void (0) : __assert_fail ("k % QK_K == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 214 | const int nb = k / QK_K256; |
| 215 | |
| 216 | block_q8_Kx4 * GGML_RESTRICT__restrict__ y = (block_q8_Kx4 *) vy; |
| 217 | |
| 218 | // scalar |
| 219 | const int blck_size_interleave = 4; |
| 220 | float srcv[4][QK_K256]; |
| 221 | float iscale[4]; |
| 222 | |
| 223 | for (int i = 0; i < nb; i++) { |
| 224 | for (int row_iter = 0; row_iter < 4; row_iter++) { |
| 225 | float amax = 0.0f; // absolute max |
| 226 | float max = 0; |
| 227 | |
| 228 | for (int j = 0; j < QK_K256; j++) { |
| 229 | srcv[row_iter][j] = x[row_iter * k + i * QK_K256 + j]; |
| 230 | // Update the maximum value of the corresponding super block |
| 231 | if(amax < fabsf(srcv[row_iter][j])) { |
| 232 | amax = fabsf(srcv[row_iter][j]); |
| 233 | max = srcv[row_iter][j]; |
| 234 | } |
| 235 | } |
| 236 | |
| 237 | iscale[row_iter] = amax ? -127.f/max : 0; |
| 238 | |
| 239 | y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; |
| 240 | } |
| 241 | |
| 242 | for (int j = 0; j < QK_K256 / 4; j++) { |
| 243 | y[i].bsums[j] = 0; |
| 244 | } |
| 245 | |
| 246 | // Quants values are interleaved in sequence of four bytes from corresponding super blocks |
| 247 | // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving |
| 248 | // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on |
| 249 | for (int j = 0; j < QK_K256 * 4; j++) { |
| 250 | int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; |
| 251 | int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; |
| 252 | src_offset += (j % blck_size_interleave); |
| 253 | int index = (((j & 15) >> 2) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); |
| 254 | |
| 255 | float x0 = srcv[src_id][src_offset] * iscale[src_id]; |
| 256 | y[i].qs[j] = nearest_int(x0); |
| 257 | y[i].bsums[index] += y[i].qs[j]; |
| 258 | } |
| 259 | } |
| 260 | } |
| 261 | |
| 262 | void ggml_quantize_mat_q8_K_4x8_generic(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t k) { |
| 263 | assert(QK_K == 256)(static_cast <bool> (256 == 256) ? void (0) : __assert_fail ("QK_K == 256", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 264 | assert(k % QK_K == 0)(static_cast <bool> (k % 256 == 0) ? void (0) : __assert_fail ("k % QK_K == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 265 | const int nb = k / QK_K256; |
| 266 | |
| 267 | block_q8_Kx4 * GGML_RESTRICT__restrict__ y = (block_q8_Kx4 *) vy; |
| 268 | |
| 269 | // scalar |
| 270 | const int blck_size_interleave = 8; |
| 271 | float srcv[4][QK_K256]; |
| 272 | float iscale[4]; |
| 273 | |
| 274 | for (int i = 0; i < nb; i++) { |
| 275 | for (int row_iter = 0; row_iter < 4; row_iter++) { |
| 276 | float amax = 0.0f; // absolute max |
| 277 | float max = 0; |
| 278 | |
| 279 | for (int j = 0; j < QK_K256; j++) { |
| 280 | srcv[row_iter][j] = x[row_iter * k + i * QK_K256 + j]; |
| 281 | // Update the maximum value of the corresponding super block |
| 282 | if(amax < fabsf(srcv[row_iter][j])) { |
| 283 | amax = fabsf(srcv[row_iter][j]); |
| 284 | max = srcv[row_iter][j]; |
| 285 | } |
| 286 | } |
| 287 | |
| 288 | iscale[row_iter] = amax ? -127.f/max : 0; |
| 289 | |
| 290 | y[i].d[row_iter] = amax ? 1/iscale[row_iter] : 0; |
| 291 | } |
| 292 | |
| 293 | for (int j = 0; j < QK_K256 / 4; j++) { |
| 294 | y[i].bsums[j] = 0; |
| 295 | } |
| 296 | |
| 297 | // Quants values are interleaved in sequence of eight bytes from corresponding super blocks |
| 298 | // Bsums values are interleaved in sequence of four bsums from each super block taken for interleaving |
| 299 | // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on |
| 300 | for (int j = 0; j < QK_K256 * 4; j++) { |
| 301 | int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; |
| 302 | int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; |
| 303 | src_offset += (j % blck_size_interleave); |
| 304 | int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); |
| 305 | |
| 306 | float x0 = srcv[src_id][src_offset] * iscale[src_id]; |
| 307 | y[i].qs[j] = nearest_int(x0); |
| 308 | y[i].bsums[index] += y[i].qs[j]; |
| 309 | } |
| 310 | } |
| 311 | } |
| 312 | |
| 313 | } // extern "C" |
| 314 | |
| 315 | template <int64_t INTER_SIZE, ggml_type PARAM_TYPE> |
| 316 | void ggml_quantize_mat_t(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row); |
| 317 | |
| 318 | template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row) { |
| 319 | assert(nrow == 4)(static_cast <bool> (nrow == 4) ? void (0) : __assert_fail ("nrow == 4", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 320 | UNUSED(nrow)(void)(nrow); |
| 321 | ggml_quantize_mat_q8_0_4x4(x, vy, n_per_row); |
| 322 | } |
| 323 | |
| 324 | template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row) { |
| 325 | assert(nrow == 4)(static_cast <bool> (nrow == 4) ? void (0) : __assert_fail ("nrow == 4", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 326 | UNUSED(nrow)(void)(nrow); |
| 327 | ggml_quantize_mat_q8_0_4x8(x, vy, n_per_row); |
| 328 | } |
| 329 | |
| 330 | template <> void ggml_quantize_mat_t<4, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row) { |
| 331 | assert(nrow == 4)(static_cast <bool> (nrow == 4) ? void (0) : __assert_fail ("nrow == 4", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 332 | UNUSED(nrow)(void)(nrow); |
| 333 | ggml_quantize_mat_q8_K_4x4(x, vy, n_per_row); |
| 334 | } |
| 335 | |
| 336 | template <> void ggml_quantize_mat_t<8, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row) { |
| 337 | assert(nrow == 4)(static_cast <bool> (nrow == 4) ? void (0) : __assert_fail ("nrow == 4", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 338 | UNUSED(nrow)(void)(nrow); |
| 339 | ggml_quantize_mat_q8_K_4x8(x, vy, n_per_row); |
| 340 | } |
| 341 | |
| 342 | #if defined __riscv_zvfh |
| 343 | template <> void ggml_quantize_mat_t<1, GGML_TYPE_Q8_0>(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row) { |
| 344 | assert(nrow == 4)(static_cast <bool> (nrow == 4) ? void (0) : __assert_fail ("nrow == 4", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 345 | UNUSED(nrow)(void)(nrow); |
| 346 | ggml_quantize_mat_q8_0_4x1(x, vy, n_per_row); |
| 347 | } |
| 348 | |
| 349 | template <> void ggml_quantize_mat_t<1, GGML_TYPE_Q8_K>(const float * GGML_RESTRICT__restrict__ x, void * GGML_RESTRICT__restrict__ vy, int64_t nrow, int64_t n_per_row) { |
| 350 | assert(nrow == 4)(static_cast <bool> (nrow == 4) ? void (0) : __assert_fail ("nrow == 4", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 351 | UNUSED(nrow)(void)(nrow); |
| 352 | ggml_quantize_mat_q8_K_4x1(x, vy, n_per_row); |
| 353 | } |
| 354 | #endif |
| 355 | |
| 356 | template <int M, int N> |
| 357 | static void ggml_gemv_q6_K_NxM_q8_K_generic_impl(int n, |
| 358 | float * GGML_RESTRICT__restrict__ s, |
| 359 | size_t bs, |
| 360 | const void * GGML_RESTRICT__restrict__ vx, |
| 361 | const void * GGML_RESTRICT__restrict__ vy, |
| 362 | int nr, |
| 363 | int nc) { |
| 364 | constexpr int blocklen = M; |
| 365 | constexpr int ncols_interleaved = N; |
| 366 | const int qk = QK_K256; |
| 367 | const int nb = n / qk; |
| 368 | const int blocks_per_half = 64 / blocklen; |
| 369 | |
| 370 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 371 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 372 | |
| 373 | UNUSED(bs)(void)(bs); |
| 374 | UNUSED(nr)(void)(nr); |
| 375 | |
| 376 | float sumf[8]; |
| 377 | |
| 378 | const block_q8_K * a_ptr = (const block_q8_K *) vy; |
| 379 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 380 | const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb); |
| 381 | |
| 382 | for (int j = 0; j < ncols_interleaved; j++) { |
| 383 | sumf[j] = 0.0f; |
| 384 | } |
| 385 | |
| 386 | for (int l = 0; l < nb; l++) { |
| 387 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 388 | const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen; |
| 389 | const int base_h = base_l + 64; |
| 390 | |
| 391 | const int scale_idx_l = base_l / 16; |
| 392 | const int scale_idx_h = base_h / 16; |
| 393 | |
| 394 | const int qh_shift_l = ((base_l % 128) / 32) * 2; |
| 395 | const int qh_shift_h = ((base_h % 128) / 32) * 2; |
| 396 | |
| 397 | const int qh_half_l = (base_l / 128) * 32; |
| 398 | const int qh_half_h = (base_h / 128) * 32; |
| 399 | |
| 400 | for (int j = 0; j < ncols_interleaved; j++) { |
| 401 | const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j]; |
| 402 | const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j]; |
| 403 | |
| 404 | int sumi_l = 0; |
| 405 | int sumi_h = 0; |
| 406 | |
| 407 | for (int i = 0; i < blocklen; i++) { |
| 408 | const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i; |
| 409 | const int l_4 = b_ptr[l].ql[ql_pos] & 0xF; |
| 410 | const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF; |
| 411 | |
| 412 | const int qh_idx_l = qh_half_l + ((base_l + i) % 32); |
| 413 | const int qh_chunk_l = qh_idx_l / blocklen; |
| 414 | const int qh_pos_l = qh_idx_l % blocklen; |
| 415 | const int qh_offset_l = qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l; |
| 416 | const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3; |
| 417 | |
| 418 | const int qh_idx_h = qh_half_h + ((base_h + i) % 32); |
| 419 | const int qh_chunk_h = qh_idx_h / blocklen; |
| 420 | const int qh_pos_h = qh_idx_h % blocklen; |
| 421 | const int qh_offset_h = qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h; |
| 422 | const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3; |
| 423 | |
| 424 | const int q_l = ((hi_2_l << 4) | l_4) - 32; |
| 425 | const int q_h = ((hi_2_h << 4) | hi_4) - 32; |
| 426 | |
| 427 | const int8_t a_l = a_ptr[l].qs[base_l + i]; |
| 428 | const int8_t a_h = a_ptr[l].qs[base_h + i]; |
| 429 | |
| 430 | sumi_l += q_l * a_l; |
| 431 | sumi_h += q_h * a_h; |
| 432 | } |
| 433 | |
| 434 | sumf[j] += |
| 435 | (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d; |
| 436 | } |
| 437 | } |
| 438 | } |
| 439 | |
| 440 | for (int j = 0; j < ncols_interleaved; j++) { |
| 441 | s[x * ncols_interleaved + j] = sumf[j]; |
| 442 | } |
| 443 | } |
| 444 | } |
| 445 | |
| 446 | template <int M, int N> |
| 447 | static void ggml_gemm_q6_K_NxM_q8_K_generic_impl(int n, |
| 448 | float * GGML_RESTRICT__restrict__ s, |
| 449 | size_t bs, |
| 450 | const void * GGML_RESTRICT__restrict__ vx, |
| 451 | const void * GGML_RESTRICT__restrict__ vy, |
| 452 | int nr, |
| 453 | int nc) { |
| 454 | constexpr int blocklen = M; |
| 455 | constexpr int ncols_interleaved = N; |
| 456 | const int qk = QK_K256; |
| 457 | const int nb = n / qk; |
| 458 | const int blocks_per_half = 64 / blocklen; |
| 459 | const int q8_half_stride = 512; |
| 460 | const int q8_low_high_step = 256; |
| 461 | |
| 462 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 463 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 464 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 465 | |
| 466 | UNUSED(bs)(void)(bs); |
| 467 | |
| 468 | float sumf[4][8]; |
| 469 | |
| 470 | for (int y = 0; y < nr / 4; y++) { |
| 471 | const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); |
| 472 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 473 | const block_q6_Kx8 * b_ptr = (const block_q6_Kx8 *) vx + (x * nb); |
| 474 | |
| 475 | for (int m = 0; m < 4; m++) { |
| 476 | for (int j = 0; j < ncols_interleaved; j++) { |
| 477 | sumf[m][j] = 0.0f; |
| 478 | } |
| 479 | } |
| 480 | |
| 481 | for (int l = 0; l < nb; l++) { |
| 482 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 483 | const int base_l = (k / blocks_per_half) * 128 + (k % blocks_per_half) * blocklen; |
| 484 | const int base_h = base_l + 64; |
| 485 | |
| 486 | const int scale_idx_l = base_l / 16; |
| 487 | const int scale_idx_h = base_h / 16; |
| 488 | |
| 489 | const int qh_shift_l = ((base_l % 128) / 32) * 2; |
| 490 | const int qh_shift_h = ((base_h % 128) / 32) * 2; |
| 491 | |
| 492 | const int qh_half_l = (base_l / 128) * 32; |
| 493 | const int qh_half_h = (base_h / 128) * 32; |
| 494 | |
| 495 | const int q8_base = (k / blocks_per_half) * q8_half_stride + (k % blocks_per_half) * (blocklen * 4); |
| 496 | |
| 497 | for (int m = 0; m < 4; m++) { |
| 498 | for (int j = 0; j < ncols_interleaved; j++) { |
| 499 | const int8_t scale_l = b_ptr[l].scales[scale_idx_l * ncols_interleaved + j]; |
| 500 | const int8_t scale_h = b_ptr[l].scales[scale_idx_h * ncols_interleaved + j]; |
| 501 | |
| 502 | int sumi_l = 0; |
| 503 | int sumi_h = 0; |
| 504 | |
| 505 | for (int i = 0; i < blocklen; i++) { |
| 506 | const int ql_pos = k * ncols_interleaved * blocklen + j * blocklen + i; |
| 507 | const int l_4 = b_ptr[l].ql[ql_pos] & 0xF; |
| 508 | const int hi_4 = (b_ptr[l].ql[ql_pos] >> 4) & 0xF; |
| 509 | |
| 510 | const int qh_idx_l = qh_half_l + ((base_l + i) % 32); |
| 511 | const int qh_chunk_l = qh_idx_l / blocklen; |
| 512 | const int qh_pos_l = qh_idx_l % blocklen; |
| 513 | const int qh_offset_l = |
| 514 | qh_chunk_l * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_l; |
| 515 | const int hi_2_l = (b_ptr[l].qh[qh_offset_l] >> qh_shift_l) & 0x3; |
| 516 | |
| 517 | const int qh_idx_h = qh_half_h + ((base_h + i) % 32); |
| 518 | const int qh_chunk_h = qh_idx_h / blocklen; |
| 519 | const int qh_pos_h = qh_idx_h % blocklen; |
| 520 | const int qh_offset_h = |
| 521 | qh_chunk_h * (blocklen * ncols_interleaved) + j * blocklen + qh_pos_h; |
| 522 | const int hi_2_h = (b_ptr[l].qh[qh_offset_h] >> qh_shift_h) & 0x3; |
| 523 | |
| 524 | const int q_l = ((hi_2_l << 4) | l_4) - 32; |
| 525 | const int q_h = ((hi_2_h << 4) | hi_4) - 32; |
| 526 | |
| 527 | const int8_t q8_l = a_ptr[l].qs[q8_base + m * blocklen + i]; |
| 528 | const int8_t q8_h = a_ptr[l].qs[q8_base + m * blocklen + i + q8_low_high_step]; |
| 529 | |
| 530 | sumi_l += q_l * q8_l; |
| 531 | sumi_h += q_h * q8_h; |
| 532 | } |
| 533 | |
| 534 | sumf[m][j] += (sumi_l * scale_l + sumi_h * scale_h) * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * |
| 535 | a_ptr[l].d[m]; |
| 536 | } |
| 537 | } |
| 538 | } |
| 539 | } |
| 540 | |
| 541 | for (int m = 0; m < 4; m++) { |
| 542 | for (int j = 0; j < ncols_interleaved; j++) { |
| 543 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 544 | } |
| 545 | } |
| 546 | } |
| 547 | } |
| 548 | } |
| 549 | |
| 550 | template <int M, int N> |
| 551 | static void ggml_gemv_q5_K_NxM_q8_K_generic_impl(int n, |
| 552 | float * GGML_RESTRICT__restrict__ s, |
| 553 | size_t bs, |
| 554 | const void * GGML_RESTRICT__restrict__ vx, |
| 555 | const void * GGML_RESTRICT__restrict__ vy, |
| 556 | int nr, |
| 557 | int nc) { |
| 558 | constexpr int blocklen = M; |
| 559 | constexpr int ncols_interleaved = N; |
| 560 | const int qk = QK_K256; |
| 561 | const int nb = n / qk; |
| 562 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 563 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 564 | static const uint32_t kmask3 = 0x03030303; |
| 565 | |
| 566 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 567 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 568 | |
| 569 | UNUSED(bs)(void)(bs); |
| 570 | UNUSED(nr)(void)(nr); |
| 571 | |
| 572 | float sumf[ncols_interleaved]; |
| 573 | float sum_minf[ncols_interleaved]; |
| 574 | uint32_t utmp[32]; |
| 575 | int sumi1; |
| 576 | int sumi2; |
| 577 | int sumi; |
| 578 | |
| 579 | const block_q8_K * a_ptr = (const block_q8_K *) vy; |
| 580 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 581 | const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb); |
| 582 | |
| 583 | for (int j = 0; j < ncols_interleaved; j++) { |
| 584 | sumf[j] = 0.0; |
| 585 | sum_minf[j] = 0.0; |
| 586 | } |
| 587 | for (int l = 0; l < nb; l++) { |
| 588 | for (int sb = 0; sb < 8; sb++) { |
| 589 | memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE12, K_SCALE_SIZE12); |
| 590 | utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); |
| 591 | const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; |
| 592 | utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); |
| 593 | utmp[sb * 4 + 2] = uaux_0; |
| 594 | utmp[sb * 4 + 0] &= kmask1; |
| 595 | } |
| 596 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 597 | constexpr int scale_stride = 32; |
| 598 | uint8_t * scales_0 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride; |
| 599 | uint8_t * scales_1 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16; |
| 600 | |
| 601 | const int qh_shift = (k / (32 / blocklen)) * 2; |
| 602 | for (int j = 0; j < ncols_interleaved; j++) { |
| 603 | sumi1 = 0; |
| 604 | sumi2 = 0; |
| 605 | sumi = 0; |
| 606 | for (int i = 0; i < blocklen; ++i) { |
| 607 | const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i; |
| 608 | |
| 609 | const int qh_idx = (k * blocklen + i) % 32; |
| 610 | const int qh_chunk = qh_idx / blocklen; |
| 611 | const int qh_pos = qh_idx % blocklen; |
| 612 | const int b_qh_offset = qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos; |
| 613 | |
| 614 | const uint8_t qh_val = b_ptr[l].qh[b_qh_offset]; |
| 615 | const uint8_t h0 = (qh_val >> qh_shift) & 1; |
| 616 | const uint8_t h1 = (qh_val >> (qh_shift + 1)) & 1; |
| 617 | |
| 618 | const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4)); |
| 619 | const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4)); |
| 620 | |
| 621 | const int q8_offset = (k / (32 / blocklen)) * 64 + (k % (32 / blocklen)) * blocklen + i; |
| 622 | |
| 623 | sumi1 = (v0 * a_ptr[l].qs[q8_offset]); |
| 624 | sumi2 = (v1 * a_ptr[l].qs[q8_offset + 32]); |
| 625 | sumi1 = sumi1 * scales_0[j]; |
| 626 | sumi2 = sumi2 * scales_1[j]; |
| 627 | sumi += sumi1 + sumi2; |
| 628 | } |
| 629 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d; |
| 630 | } |
| 631 | } |
| 632 | for (int sb = 0; sb < 8; sb++) { |
| 633 | uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16; |
| 634 | for (int j = 0; j < ncols_interleaved; j++) { |
| 635 | sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * |
| 636 | GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d; |
| 637 | } |
| 638 | } |
| 639 | } |
| 640 | for (int j = 0; j < ncols_interleaved; j++) { |
| 641 | s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; |
| 642 | } |
| 643 | } |
| 644 | } |
| 645 | |
| 646 | template <int M, int N> |
| 647 | static void ggml_gemm_q5_K_NxM_q8_K_generic_impl(int n, |
| 648 | float * GGML_RESTRICT__restrict__ s, |
| 649 | size_t bs, |
| 650 | const void * GGML_RESTRICT__restrict__ vx, |
| 651 | const void * GGML_RESTRICT__restrict__ vy, |
| 652 | int nr, |
| 653 | int nc) { |
| 654 | constexpr int blocklen = M; |
| 655 | constexpr int ncols_interleaved = N; |
| 656 | const int qk = QK_K256; |
| 657 | const int nb = n / qk; |
| 658 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 659 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 660 | static const uint32_t kmask3 = 0x03030303; |
| 661 | |
| 662 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 663 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 664 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 665 | |
| 666 | float sumf[4][ncols_interleaved]; |
| 667 | float sum_minf[4][ncols_interleaved]; |
| 668 | uint32_t utmp[32]; |
| 669 | int sumi1; |
| 670 | int sumi2; |
| 671 | int sumi; |
| 672 | |
| 673 | for (int y = 0; y < nr / 4; y++) { |
| 674 | const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); |
| 675 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 676 | const block_q5_Kx8 * b_ptr = (const block_q5_Kx8 *) vx + (x * nb); |
| 677 | for (int m = 0; m < 4; m++) { |
| 678 | for (int j = 0; j < ncols_interleaved; j++) { |
| 679 | sumf[m][j] = 0.0; |
| 680 | sum_minf[m][j] = 0.0; |
| 681 | } |
| 682 | } |
| 683 | for (int l = 0; l < nb; l++) { |
| 684 | for (int sb = 0; sb < 8; sb++) { |
| 685 | memcpy(utmp + sb * 4, b_ptr[l].scales + sb * K_SCALE_SIZE12, K_SCALE_SIZE12); |
| 686 | utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); |
| 687 | const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; |
| 688 | utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); |
| 689 | utmp[sb * 4 + 2] = uaux_0; |
| 690 | utmp[sb * 4 + 0] &= kmask1; |
| 691 | } |
| 692 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 693 | constexpr int scale_stride = 32; |
| 694 | uint8_t * scales_0 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride; |
| 695 | uint8_t * scales_1 = (uint8_t *) utmp + (k / (32 / blocklen)) * scale_stride + 16; |
| 696 | |
| 697 | const int qh_shift = (k / (32 / blocklen)) * 2; |
| 698 | for (int m = 0; m < 4; m++) { |
| 699 | for (int j = 0; j < ncols_interleaved; j++) { |
| 700 | sumi1 = 0; |
| 701 | sumi2 = 0; |
| 702 | sumi = 0; |
| 703 | for (int i = 0; i < blocklen; ++i) { |
| 704 | const int b_qs_offset = k * ncols_interleaved * blocklen + j * blocklen + i; |
| 705 | |
| 706 | const int qh_idx = (k * blocklen + i) % 32; |
| 707 | const int qh_chunk = qh_idx / blocklen; |
| 708 | const int qh_pos = qh_idx % blocklen; |
| 709 | const int b_qh_offset = |
| 710 | qh_chunk * (blocklen * ncols_interleaved) + j * blocklen + qh_pos; |
| 711 | |
| 712 | const uint8_t qh_val = b_ptr[l].qh[b_qh_offset]; |
| 713 | const uint8_t h0 = (qh_val >> qh_shift) & 1; |
| 714 | const uint8_t h1 = (qh_val >> (qh_shift + 1)) & 1; |
| 715 | |
| 716 | const int v0 = (int8_t) ((b_ptr[l].qs[b_qs_offset] & 0xF) | (h0 << 4)); |
| 717 | const int v1 = (int8_t) ((b_ptr[l].qs[b_qs_offset] >> 4) | (h1 << 4)); |
| 718 | |
| 719 | const int q8_offset = (k / (32 / blocklen)) * 256 + |
| 720 | (k % (32 / blocklen)) * 4 * blocklen + m * blocklen + i; |
| 721 | |
| 722 | sumi1 = (v0 * a_ptr[l].qs[q8_offset]); |
| 723 | sumi2 = (v1 * a_ptr[l].qs[q8_offset + 128]); |
| 724 | sumi1 = sumi1 * scales_0[j]; |
| 725 | sumi2 = sumi2 * scales_1[j]; |
| 726 | sumi += sumi1 + sumi2; |
| 727 | } |
| 728 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d[m]; |
| 729 | } |
| 730 | } |
| 731 | } |
| 732 | for (int sb = 0; sb < 8; sb++) { |
| 733 | uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16; |
| 734 | for (int m = 0; m < 4; m++) { |
| 735 | const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); |
| 736 | for (int j = 0; j < ncols_interleaved; j++) { |
| 737 | sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * |
| 738 | GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; |
| 739 | } |
| 740 | } |
| 741 | } |
| 742 | } |
| 743 | for (int m = 0; m < 4; m++) { |
| 744 | for (int j = 0; j < ncols_interleaved; j++) { |
| 745 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; |
| 746 | } |
| 747 | } |
| 748 | } |
| 749 | } |
| 750 | } |
| 751 | |
| 752 | extern "C" { |
| 753 | |
| 754 | void ggml_gemv_q4_0_4x4_q8_0_genericggml_gemv_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 755 | const int qk = QK8_032; |
| 756 | const int nb = n / qk; |
| 757 | const int ncols_interleaved = 4; |
| 758 | const int blocklen = 4; |
| 759 | |
| 760 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 761 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 762 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 763 | |
| 764 | UNUSED(s)(void)(s); |
| 765 | UNUSED(bs)(void)(bs); |
| 766 | UNUSED(vx)(void)(vx); |
| 767 | UNUSED(vy)(void)(vy); |
| 768 | UNUSED(nr)(void)(nr); |
| 769 | UNUSED(nc)(void)(nc); |
| 770 | UNUSED(nb)(void)(nb); |
| 771 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 772 | UNUSED(blocklen)(void)(blocklen); |
| 773 | |
| 774 | float sumf[4]; |
| 775 | int sumi; |
| 776 | |
| 777 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 778 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 779 | const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); |
| 780 | |
| 781 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 782 | for (int l = 0; l < nb; l++) { |
| 783 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 784 | for (int j = 0; j < ncols_interleaved; j++) { |
| 785 | sumi = 0; |
| 786 | for (int i = 0; i < blocklen; ++i) { |
| 787 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 788 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 789 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; |
| 790 | } |
| 791 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 792 | } |
| 793 | } |
| 794 | } |
| 795 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 796 | } |
| 797 | } |
| 798 | |
| 799 | void ggml_gemv_q4_0_4x8_q8_0_genericggml_gemv_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 800 | const int qk = QK8_032; |
| 801 | const int nb = n / qk; |
| 802 | const int ncols_interleaved = 4; |
| 803 | const int blocklen = 8; |
| 804 | |
| 805 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 806 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 807 | |
| 808 | UNUSED(s)(void)(s); |
| 809 | UNUSED(bs)(void)(bs); |
| 810 | UNUSED(vx)(void)(vx); |
| 811 | UNUSED(vy)(void)(vy); |
| 812 | UNUSED(nr)(void)(nr); |
| 813 | UNUSED(nc)(void)(nc); |
| 814 | UNUSED(nb)(void)(nb); |
| 815 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 816 | UNUSED(blocklen)(void)(blocklen); |
| 817 | |
| 818 | float sumf[4]; |
| 819 | int sumi; |
| 820 | |
| 821 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 822 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 823 | const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); |
| 824 | |
| 825 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 826 | for (int l = 0; l < nb; l++) { |
| 827 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 828 | for (int j = 0; j < ncols_interleaved; j++) { |
| 829 | sumi = 0; |
| 830 | for (int i = 0; i < blocklen; ++i) { |
| 831 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 832 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 833 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; |
| 834 | } |
| 835 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 836 | } |
| 837 | } |
| 838 | } |
| 839 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 840 | } |
| 841 | } |
| 842 | |
| 843 | void ggml_gemv_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 844 | const int qk = QK8_032; |
| 845 | const int nb = n / qk; |
| 846 | const int ncols_interleaved = 8; |
| 847 | const int blocklen = 8; |
| 848 | |
| 849 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 850 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 851 | |
| 852 | UNUSED(s)(void)(s); |
| 853 | UNUSED(bs)(void)(bs); |
| 854 | UNUSED(vx)(void)(vx); |
| 855 | UNUSED(vy)(void)(vy); |
| 856 | UNUSED(nr)(void)(nr); |
| 857 | UNUSED(nc)(void)(nc); |
| 858 | UNUSED(nb)(void)(nb); |
| 859 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 860 | UNUSED(blocklen)(void)(blocklen); |
| 861 | |
| 862 | float sumf[8]; |
| 863 | int sumi; |
| 864 | |
| 865 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 866 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 867 | const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); |
| 868 | |
| 869 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 870 | for (int l = 0; l < nb; l++) { |
| 871 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 872 | for (int j = 0; j < ncols_interleaved; j++) { |
| 873 | sumi = 0; |
| 874 | for (int i = 0; i < blocklen; ++i) { |
| 875 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 876 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 877 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; |
| 878 | } |
| 879 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 880 | } |
| 881 | } |
| 882 | } |
| 883 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 884 | } |
| 885 | } |
| 886 | |
| 887 | void ggml_gemv_q4_K_8x4_q8_K_genericggml_gemv_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 888 | const int qk = QK_K256; |
| 889 | const int nb = n / qk; |
| 890 | const int ncols_interleaved = 8; |
| 891 | const int blocklen = 4; |
| 892 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 893 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 894 | static const uint32_t kmask3 = 0x03030303; |
| 895 | |
| 896 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 897 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 898 | |
| 899 | UNUSED(bs)(void)(bs); |
| 900 | UNUSED(nr)(void)(nr); |
| 901 | |
| 902 | float sumf[8]; |
| 903 | float sum_minf[8]; |
| 904 | uint32_t utmp[32]; |
| 905 | int sumi1; |
| 906 | int sumi2; |
| 907 | int sumi; |
| 908 | |
| 909 | const block_q8_K * a_ptr = (const block_q8_K *) vy; |
| 910 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 911 | const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); |
| 912 | |
| 913 | for (int j = 0; j < ncols_interleaved; j++) { |
| 914 | sumf[j] = 0.0; |
| 915 | sum_minf[j] = 0.0; |
| 916 | } |
| 917 | for (int l = 0; l < nb; l++) { |
| 918 | for (int sb = 0; sb < 8; sb++) { |
| 919 | memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); |
| 920 | utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); |
| 921 | const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; |
| 922 | utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); |
| 923 | utmp[sb * 4 + 2] = uaux_0; |
| 924 | utmp[sb * 4 + 0] &= kmask1; |
| 925 | } |
| 926 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 927 | uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32; |
| 928 | uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16; |
| 929 | for (int j = 0; j < ncols_interleaved; j++) { |
| 930 | sumi1 = 0; |
| 931 | sumi2 = 0; |
Value stored to 'sumi2' is never read | |
| 932 | sumi = 0; |
| 933 | for (int i = 0; i < blocklen; ++i) { |
| 934 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); |
| 935 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); |
| 936 | sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i]); |
| 937 | sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 64 + (k % 8) * blocklen + i + 32]); |
| 938 | sumi1 = sumi1 * scales_0[j]; |
| 939 | sumi2 = sumi2 * scales_1[j]; |
| 940 | sumi += sumi1 + sumi2; |
| 941 | } |
| 942 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d; |
| 943 | } |
| 944 | } |
| 945 | for (int sb = 0; sb < 8; sb++) { |
| 946 | uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16; |
| 947 | for (int j = 0; j < ncols_interleaved; j++) { |
| 948 | sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d; |
| 949 | } |
| 950 | } |
| 951 | } |
| 952 | for (int j = 0; j < ncols_interleaved; j++) { |
| 953 | s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; |
| 954 | } |
| 955 | } |
| 956 | } |
| 957 | |
| 958 | void ggml_gemv_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 959 | const int qk = QK_K256; |
| 960 | const int nb = n / qk; |
| 961 | const int ncols_interleaved = 8; |
| 962 | const int blocklen = 8; |
| 963 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 964 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 965 | static const uint32_t kmask3 = 0x03030303; |
| 966 | |
| 967 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 968 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 969 | |
| 970 | UNUSED(bs)(void)(bs); |
| 971 | UNUSED(nr)(void)(nr); |
| 972 | |
| 973 | float sumf[8]; |
| 974 | float sum_minf[8]; |
| 975 | uint32_t utmp[32]; |
| 976 | int sumi1; |
| 977 | int sumi2; |
| 978 | int sumi; |
| 979 | |
| 980 | const block_q8_K * a_ptr = (const block_q8_K *) vy; |
| 981 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 982 | const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); |
| 983 | |
| 984 | for (int j = 0; j < ncols_interleaved; j++) { |
| 985 | sumf[j] = 0.0; |
| 986 | sum_minf[j] = 0.0; |
| 987 | } |
| 988 | for (int l = 0; l < nb; l++) { |
| 989 | for (int sb = 0; sb < 8; sb++) { |
| 990 | memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); |
| 991 | utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); |
| 992 | const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; |
| 993 | utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); |
| 994 | utmp[sb * 4 + 2] = uaux_0; |
| 995 | utmp[sb * 4 + 0] &= kmask1; |
| 996 | } |
| 997 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 998 | uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; |
| 999 | uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; |
| 1000 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1001 | sumi1 = 0; |
| 1002 | sumi2 = 0; |
| 1003 | sumi = 0; |
| 1004 | for (int i = 0; i < blocklen; ++i) { |
| 1005 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); |
| 1006 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); |
| 1007 | sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i]); |
| 1008 | sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 64 + (k % 4) * blocklen + i + 32]); |
| 1009 | sumi1 = sumi1 * scales_0[j]; |
| 1010 | sumi2 = sumi2 * scales_1[j]; |
| 1011 | sumi += sumi1 + sumi2; |
| 1012 | } |
| 1013 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d; |
| 1014 | } |
| 1015 | } |
| 1016 | for (int sb = 0; sb < 8; sb++) { |
| 1017 | uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; |
| 1018 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1019 | sum_minf[j] += mins[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d; |
| 1020 | } |
| 1021 | } |
| 1022 | } |
| 1023 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1024 | s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; |
| 1025 | } |
| 1026 | } |
| 1027 | } |
| 1028 | |
| 1029 | void ggml_gemv_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1030 | const int qk = QK_K256; |
| 1031 | const int nb = n / qk; |
| 1032 | const int ncols_interleaved = 8; |
| 1033 | const int blocklen = 8; |
| 1034 | |
| 1035 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1036 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1037 | |
| 1038 | UNUSED(s)(void)(s); |
| 1039 | UNUSED(bs)(void)(bs); |
| 1040 | UNUSED(vx)(void)(vx); |
| 1041 | UNUSED(vy)(void)(vy); |
| 1042 | UNUSED(nr)(void)(nr); |
| 1043 | UNUSED(nc)(void)(nc); |
| 1044 | UNUSED(nb)(void)(nb); |
| 1045 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1046 | UNUSED(blocklen)(void)(blocklen); |
| 1047 | |
| 1048 | float sumf[8]; |
| 1049 | float sum_minf[8]; |
| 1050 | int sumi1,sumi2,sumi3,sumi4; |
| 1051 | int sumi; |
| 1052 | |
| 1053 | const block_q8_K * a_ptr = (const block_q8_K *)vy; |
| 1054 | for(int x = 0; x < nc / ncols_interleaved; x++) { |
| 1055 | const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb); |
| 1056 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1057 | sumf[j] = 0.0; |
| 1058 | sum_minf[j] = 0.0; |
| 1059 | } |
| 1060 | for (int l = 0; l < nb; l++) { |
| 1061 | for (int k = 0; k < (qk / (4 * blocklen)); k++) { |
| 1062 | const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ; |
| 1063 | const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16; |
| 1064 | const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32; |
| 1065 | const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48; |
| 1066 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1067 | sumi1 = 0; |
| 1068 | sumi2 = 0; |
| 1069 | sumi3 = 0; |
| 1070 | sumi4 = 0; |
| 1071 | sumi = 0; |
| 1072 | int offset = ((k / 2) % 2) + j * 2; |
| 1073 | for (int i = 0; i < blocklen; ++i){ |
| 1074 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3); |
| 1075 | const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3); |
| 1076 | const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3); |
| 1077 | const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3); |
| 1078 | sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i]); |
| 1079 | sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 32]); |
| 1080 | sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 64]); |
| 1081 | sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 128 + (k % 4) * blocklen + i + 96]); |
| 1082 | |
| 1083 | sumi1 = sumi1 * (scales_0[offset] & 0xF); |
| 1084 | sumi2 = sumi2 * (scales_1[offset] & 0xF); |
| 1085 | sumi3 = sumi3 * (scales_2[offset] & 0xF); |
| 1086 | sumi4 = sumi4 * (scales_3[offset] & 0xF); |
| 1087 | sumi += sumi1 + sumi2 + sumi3 + sumi4; |
| 1088 | } |
| 1089 | sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j])ggml_compute_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d; |
| 1090 | } |
| 1091 | } |
| 1092 | for(int sb = 0; sb < 8; sb++) { |
| 1093 | const uint8_t *mins = b_ptr[l].scales + sb * 16; |
| 1094 | for(int j = 0; j < ncols_interleaved; j++){ |
| 1095 | sum_minf[j] += ((mins[j * 2] >> 4) * a_ptr[l].bsums[sb * 2] + (mins[(j * 2)+ 1] >> 4) * a_ptr[l].bsums[sb * 2 + 1]) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_compute_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d; |
| 1096 | } |
| 1097 | } |
| 1098 | } |
| 1099 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1100 | s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; |
| 1101 | } |
| 1102 | } |
| 1103 | } |
| 1104 | |
| 1105 | void ggml_gemv_q5_K_8x4_q8_K_genericggml_gemv_q5_K_8x4_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1106 | ggml_gemv_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc); |
| 1107 | } |
| 1108 | |
| 1109 | void ggml_gemv_q5_K_8x8_q8_K_genericggml_gemv_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1110 | ggml_gemv_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc); |
| 1111 | } |
| 1112 | |
| 1113 | |
| 1114 | void ggml_gemv_q6_K_8x4_q8_K_genericggml_gemv_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1115 | ggml_gemv_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc); |
| 1116 | } |
| 1117 | |
| 1118 | void ggml_gemv_q6_K_8x8_q8_K_genericggml_gemv_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1119 | ggml_gemv_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc); |
| 1120 | } |
| 1121 | |
| 1122 | void ggml_gemv_iq4_nl_4x4_q8_0_genericggml_gemv_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1123 | const int qk = QK8_032; |
| 1124 | const int nb = n / qk; |
| 1125 | const int ncols_interleaved = 4; |
| 1126 | const int blocklen = 4; |
| 1127 | |
| 1128 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1129 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1130 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1131 | |
| 1132 | UNUSED(bs)(void)(bs); |
| 1133 | UNUSED(nr)(void)(nr); |
| 1134 | |
| 1135 | float sumf[4]; |
| 1136 | int sumi; |
| 1137 | |
| 1138 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1139 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1140 | const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); |
| 1141 | |
| 1142 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 1143 | for (int l = 0; l < nb; l++) { |
| 1144 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1145 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1146 | sumi = 0; |
| 1147 | for (int i = 0; i < blocklen; ++i) { |
| 1148 | const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 1149 | const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 1150 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); |
| 1151 | } |
| 1152 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1153 | } |
| 1154 | } |
| 1155 | } |
| 1156 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 1157 | } |
| 1158 | } |
| 1159 | |
| 1160 | void ggml_gemv_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1161 | const int qk = QK8_032; |
| 1162 | const int nb = n / qk; |
| 1163 | const int ncols_interleaved = 8; |
| 1164 | const int blocklen = 8; |
| 1165 | |
| 1166 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1167 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1168 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1169 | |
| 1170 | UNUSED(bs)(void)(bs); |
| 1171 | UNUSED(nr)(void)(nr); |
| 1172 | |
| 1173 | float sumf[8]; |
| 1174 | int sumi; |
| 1175 | |
| 1176 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1177 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1178 | const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb); |
| 1179 | |
| 1180 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 1181 | for (int l = 0; l < nb; l++) { |
| 1182 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1183 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1184 | sumi = 0; |
| 1185 | for (int i = 0; i < blocklen; ++i) { |
| 1186 | const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 1187 | const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 1188 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); |
| 1189 | } |
| 1190 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1191 | } |
| 1192 | } |
| 1193 | } |
| 1194 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 1195 | } |
| 1196 | } |
| 1197 | |
| 1198 | void ggml_gemv_mxfp4_4x4_q8_0_genericggml_gemv_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1199 | const int qk = QK8_032; |
| 1200 | const int nb = n / qk; |
| 1201 | const int ncols_interleaved = 4; |
| 1202 | const int blocklen = 4; |
| 1203 | |
| 1204 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1205 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1206 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1207 | |
| 1208 | UNUSED(bs)(void)(bs); |
| 1209 | UNUSED(nr)(void)(nr); |
| 1210 | |
| 1211 | float sumf[4]; |
| 1212 | int sumi; |
| 1213 | |
| 1214 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1215 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1216 | const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb); |
| 1217 | |
| 1218 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 1219 | for (int l = 0; l < nb; l++) { |
| 1220 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1221 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1222 | sumi = 0; |
| 1223 | for (int i = 0; i < blocklen; ++i) { |
| 1224 | const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 1225 | const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 1226 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); |
| 1227 | } |
| 1228 | sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j])ggml_table_f32_e8m0_half[(uint8_t)(b_ptr[l].e[j])] * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1229 | } |
| 1230 | } |
| 1231 | } |
| 1232 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 1233 | } |
| 1234 | } |
| 1235 | |
| 1236 | void ggml_gemv_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1237 | const int qk = QK8_032; |
| 1238 | const int nb = n / qk; |
| 1239 | const int ncols_interleaved = 8; |
| 1240 | const int blocklen = 8; |
| 1241 | |
| 1242 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1243 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1244 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1245 | |
| 1246 | UNUSED(bs)(void)(bs); |
| 1247 | UNUSED(nr)(void)(nr); |
| 1248 | |
| 1249 | float sumf[8]; |
| 1250 | int sumi; |
| 1251 | |
| 1252 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1253 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1254 | const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb); |
| 1255 | |
| 1256 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 1257 | for (int l = 0; l < nb; l++) { |
| 1258 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1259 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1260 | sumi = 0; |
| 1261 | for (int i = 0; i < blocklen; ++i) { |
| 1262 | const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 1263 | const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 1264 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); |
| 1265 | } |
| 1266 | sumf[j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j])ggml_table_f32_e8m0_half[(uint8_t)(b_ptr[l].e[j])] * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1267 | } |
| 1268 | } |
| 1269 | } |
| 1270 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 1271 | } |
| 1272 | } |
| 1273 | |
| 1274 | void ggml_gemv_q8_0_4x4_q8_0_genericggml_gemv_q8_0_4x4_q8_0(int n, |
| 1275 | float * GGML_RESTRICT__restrict__ s, |
| 1276 | size_t bs, |
| 1277 | const void * GGML_RESTRICT__restrict__ vx, |
| 1278 | const void * GGML_RESTRICT__restrict__ vy, |
| 1279 | int nr, |
| 1280 | int nc) { |
| 1281 | const int qk = QK8_032; |
| 1282 | const int nb = n / qk; |
| 1283 | const int ncols_interleaved = 4; |
| 1284 | const int blocklen = 4; |
| 1285 | |
| 1286 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1287 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1288 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1289 | |
| 1290 | UNUSED(bs)(void)(bs); |
| 1291 | UNUSED(nr)(void)(nr); |
| 1292 | |
| 1293 | float sumf[4]; |
| 1294 | int sumi; |
| 1295 | |
| 1296 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1297 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1298 | const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); |
| 1299 | |
| 1300 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1301 | sumf[j] = 0.0; |
| 1302 | } |
| 1303 | for (int l = 0; l < nb; l++) { |
| 1304 | for (int k = 0; k < (qk / blocklen); k++) { |
| 1305 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1306 | sumi = 0; |
| 1307 | for (int i = 0; i < blocklen; ++i) { |
| 1308 | const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; |
| 1309 | sumi += v0 * a_ptr[l].qs[k * blocklen + i]; |
| 1310 | } |
| 1311 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1312 | } |
| 1313 | } |
| 1314 | } |
| 1315 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1316 | s[x * ncols_interleaved + j] = sumf[j]; |
| 1317 | } |
| 1318 | } |
| 1319 | } |
| 1320 | |
| 1321 | void ggml_gemv_q8_0_4x8_q8_0_genericggml_gemv_q8_0_4x8_q8_0(int n, |
| 1322 | float * GGML_RESTRICT__restrict__ s, |
| 1323 | size_t bs, |
| 1324 | const void * GGML_RESTRICT__restrict__ vx, |
| 1325 | const void * GGML_RESTRICT__restrict__ vy, |
| 1326 | int nr, |
| 1327 | int nc) { |
| 1328 | const int qk = QK8_032; |
| 1329 | const int nb = n / qk; |
| 1330 | const int ncols_interleaved = 4; |
| 1331 | const int blocklen = 8; |
| 1332 | |
| 1333 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1334 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1335 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1336 | |
| 1337 | UNUSED(bs)(void)(bs); |
| 1338 | UNUSED(nr)(void)(nr); |
| 1339 | |
| 1340 | float sumf[4]; |
| 1341 | int sumi; |
| 1342 | |
| 1343 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1344 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1345 | const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); |
| 1346 | |
| 1347 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1348 | sumf[j] = 0.0; |
| 1349 | } |
| 1350 | for (int l = 0; l < nb; l++) { |
| 1351 | for (int k = 0; k < (qk / blocklen); k++) { |
| 1352 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1353 | sumi = 0; |
| 1354 | for (int i = 0; i < blocklen; ++i) { |
| 1355 | const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; |
| 1356 | sumi += v0 * a_ptr[l].qs[k * blocklen + i]; |
| 1357 | } |
| 1358 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1359 | } |
| 1360 | } |
| 1361 | } |
| 1362 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1363 | s[x * ncols_interleaved + j] = sumf[j]; |
| 1364 | } |
| 1365 | } |
| 1366 | } |
| 1367 | |
| 1368 | // Only enable these for RISC-V. |
| 1369 | #if defined __riscv_zvfh |
| 1370 | void ggml_gemv_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1371 | const int qk = QK8_032; |
| 1372 | const int nb = n / qk; |
| 1373 | const int ncols_interleaved = 16; |
| 1374 | const int blocklen = 1; |
| 1375 | |
| 1376 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1377 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1378 | |
| 1379 | UNUSED(s)(void)(s); |
| 1380 | UNUSED(bs)(void)(bs); |
| 1381 | UNUSED(vx)(void)(vx); |
| 1382 | UNUSED(vy)(void)(vy); |
| 1383 | UNUSED(nr)(void)(nr); |
| 1384 | UNUSED(nc)(void)(nc); |
| 1385 | UNUSED(nb)(void)(nb); |
| 1386 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1387 | UNUSED(blocklen)(void)(blocklen); |
| 1388 | |
| 1389 | float sumf[16]; |
| 1390 | int sumi; |
| 1391 | |
| 1392 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1393 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1394 | const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb); |
| 1395 | |
| 1396 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 1397 | for (int l = 0; l < nb; l++) { |
| 1398 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1399 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1400 | sumi = 0; |
| 1401 | for (int i = 0; i < blocklen; ++i) { |
| 1402 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 1403 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 1404 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4; |
| 1405 | } |
| 1406 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1407 | } |
| 1408 | } |
| 1409 | } |
| 1410 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 1411 | } |
| 1412 | } |
| 1413 | |
| 1414 | void ggml_gemv_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1415 | const int qk = QK_K256; |
| 1416 | const int nb = n / qk; |
| 1417 | const int ncols_interleaved = 16; |
| 1418 | const int blocklen = 1; |
| 1419 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1420 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1421 | UNUSED(s)(void)(s); |
| 1422 | UNUSED(bs)(void)(bs); |
| 1423 | UNUSED(vx)(void)(vx); |
| 1424 | UNUSED(vy)(void)(vy); |
| 1425 | UNUSED(nr)(void)(nr); |
| 1426 | UNUSED(nc)(void)(nc); |
| 1427 | UNUSED(nb)(void)(nb); |
| 1428 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1429 | UNUSED(blocklen)(void)(blocklen); |
| 1430 | float sumf[16]; |
| 1431 | float sum_minf[16]; |
| 1432 | uint8_t scales[128]; |
| 1433 | uint8_t mins[128]; |
| 1434 | int sumi1; |
| 1435 | int sumi2; |
| 1436 | int sumi; |
| 1437 | const block_q8_K * a_ptr = (const block_q8_K *) vy; |
| 1438 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1439 | const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb); |
| 1440 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1441 | sumf[j] = 0.0f; |
| 1442 | sum_minf[j] = 0.0f; |
| 1443 | } |
| 1444 | for (int l = 0; l < nb; l++) { |
| 1445 | for (int i = 0; i < 128; i++) { |
| 1446 | scales[i] = b_ptr[l].scales[i] & 0x0F; |
| 1447 | mins[i] = b_ptr[l].scales[i] >> 4; |
| 1448 | } |
| 1449 | for (int i = 0; i < 64; i++) { |
| 1450 | scales[i] |= (b_ptr[l].scales[128 + i] & 0x03) << 4; |
| 1451 | mins[i] |= (b_ptr[l].scales[128 + i] & 0x0C) << 2; |
| 1452 | scales[i + 64] |= (b_ptr[l].scales[128 + i] & 0x30); |
| 1453 | mins[i + 64] |= (b_ptr[l].scales[128 + i] & 0xC0) >> 2; |
| 1454 | } |
| 1455 | for (int sb = 0; sb < 8; sb++) { |
| 1456 | uint8_t *min = &mins[sb * 16]; |
| 1457 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1458 | sum_minf[j] += min[j] * (a_ptr[l].bsums[sb * 2] + a_ptr[l].bsums[sb * 2 + 1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d; |
| 1459 | } |
| 1460 | } |
| 1461 | for (int sb = 0; sb < 8; sb += 2) { |
| 1462 | uint8_t *scales_0 = &scales[sb * 16]; |
| 1463 | uint8_t *scales_1 = &scales[(sb + 1) * 16]; |
| 1464 | for (int i = 0; i < QK4_032; i++) { |
| 1465 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1466 | sumi1 = 0; |
| 1467 | sumi2 = 0; |
| 1468 | sumi = 0; |
| 1469 | const int v0 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] & 0xF); |
| 1470 | const int v1 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] >> 4); |
| 1471 | sumi1 = (v0 * a_ptr[l].qs[sb * 32 + i]); |
| 1472 | sumi2 = (v1 * a_ptr[l].qs[sb * 32 + 32 + i]); |
| 1473 | sumi1 = sumi1 * scales_0[j]; |
| 1474 | sumi2 = sumi2 * scales_1[j]; |
| 1475 | sumi += sumi1 + sumi2; |
| 1476 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d; |
| 1477 | } |
| 1478 | } |
| 1479 | } |
| 1480 | } |
| 1481 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1482 | s[x * ncols_interleaved + j] = sumf[j] - sum_minf[j]; |
| 1483 | } |
| 1484 | } |
| 1485 | } |
| 1486 | |
| 1487 | void ggml_gemv_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1488 | const int qk = QK8_032; |
| 1489 | const int nb = n / qk; |
| 1490 | const int ncols_interleaved = 16; |
| 1491 | const int blocklen = 1; |
| 1492 | |
| 1493 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1494 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1495 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1496 | |
| 1497 | UNUSED(bs)(void)(bs); |
| 1498 | UNUSED(nr)(void)(nr); |
| 1499 | |
| 1500 | float sumf[16]; |
| 1501 | int sumi; |
| 1502 | |
| 1503 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1504 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1505 | const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb); |
| 1506 | |
| 1507 | for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0; |
| 1508 | for (int l = 0; l < nb; l++) { |
| 1509 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1510 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1511 | sumi = 0; |
| 1512 | for (int i = 0; i < blocklen; ++i) { |
| 1513 | const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 1514 | const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 1515 | sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])); |
| 1516 | } |
| 1517 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1518 | } |
| 1519 | } |
| 1520 | } |
| 1521 | for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j]; |
| 1522 | } |
| 1523 | } |
| 1524 | |
| 1525 | void ggml_gemv_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1526 | const int qk = QK8_032; |
| 1527 | const int nb = n / qk; |
| 1528 | const int ncols_interleaved = 16; |
| 1529 | const int blocklen = 1; |
| 1530 | |
| 1531 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1532 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1533 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1534 | |
| 1535 | UNUSED(bs)(void)(bs); |
| 1536 | UNUSED(nr)(void)(nr); |
| 1537 | |
| 1538 | float sumf[16]; |
| 1539 | int sumi; |
| 1540 | |
| 1541 | const block_q8_0 * a_ptr = (const block_q8_0 *) vy; |
| 1542 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1543 | const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb); |
| 1544 | |
| 1545 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1546 | sumf[j] = 0.0; |
| 1547 | } |
| 1548 | for (int l = 0; l < nb; l++) { |
| 1549 | for (int k = 0; k < (qk / blocklen); k++) { |
| 1550 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1551 | sumi = 0; |
| 1552 | for (int i = 0; i < blocklen; ++i) { |
| 1553 | const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; |
| 1554 | sumi += v0 * a_ptr[l].qs[k * blocklen + i]; |
| 1555 | } |
| 1556 | sumf[j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d)ggml_lookup_fp16_to_fp32(a_ptr[l].d); |
| 1557 | } |
| 1558 | } |
| 1559 | } |
| 1560 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1561 | s[x * ncols_interleaved + j] = sumf[j]; |
| 1562 | } |
| 1563 | } |
| 1564 | } |
| 1565 | |
| 1566 | void ggml_gemv_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1567 | assert(n % QK_K == 0)(static_cast <bool> (n % 256 == 0) ? void (0) : __assert_fail ("n % QK_K == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1568 | assert(nr == 1)(static_cast <bool> (nr == 1) ? void (0) : __assert_fail ("nr == 1", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1569 | assert(nc % 16 == 0)(static_cast <bool> (nc % 16 == 0) ? void (0) : __assert_fail ("nc % 16 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1570 | |
| 1571 | UNUSED(bs)(void)(bs); |
| 1572 | UNUSED(nr)(void)(nr); |
| 1573 | |
| 1574 | const int nb = n / QK_K256; |
| 1575 | const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx; |
| 1576 | const block_q8_K * y = (const block_q8_K *)vy; |
| 1577 | |
| 1578 | // Layout: Even-Low(0,2,4,6), Odd-Low(1,3,5,7), Even-High(8...), Odd-High(9...) |
| 1579 | const int sb_perm[16] = { |
| 1580 | 0, 4, 1, 5, 2, 6, 3, 7, // 0-7 |
| 1581 | 8, 12, 9, 13, 10, 14, 11, 15 // 8-15 |
| 1582 | }; |
| 1583 | |
| 1584 | for (int col_tile = 0; col_tile < nc; col_tile += 16) { |
| 1585 | const block_q2_Kx16 * x_ptr = x + (col_tile / 16) * nb; |
| 1586 | const block_q8_K * y_ptr = y; |
| 1587 | |
| 1588 | float sumf[16] = {0}; |
| 1589 | |
| 1590 | // Loop over K-blocks |
| 1591 | for (int k_block = 0; k_block < nb; ++k_block) { |
| 1592 | int32_t isum[16] = {0}; |
| 1593 | int32_t summs[16] = {0}; |
| 1594 | |
| 1595 | const uint8_t * qs_rhs = x_ptr[k_block].qs; |
| 1596 | const uint8_t * sc_rhs = x_ptr[k_block].scales; |
| 1597 | const int8_t * qs_lhs = y_ptr[k_block].qs; |
| 1598 | const int16_t * bs_lhs = y_ptr[k_block].bsums; |
| 1599 | |
| 1600 | // Iterate over sub-blocks 0..15 |
| 1601 | for (int sb = 0; sb < 16; ++sb) { |
| 1602 | // Correction Term |
| 1603 | int16_t bsum = bs_lhs[sb]; |
| 1604 | int scale_offset = sb_perm[sb] * 16; |
| 1605 | |
| 1606 | for (int col = 0; col < 16; ++col) { |
| 1607 | uint8_t sc_val = sc_rhs[scale_offset + col]; |
| 1608 | summs[col] += bsum * (sc_val >> 4); // Min is high 4 bits |
| 1609 | } |
| 1610 | |
| 1611 | // Main Dot Product |
| 1612 | // Calculate base offsets for Q2 unpacking based on SB |
| 1613 | int byte_base; |
| 1614 | if (sb < 8) byte_base = (sb % 2 == 0) ? 0 : 16; |
| 1615 | else byte_base = (sb % 2 == 0) ? 32 : 48; |
| 1616 | |
| 1617 | int shift = ((sb / 2) % 4) * 2; |
| 1618 | |
| 1619 | for (int col = 0; col < 16; ++col) { |
| 1620 | uint8_t sc_val = sc_rhs[scale_offset + col]; |
| 1621 | int32_t d_sb = sc_val & 0xF; // Scale is low 4 bits |
| 1622 | |
| 1623 | // Process 16 elements (l=0..15) |
| 1624 | for (int l = 0; l < 16; ++l) { |
| 1625 | // Q2: Interleaved by column. Byte `l` contains 4 k-values. |
| 1626 | int qs_idx = (byte_base + l) * 16 + col; |
| 1627 | uint8_t q2_val = (qs_rhs[qs_idx] >> shift) & 3; |
| 1628 | |
| 1629 | // Q8: Linear access |
| 1630 | int k = sb * 16 + l; |
| 1631 | int8_t q8_val = qs_lhs[k]; |
| 1632 | |
| 1633 | isum[col] += q8_val * q2_val * d_sb; |
| 1634 | } |
| 1635 | } |
| 1636 | } |
| 1637 | |
| 1638 | // Finalize K-Block |
| 1639 | for (int col = 0; col < 16; ++col) { |
| 1640 | float d_lhs = y_ptr[k_block].d; |
| 1641 | float d_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].d[col])ggml_compute_fp16_to_fp32(x_ptr[k_block].d[col]); |
| 1642 | float dm_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].dmin[col])ggml_compute_fp16_to_fp32(x_ptr[k_block].dmin[col]); |
| 1643 | |
| 1644 | float d_all = d_lhs * d_rhs; |
| 1645 | float d_min = d_lhs * dm_rhs; |
| 1646 | |
| 1647 | sumf[col] += (isum[col] * d_all) - (summs[col] * d_min); |
| 1648 | } |
| 1649 | } |
| 1650 | |
| 1651 | for (int col = 0; col < 16; ++col) { |
| 1652 | s[col_tile + col] = sumf[col]; |
| 1653 | } |
| 1654 | } |
| 1655 | } |
| 1656 | #endif |
| 1657 | |
| 1658 | void ggml_gemm_q4_0_4x4_q8_0_genericggml_gemm_q4_0_4x4_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1659 | const int qk = QK8_032; |
| 1660 | const int nb = n / qk; |
| 1661 | const int ncols_interleaved = 4; |
| 1662 | const int blocklen = 4; |
| 1663 | |
| 1664 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1665 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1666 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1667 | |
| 1668 | UNUSED(s)(void)(s); |
| 1669 | UNUSED(bs)(void)(bs); |
| 1670 | UNUSED(vx)(void)(vx); |
| 1671 | UNUSED(vy)(void)(vy); |
| 1672 | UNUSED(nr)(void)(nr); |
| 1673 | UNUSED(nc)(void)(nc); |
| 1674 | UNUSED(nb)(void)(nb); |
| 1675 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1676 | UNUSED(blocklen)(void)(blocklen); |
| 1677 | |
| 1678 | { |
| 1679 | float sumf[4][4]; |
| 1680 | int sumi; |
| 1681 | |
| 1682 | for (int y = 0; y < nr / 4; y++) { |
| 1683 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 1684 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1685 | const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); |
| 1686 | for (int m = 0; m < 4; m++) { |
| 1687 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 1688 | } |
| 1689 | for (int l = 0; l < nb; l++) { |
| 1690 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1691 | for (int m = 0; m < 4; m++) { |
| 1692 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1693 | sumi = 0; |
| 1694 | for (int i = 0; i < blocklen; ++i) { |
| 1695 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 1696 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 1697 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 1698 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; |
| 1699 | } |
| 1700 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 1701 | } |
| 1702 | } |
| 1703 | } |
| 1704 | } |
| 1705 | for (int m = 0; m < 4; m++) { |
| 1706 | for (int j = 0; j < ncols_interleaved; j++) |
| 1707 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 1708 | } |
| 1709 | } |
| 1710 | } |
| 1711 | } |
| 1712 | } |
| 1713 | |
| 1714 | void ggml_gemm_q4_0_4x8_q8_0_genericggml_gemm_q4_0_4x8_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1715 | const int qk = QK8_032; |
| 1716 | const int nb = n / qk; |
| 1717 | const int ncols_interleaved = 4; |
| 1718 | const int blocklen = 8; |
| 1719 | |
| 1720 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1721 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1722 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1723 | |
| 1724 | UNUSED(s)(void)(s); |
| 1725 | UNUSED(bs)(void)(bs); |
| 1726 | UNUSED(vx)(void)(vx); |
| 1727 | UNUSED(vy)(void)(vy); |
| 1728 | UNUSED(nr)(void)(nr); |
| 1729 | UNUSED(nc)(void)(nc); |
| 1730 | UNUSED(nb)(void)(nb); |
| 1731 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1732 | UNUSED(blocklen)(void)(blocklen); |
| 1733 | |
| 1734 | float sumf[4][4]; |
| 1735 | int sumi; |
| 1736 | |
| 1737 | for (int y = 0; y < nr / 4; y++) { |
| 1738 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 1739 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1740 | const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb); |
| 1741 | for (int m = 0; m < 4; m++) { |
| 1742 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 1743 | } |
| 1744 | for (int l = 0; l < nb; l++) { |
| 1745 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1746 | for (int m = 0; m < 4; m++) { |
| 1747 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1748 | sumi = 0; |
| 1749 | for (int i = 0; i < blocklen; ++i) { |
| 1750 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 1751 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 1752 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 1753 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; |
| 1754 | } |
| 1755 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 1756 | } |
| 1757 | } |
| 1758 | } |
| 1759 | } |
| 1760 | for (int m = 0; m < 4; m++) { |
| 1761 | for (int j = 0; j < ncols_interleaved; j++) |
| 1762 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 1763 | } |
| 1764 | } |
| 1765 | } |
| 1766 | } |
| 1767 | |
| 1768 | void ggml_gemm_q4_0_8x8_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1769 | const int qk = QK8_032; |
| 1770 | const int nb = n / qk; |
| 1771 | const int ncols_interleaved = 8; |
| 1772 | const int blocklen = 8; |
| 1773 | |
| 1774 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1775 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1776 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1777 | |
| 1778 | UNUSED(s)(void)(s); |
| 1779 | UNUSED(bs)(void)(bs); |
| 1780 | UNUSED(vx)(void)(vx); |
| 1781 | UNUSED(vy)(void)(vy); |
| 1782 | UNUSED(nr)(void)(nr); |
| 1783 | UNUSED(nc)(void)(nc); |
| 1784 | UNUSED(nb)(void)(nb); |
| 1785 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1786 | UNUSED(blocklen)(void)(blocklen); |
| 1787 | |
| 1788 | float sumf[4][8]; |
| 1789 | int sumi; |
| 1790 | |
| 1791 | for (int y = 0; y < nr / 4; y++) { |
| 1792 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 1793 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1794 | const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb); |
| 1795 | for (int m = 0; m < 4; m++) { |
| 1796 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 1797 | } |
| 1798 | for (int l = 0; l < nb; l++) { |
| 1799 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1800 | for (int m = 0; m < 4; m++) { |
| 1801 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1802 | sumi = 0; |
| 1803 | for (int i = 0; i < blocklen; ++i) { |
| 1804 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 1805 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 1806 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 1807 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; |
| 1808 | } |
| 1809 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 1810 | } |
| 1811 | } |
| 1812 | } |
| 1813 | } |
| 1814 | for (int m = 0; m < 4; m++) { |
| 1815 | for (int j = 0; j < ncols_interleaved; j++) |
| 1816 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 1817 | } |
| 1818 | } |
| 1819 | } |
| 1820 | } |
| 1821 | |
| 1822 | void ggml_gemm_q4_K_8x4_q8_K_genericggml_gemm_q4_K_8x4_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1823 | const int qk = QK_K256; |
| 1824 | const int nb = n / qk; |
| 1825 | const int ncols_interleaved = 8; |
| 1826 | const int blocklen = 4; |
| 1827 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 1828 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 1829 | static const uint32_t kmask3 = 0x03030303; |
| 1830 | |
| 1831 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1832 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1833 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1834 | |
| 1835 | UNUSED(nb)(void)(nb); |
| 1836 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 1837 | UNUSED(blocklen)(void)(blocklen); |
| 1838 | |
| 1839 | float sumf[4][8]; |
| 1840 | float sum_minf[4][8]; |
| 1841 | uint32_t utmp[32]; |
| 1842 | int sumi1; |
| 1843 | int sumi2; |
| 1844 | int sumi; |
| 1845 | |
| 1846 | for (int y = 0; y < nr / 4; y++) { |
| 1847 | const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); |
| 1848 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1849 | const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); |
| 1850 | for (int m = 0; m < 4; m++) { |
| 1851 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1852 | sumf[m][j] = 0.0; |
| 1853 | sum_minf[m][j] = 0.0; |
| 1854 | } |
| 1855 | } |
| 1856 | for (int l = 0; l < nb; l++) { |
| 1857 | for (int sb = 0; sb < 8; sb++) { |
| 1858 | memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); |
| 1859 | utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); |
| 1860 | const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; |
| 1861 | utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); |
| 1862 | utmp[sb * 4 + 2] = uaux_0; |
| 1863 | utmp[sb * 4 + 0] &= kmask1; |
| 1864 | } |
| 1865 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1866 | uint8_t * scales_0 = (uint8_t *) utmp + (k / 8) * 32; |
| 1867 | uint8_t * scales_1 = (uint8_t *) utmp + (k / 8) * 32 + 16; |
| 1868 | for (int m = 0; m < 4; m++) { |
| 1869 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1870 | sumi1 = 0; |
| 1871 | sumi2 = 0; |
| 1872 | sumi = 0; |
| 1873 | for (int i = 0; i < blocklen; ++i) { |
| 1874 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); |
| 1875 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); |
| 1876 | sumi1 = (v0 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i]); |
| 1877 | sumi2 = (v1 * a_ptr[l].qs[(k / 8) * 256 + (k % 8) * 4 * blocklen + m * blocklen + i + 128]); |
| 1878 | sumi1 = sumi1 * scales_0[j]; |
| 1879 | sumi2 = sumi2 * scales_1[j]; |
| 1880 | sumi += sumi1 + sumi2; |
| 1881 | } |
| 1882 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d[m]; |
| 1883 | } |
| 1884 | } |
| 1885 | } |
| 1886 | for (int sb = 0; sb < 8; sb++) { |
| 1887 | uint8_t * mins = (uint8_t *) utmp + 8 + sb * 16; |
| 1888 | for(int m = 0; m < 4; m++) { |
| 1889 | const int16_t * bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); |
| 1890 | for(int j = 0; j < ncols_interleaved; j++) { |
| 1891 | sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; |
| 1892 | } |
| 1893 | } |
| 1894 | } |
| 1895 | } |
| 1896 | for (int m = 0; m < 4; m++) { |
| 1897 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1898 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; |
| 1899 | } |
| 1900 | } |
| 1901 | } |
| 1902 | } |
| 1903 | } |
| 1904 | |
| 1905 | void ggml_gemm_q4_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1906 | const int qk = QK_K256; |
| 1907 | const int nb = n / qk; |
| 1908 | const int ncols_interleaved = 8; |
| 1909 | const int blocklen = 8; |
| 1910 | static const uint32_t kmask1 = 0x3f3f3f3f; |
| 1911 | static const uint32_t kmask2 = 0x0f0f0f0f; |
| 1912 | static const uint32_t kmask3 = 0x03030303; |
| 1913 | |
| 1914 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1915 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1916 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1917 | |
| 1918 | UNUSED(bs)(void)(bs); |
| 1919 | |
| 1920 | float sumf[4][8]; |
| 1921 | float sum_minf[4][8]; |
| 1922 | uint32_t utmp[32]; |
| 1923 | int sumi1; |
| 1924 | int sumi2; |
| 1925 | int sumi; |
| 1926 | |
| 1927 | for (int y = 0; y < nr / 4; y++) { |
| 1928 | const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); |
| 1929 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 1930 | const block_q4_Kx8 * b_ptr = (const block_q4_Kx8 *) vx + (x * nb); |
| 1931 | for (int m = 0; m < 4; m++) { |
| 1932 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1933 | sumf[m][j] = 0.0; |
| 1934 | sum_minf[m][j] = 0.0; |
| 1935 | } |
| 1936 | } |
| 1937 | for (int l = 0; l < nb; l++) { |
| 1938 | for (int sb = 0; sb < 8; sb++) { |
| 1939 | memcpy(utmp + sb * 4, b_ptr[l].scales + sb * 12, 12); |
| 1940 | utmp[sb * 4 + 3] = ((utmp[sb * 4 + 2] >> 4) & kmask2) | (((utmp[sb * 4 + 1] >> 6) & kmask3) << 4); |
| 1941 | const uint32_t uaux_0 = utmp[sb * 4 + 1] & kmask1; |
| 1942 | utmp[sb * 4 + 1] = (utmp[sb * 4 + 2] & kmask2) | (((utmp[sb * 4 + 0] >> 6) & kmask3) << 4); |
| 1943 | utmp[sb * 4 + 2] = uaux_0; |
| 1944 | utmp[sb * 4 + 0] &= kmask1; |
| 1945 | } |
| 1946 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 1947 | uint8_t *scales_0 = (uint8_t*) utmp + (k / 4) * 32; |
| 1948 | uint8_t *scales_1 = (uint8_t*) utmp + (k / 4) * 32 + 16; |
| 1949 | for (int m = 0; m < 4; m++) { |
| 1950 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1951 | sumi1 = 0; |
| 1952 | sumi2 = 0; |
| 1953 | sumi = 0; |
| 1954 | for (int i = 0; i < blocklen; ++i) { |
| 1955 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF); |
| 1956 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4); |
| 1957 | sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i]); |
| 1958 | sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 256 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]); |
| 1959 | sumi1 = sumi1 * scales_0[j]; |
| 1960 | sumi2 = sumi2 * scales_1[j]; |
| 1961 | sumi += sumi1 + sumi2; |
| 1962 | } |
| 1963 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d[m]; |
| 1964 | } |
| 1965 | } |
| 1966 | } |
| 1967 | for (int sb = 0; sb < 8; sb++) { |
| 1968 | uint8_t *mins = (uint8_t*) utmp + 8 + sb * 16; |
| 1969 | for(int m = 0; m < 4; m++) { |
| 1970 | const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); |
| 1971 | for(int j = 0; j < ncols_interleaved; j++) { |
| 1972 | sum_minf[m][j] += mins[j] * (bsums[0] + bsums[1]) * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; |
| 1973 | } |
| 1974 | } |
| 1975 | } |
| 1976 | } |
| 1977 | for (int m = 0; m < 4; m++) { |
| 1978 | for (int j = 0; j < ncols_interleaved; j++) { |
| 1979 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; |
| 1980 | } |
| 1981 | } |
| 1982 | } |
| 1983 | } |
| 1984 | } |
| 1985 | |
| 1986 | void ggml_gemm_q2_K_8x8_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 1987 | const int qk = QK_K256; |
| 1988 | const int nb = n / qk; |
| 1989 | const int ncols_interleaved = 8; |
| 1990 | const int blocklen = 8; |
| 1991 | |
| 1992 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1993 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1994 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 1995 | |
| 1996 | UNUSED(s)(void)(s); |
| 1997 | UNUSED(bs)(void)(bs); |
| 1998 | UNUSED(vx)(void)(vx); |
| 1999 | UNUSED(vy)(void)(vy); |
| 2000 | UNUSED(nr)(void)(nr); |
| 2001 | UNUSED(nc)(void)(nc); |
| 2002 | UNUSED(nb)(void)(nb); |
| 2003 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 2004 | UNUSED(blocklen)(void)(blocklen); |
| 2005 | |
| 2006 | float sumf[4][8]; |
| 2007 | float sum_minf[4][8]; |
| 2008 | int sumi1, sumi2, sumi3, sumi4; |
| 2009 | int sumi; |
| 2010 | |
| 2011 | for (int y = 0; y < nr / 4; y++) { |
| 2012 | const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); |
| 2013 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2014 | const block_q2_Kx8 * b_ptr = (const block_q2_Kx8 *) vx + (x * nb); |
| 2015 | for (int m = 0; m < 4; m++) { |
| 2016 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2017 | sumf[m][j] = 0.0; |
| 2018 | sum_minf[m][j] = 0.0; |
| 2019 | } |
| 2020 | } |
| 2021 | for (int l = 0; l < nb; l++) { |
| 2022 | for (int k = 0; k < (qk / (4 * blocklen)); k++) { |
| 2023 | |
| 2024 | const uint8_t *scales_0 = b_ptr[l].scales + (k / 4) * 64 ; |
| 2025 | const uint8_t *scales_1 = b_ptr[l].scales + (k / 4) * 64 + 16; |
| 2026 | const uint8_t *scales_2 = b_ptr[l].scales + (k / 4) * 64 + 32; |
| 2027 | const uint8_t *scales_3 = b_ptr[l].scales + (k / 4) * 64 + 48; |
| 2028 | for (int m = 0; m < 4; m++) { |
| 2029 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2030 | sumi1 = 0; |
| 2031 | sumi2 = 0; |
| 2032 | sumi3 = 0; |
| 2033 | sumi4 = 0; |
| 2034 | sumi = 0; |
| 2035 | int offset = ((k / 2) % 2) + j * 2; |
| 2036 | for (int i = 0; i < blocklen; ++i){ |
| 2037 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 3); |
| 2038 | const int v1 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 2 ) & 3); |
| 2039 | const int v2 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4 ) & 3); |
| 2040 | const int v3 = (int8_t) ((b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 6 ) & 3); |
| 2041 | sumi1 = (v0 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i]); |
| 2042 | sumi2 = (v1 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 128]); |
| 2043 | sumi3 = (v2 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 256]); |
| 2044 | sumi4 = (v3 * a_ptr[l].qs[(k >> 2) * 512 + (k % 4) * 4 * blocklen + m * blocklen + i + 384]); |
| 2045 | sumi1 = sumi1 * (scales_0[offset] & 0xF); |
| 2046 | sumi2 = sumi2 * (scales_1[offset] & 0xF); |
| 2047 | sumi3 = sumi3 * (scales_2[offset] & 0xF); |
| 2048 | sumi4 = sumi4 * (scales_3[offset] & 0xF); |
| 2049 | sumi += sumi1 + sumi2 + sumi3 + sumi4; |
| 2050 | } |
| 2051 | sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j])ggml_compute_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d[m]; |
| 2052 | } |
| 2053 | } |
| 2054 | } |
| 2055 | for(int sb = 0; sb < 8; sb++) { |
| 2056 | const uint8_t *mins = b_ptr[l].scales + sb * 16; |
| 2057 | for(int m = 0; m < 4; m++) { |
| 2058 | const int16_t *bsums = a_ptr[l].bsums + (sb * 8) + (m * 4) - ((sb % 2) * 6); |
| 2059 | for(int j = 0; j < ncols_interleaved; j++) { |
| 2060 | int mins_prod = ((mins[j * 2] >> 4) * bsums[0] + (mins[(j * 2)+ 1] >> 4) * bsums[1]); |
| 2061 | sum_minf[m][j] += (mins_prod) * GGML_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_compute_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; |
| 2062 | } |
| 2063 | } |
| 2064 | } |
| 2065 | } |
| 2066 | |
| 2067 | for (int m = 0; m < 4; m++) { |
| 2068 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2069 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; |
| 2070 | } |
| 2071 | } |
| 2072 | } |
| 2073 | } |
| 2074 | } |
| 2075 | |
| 2076 | void ggml_gemm_q5_K_8x4_q8_K_genericggml_gemm_q5_K_8x4_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2077 | ggml_gemm_q5_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc); |
| 2078 | } |
| 2079 | |
| 2080 | void ggml_gemm_q5_K_8x8_q8_K_genericggml_gemm_q5_K_8x8_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2081 | ggml_gemm_q5_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc); |
| 2082 | } |
| 2083 | |
| 2084 | void ggml_gemm_q6_K_8x4_q8_K_genericggml_gemm_q6_K_8x4_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2085 | ggml_gemm_q6_K_NxM_q8_K_generic_impl<4, 8>(n, s, bs, vx, vy, nr, nc); |
| 2086 | } |
| 2087 | |
| 2088 | void ggml_gemm_q6_K_8x8_q8_K_genericggml_gemm_q6_K_8x8_q8_K(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2089 | ggml_gemm_q6_K_NxM_q8_K_generic_impl<8, 8>(n, s, bs, vx, vy, nr, nc); |
| 2090 | } |
| 2091 | |
| 2092 | void ggml_gemm_iq4_nl_4x4_q8_0_genericggml_gemm_iq4_nl_4x4_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2093 | const int qk = QK8_032; |
| 2094 | const int nb = n / qk; |
| 2095 | const int ncols_interleaved = 4; |
| 2096 | const int blocklen = 4; |
| 2097 | |
| 2098 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2099 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2100 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2101 | |
| 2102 | UNUSED(s)(void)(s); |
| 2103 | UNUSED(bs)(void)(bs); |
| 2104 | UNUSED(vx)(void)(vx); |
| 2105 | UNUSED(vy)(void)(vy); |
| 2106 | UNUSED(nr)(void)(nr); |
| 2107 | UNUSED(nc)(void)(nc); |
| 2108 | UNUSED(nb)(void)(nb); |
| 2109 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 2110 | UNUSED(blocklen)(void)(blocklen); |
| 2111 | |
| 2112 | { |
| 2113 | float sumf[4][4]; |
| 2114 | int sumi; |
| 2115 | |
| 2116 | for (int y = 0; y < nr / 4; y++) { |
| 2117 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2118 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2119 | const block_iq4_nlx4 * b_ptr = (const block_iq4_nlx4 *) vx + (x * nb); |
| 2120 | for (int m = 0; m < 4; m++) { |
| 2121 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 2122 | } |
| 2123 | for (int l = 0; l < nb; l++) { |
| 2124 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 2125 | for (int m = 0; m < 4; m++) { |
| 2126 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2127 | sumi = 0; |
| 2128 | for (int i = 0; i < blocklen; ++i) { |
| 2129 | const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 2130 | const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 2131 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 2132 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); |
| 2133 | } |
| 2134 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2135 | } |
| 2136 | } |
| 2137 | } |
| 2138 | } |
| 2139 | for (int m = 0; m < 4; m++) { |
| 2140 | for (int j = 0; j < ncols_interleaved; j++) |
| 2141 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2142 | } |
| 2143 | } |
| 2144 | } |
| 2145 | } |
| 2146 | } |
| 2147 | |
| 2148 | void ggml_gemm_iq4_nl_8x8_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2149 | const int qk = QK8_032; |
| 2150 | const int nb = n / qk; |
| 2151 | const int ncols_interleaved = 8; |
| 2152 | const int blocklen = 8; |
| 2153 | |
| 2154 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2155 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2156 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2157 | |
| 2158 | float sumf[4][8]; |
| 2159 | int sumi; |
| 2160 | |
| 2161 | for (int y = 0; y < nr / 4; y++) { |
| 2162 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2163 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2164 | const block_iq4_nlx8 * b_ptr = (const block_iq4_nlx8 *) vx + (x * nb); |
| 2165 | for (int m = 0; m < 4; m++) { |
| 2166 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 2167 | } |
| 2168 | for (int l = 0; l < nb; l++) { |
| 2169 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 2170 | for (int m = 0; m < 4; m++) { |
| 2171 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2172 | sumi = 0; |
| 2173 | for (int i = 0; i < blocklen; ++i) { |
| 2174 | const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 2175 | const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 2176 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 2177 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); |
| 2178 | } |
| 2179 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2180 | } |
| 2181 | } |
| 2182 | } |
| 2183 | } |
| 2184 | for (int m = 0; m < 4; m++) { |
| 2185 | for (int j = 0; j < ncols_interleaved; j++) |
| 2186 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2187 | } |
| 2188 | } |
| 2189 | } |
| 2190 | } |
| 2191 | |
| 2192 | void ggml_gemm_mxfp4_4x4_q8_0_genericggml_gemm_mxfp4_4x4_q8_0(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2193 | const int qk = QK8_032; |
| 2194 | const int nb = n / qk; |
| 2195 | const int ncols_interleaved = 4; |
| 2196 | const int blocklen = 4; |
| 2197 | |
| 2198 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2199 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2200 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2201 | |
| 2202 | float sumf[4][4]; |
| 2203 | int sumi; |
| 2204 | |
| 2205 | for (int y = 0; y < nr / 4; y++) { |
| 2206 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2207 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2208 | const block_mxfp4x4 * b_ptr = (const block_mxfp4x4 *) vx + (x * nb); |
| 2209 | for (int m = 0; m < 4; m++) { |
| 2210 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 2211 | } |
| 2212 | for (int l = 0; l < nb; l++) { |
| 2213 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 2214 | for (int m = 0; m < 4; m++) { |
| 2215 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2216 | sumi = 0; |
| 2217 | for (int i = 0; i < blocklen; ++i) { |
| 2218 | const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 2219 | const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 2220 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 2221 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); |
| 2222 | } |
| 2223 | sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j])ggml_table_f32_e8m0_half[(uint8_t)(b_ptr[l].e[j])] * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2224 | } |
| 2225 | } |
| 2226 | } |
| 2227 | } |
| 2228 | for (int m = 0; m < 4; m++) { |
| 2229 | for (int j = 0; j < ncols_interleaved; j++) |
| 2230 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2231 | } |
| 2232 | } |
| 2233 | } |
| 2234 | } |
| 2235 | |
| 2236 | void ggml_gemm_mxfp4_8x8_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2237 | const int qk = QK8_032; |
| 2238 | const int nb = n / qk; |
| 2239 | const int ncols_interleaved = 8; |
| 2240 | const int blocklen = 8; |
| 2241 | |
| 2242 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2243 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2244 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2245 | |
| 2246 | float sumf[4][8]; |
| 2247 | int sumi; |
| 2248 | |
| 2249 | for (int y = 0; y < nr / 4; y++) { |
| 2250 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2251 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2252 | const block_mxfp4x8 * b_ptr = (const block_mxfp4x8 *) vx + (x * nb); |
| 2253 | for (int m = 0; m < 4; m++) { |
| 2254 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 2255 | } |
| 2256 | for (int l = 0; l < nb; l++) { |
| 2257 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 2258 | for (int m = 0; m < 4; m++) { |
| 2259 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2260 | sumi = 0; |
| 2261 | for (int i = 0; i < blocklen; ++i) { |
| 2262 | const int v0 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 2263 | const int v1 = kvalues_mxfp4[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 2264 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 2265 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])); |
| 2266 | } |
| 2267 | sumf[m][j] += sumi * GGML_CPU_E8M0_TO_FP32_HALF(b_ptr[l].e[j])ggml_table_f32_e8m0_half[(uint8_t)(b_ptr[l].e[j])] * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2268 | } |
| 2269 | } |
| 2270 | } |
| 2271 | } |
| 2272 | for (int m = 0; m < 4; m++) { |
| 2273 | for (int j = 0; j < ncols_interleaved; j++) |
| 2274 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2275 | } |
| 2276 | } |
| 2277 | } |
| 2278 | } |
| 2279 | |
| 2280 | void ggml_gemm_q8_0_4x4_q8_0_genericggml_gemm_q8_0_4x4_q8_0(int n, |
| 2281 | float * GGML_RESTRICT__restrict__ s, |
| 2282 | size_t bs, |
| 2283 | const void * GGML_RESTRICT__restrict__ vx, |
| 2284 | const void * GGML_RESTRICT__restrict__ vy, |
| 2285 | int nr, |
| 2286 | int nc) { |
| 2287 | const int qk = QK8_032; |
| 2288 | const int nb = n / qk; |
| 2289 | const int ncols_interleaved = 4; |
| 2290 | const int blocklen = 4; |
| 2291 | |
| 2292 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2293 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2294 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2295 | |
| 2296 | float sumf[4][4]; |
| 2297 | int sumi; |
| 2298 | |
| 2299 | for (int y = 0; y < nr / 4; y++) { |
| 2300 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2301 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2302 | const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); |
| 2303 | for (int m = 0; m < 4; m++) { |
| 2304 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2305 | sumf[m][j] = 0.0; |
| 2306 | } |
| 2307 | } |
| 2308 | for (int l = 0; l < nb; l++) { |
| 2309 | for (int k = 0; k < (qk / blocklen); k++) { |
| 2310 | for (int m = 0; m < 4; m++) { |
| 2311 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2312 | sumi = 0; |
| 2313 | for (int i = 0; i < blocklen; ++i) { |
| 2314 | const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; |
| 2315 | sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; |
| 2316 | } |
| 2317 | sumf[m][j] += |
| 2318 | sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2319 | } |
| 2320 | } |
| 2321 | } |
| 2322 | } |
| 2323 | for (int m = 0; m < 4; m++) { |
| 2324 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2325 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2326 | } |
| 2327 | } |
| 2328 | } |
| 2329 | } |
| 2330 | } |
| 2331 | |
| 2332 | |
| 2333 | |
| 2334 | void ggml_gemm_q8_0_4x8_q8_0_genericggml_gemm_q8_0_4x8_q8_0(int n, |
| 2335 | float * GGML_RESTRICT__restrict__ s, |
| 2336 | size_t bs, |
| 2337 | const void * GGML_RESTRICT__restrict__ vx, |
| 2338 | const void * GGML_RESTRICT__restrict__ vy, |
| 2339 | int nr, |
| 2340 | int nc) { |
| 2341 | const int qk = QK8_032; |
| 2342 | const int nb = n / qk; |
| 2343 | const int ncols_interleaved = 4; |
| 2344 | const int blocklen = 8; |
| 2345 | |
| 2346 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2347 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2348 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2349 | |
| 2350 | float sumf[4][4]; |
| 2351 | int sumi; |
| 2352 | |
| 2353 | for (int y = 0; y < nr / 4; y++) { |
| 2354 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2355 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2356 | const block_q8_0x4 * b_ptr = (const block_q8_0x4 *) vx + (x * nb); |
| 2357 | for (int m = 0; m < 4; m++) { |
| 2358 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2359 | sumf[m][j] = 0.0; |
| 2360 | } |
| 2361 | } |
| 2362 | for (int l = 0; l < nb; l++) { |
| 2363 | for (int k = 0; k < (qk / blocklen); k++) { |
| 2364 | for (int m = 0; m < 4; m++) { |
| 2365 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2366 | sumi = 0; |
| 2367 | for (int i = 0; i < blocklen; ++i) { |
| 2368 | const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; |
| 2369 | sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; |
| 2370 | } |
| 2371 | sumf[m][j] += |
| 2372 | sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2373 | } |
| 2374 | } |
| 2375 | } |
| 2376 | } |
| 2377 | for (int m = 0; m < 4; m++) { |
| 2378 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2379 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2380 | } |
| 2381 | } |
| 2382 | } |
| 2383 | } |
| 2384 | } |
| 2385 | |
| 2386 | // Only enable these for RISC-V. |
| 2387 | #if defined __riscv_zvfh |
| 2388 | void ggml_gemm_q4_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2389 | const int qk = QK8_032; |
| 2390 | const int nb = n / qk; |
| 2391 | const int ncols_interleaved = 16; |
| 2392 | const int blocklen = 1; |
| 2393 | |
| 2394 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2395 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2396 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2397 | |
| 2398 | UNUSED(s)(void)(s); |
| 2399 | UNUSED(bs)(void)(bs); |
| 2400 | UNUSED(vx)(void)(vx); |
| 2401 | UNUSED(vy)(void)(vy); |
| 2402 | UNUSED(nr)(void)(nr); |
| 2403 | UNUSED(nc)(void)(nc); |
| 2404 | UNUSED(nb)(void)(nb); |
| 2405 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 2406 | UNUSED(blocklen)(void)(blocklen); |
| 2407 | |
| 2408 | float sumf[4][16]; |
| 2409 | int sumi; |
| 2410 | |
| 2411 | for (int y = 0; y < nr / 4; y++) { |
| 2412 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2413 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2414 | const block_q4_0x16 * b_ptr = (const block_q4_0x16 *) vx + (x * nb); |
| 2415 | for (int m = 0; m < 4; m++) { |
| 2416 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 2417 | } |
| 2418 | for (int l = 0; l < nb; l++) { |
| 2419 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 2420 | for (int m = 0; m < 4; m++) { |
| 2421 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2422 | sumi = 0; |
| 2423 | for (int i = 0; i < blocklen; ++i) { |
| 2424 | const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4); |
| 2425 | const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0); |
| 2426 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 2427 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4; |
| 2428 | } |
| 2429 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2430 | } |
| 2431 | } |
| 2432 | } |
| 2433 | } |
| 2434 | for (int m = 0; m < 4; m++) { |
| 2435 | for (int j = 0; j < ncols_interleaved; j++) |
| 2436 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2437 | } |
| 2438 | } |
| 2439 | } |
| 2440 | } |
| 2441 | |
| 2442 | void ggml_gemm_q4_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2443 | const int qk = QK_K256; |
| 2444 | const int nb = n / qk; |
| 2445 | const int ncols_interleaved = 16; |
| 2446 | const int blocklen = 1; |
| 2447 | |
| 2448 | assert (n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2449 | assert (nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2450 | assert (nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2451 | |
| 2452 | UNUSED(s)(void)(s); |
| 2453 | UNUSED(bs)(void)(bs); |
| 2454 | UNUSED(vx)(void)(vx); |
| 2455 | UNUSED(vy)(void)(vy); |
| 2456 | UNUSED(nr)(void)(nr); |
| 2457 | UNUSED(nc)(void)(nc); |
| 2458 | UNUSED(nb)(void)(nb); |
| 2459 | UNUSED(ncols_interleaved)(void)(ncols_interleaved); |
| 2460 | UNUSED(blocklen)(void)(blocklen); |
| 2461 | |
| 2462 | float sumf[4][16]; |
| 2463 | float sum_minf[4][16]; |
| 2464 | uint8_t scales[128]; |
| 2465 | uint8_t mins[128]; |
| 2466 | int sumi1; |
| 2467 | int sumi2; |
| 2468 | int sumi; |
| 2469 | |
| 2470 | for (int y = 0; y < nr / 4; y++) { |
| 2471 | const block_q8_Kx4 * a_ptr = (const block_q8_Kx4 *) vy + (y * nb); |
| 2472 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2473 | const block_q4_Kx16 * b_ptr = (const block_q4_Kx16 *) vx + (x * nb); |
| 2474 | for (int m = 0; m < 4; m++) { |
| 2475 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2476 | sumf[m][j] = 0.0; |
| 2477 | sum_minf[m][j] = 0.0; |
| 2478 | } |
| 2479 | } |
| 2480 | for (int l = 0; l < nb; l++) { |
| 2481 | for (int i = 0; i < 128; i++) { |
| 2482 | scales[i] = b_ptr[l].scales[i] & 0x0F; |
| 2483 | mins[i] = b_ptr[l].scales[i] >> 4; |
| 2484 | } |
| 2485 | for (int i = 0; i < 64; i++) { |
| 2486 | scales[i] |= (b_ptr[l].scales[128 + i] & 0x03) << 4; |
| 2487 | mins[i] |= (b_ptr[l].scales[128 + i] & 0x0C) << 2; |
| 2488 | scales[i + 64] |= (b_ptr[l].scales[128 + i] & 0x30); |
| 2489 | mins[i + 64] |= (b_ptr[l].scales[128 + i] & 0xC0) >> 2; |
| 2490 | } |
| 2491 | |
| 2492 | for (int sb = 0; sb < 8; sb++) { |
| 2493 | uint8_t *min = &mins[sb * 16]; |
| 2494 | for(int m = 0; m < 4; m++) { |
| 2495 | const int16_t bsums = a_ptr[l].bsums[sb * 8 + m] + a_ptr[l].bsums[sb * 8 + m + 4]; |
| 2496 | for(int j = 0; j < ncols_interleaved; j++) { |
| 2497 | sum_minf[m][j] += min[j] * bsums * GGML_CPU_FP16_TO_FP32(b_ptr[l].dmin[j])ggml_lookup_fp16_to_fp32(b_ptr[l].dmin[j]) * a_ptr[l].d[m]; |
| 2498 | } |
| 2499 | } |
| 2500 | } |
| 2501 | |
| 2502 | for (int sb = 0; sb < 8; sb += 2) { |
| 2503 | uint8_t *scales_0 = &scales[sb * 16]; |
| 2504 | uint8_t *scales_1 = &scales[(sb + 1) * 16]; |
| 2505 | |
| 2506 | for (int i = 0; i < QK4_032; i++) { |
| 2507 | for (int m = 0; m < 4; m++) { |
| 2508 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2509 | sumi1 = 0; |
| 2510 | sumi2 = 0; |
| 2511 | sumi = 0; |
| 2512 | |
| 2513 | const int v0 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] & 0xF); |
| 2514 | const int v1 = (int8_t) (b_ptr[l].qs[sb * 256 + i * 16 + j] >> 4); |
| 2515 | sumi1 = (v0 * a_ptr[l].qs[sb * 4 * 32 + i * 4 + m]); |
| 2516 | sumi2 = (v1 * a_ptr[l].qs[sb * 4 * 32 + 32 * 4 + i * 4 + m]); |
| 2517 | sumi1 = sumi1 * scales_0[j]; |
| 2518 | sumi2 = sumi2 * scales_1[j]; |
| 2519 | sumi += sumi1 + sumi2; |
| 2520 | |
| 2521 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * a_ptr[l].d[m]; |
| 2522 | } |
| 2523 | } |
| 2524 | } |
| 2525 | } |
| 2526 | } |
| 2527 | for (int m = 0; m < 4; m++) { |
| 2528 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2529 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j] - sum_minf[m][j]; |
| 2530 | } |
| 2531 | } |
| 2532 | } |
| 2533 | } |
| 2534 | } |
| 2535 | |
| 2536 | void ggml_gemm_iq4_nl_16x1_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2537 | const int qk = QK8_032; |
| 2538 | const int nb = n / qk; |
| 2539 | const int ncols_interleaved = 16; |
| 2540 | const int blocklen = 1; |
| 2541 | |
| 2542 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2543 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2544 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2545 | |
| 2546 | float sumf[4][16]; |
| 2547 | int sumi; |
| 2548 | |
| 2549 | for (int y = 0; y < nr / 4; y++) { |
| 2550 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2551 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2552 | const block_iq4_nlx16 * b_ptr = (const block_iq4_nlx16 *) vx + (x * nb); |
| 2553 | for (int m = 0; m < 4; m++) { |
| 2554 | for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0; |
| 2555 | } |
| 2556 | for (int l = 0; l < nb; l++) { |
| 2557 | for (int k = 0; k < (qk / (2 * blocklen)); k++) { |
| 2558 | for (int m = 0; m < 4; m++) { |
| 2559 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2560 | sumi = 0; |
| 2561 | for (int i = 0; i < blocklen; ++i) { |
| 2562 | const int v0 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0x0F]; |
| 2563 | const int v1 = kvalues_iq4nl[b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] >> 4]; |
| 2564 | sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) + |
| 2565 | (v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + (qk / 2) * 4])); |
| 2566 | } |
| 2567 | sumf[m][j] += sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2568 | } |
| 2569 | } |
| 2570 | } |
| 2571 | } |
| 2572 | for (int m = 0; m < 4; m++) { |
| 2573 | for (int j = 0; j < ncols_interleaved; j++) |
| 2574 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2575 | } |
| 2576 | } |
| 2577 | } |
| 2578 | } |
| 2579 | |
| 2580 | void ggml_gemm_q8_0_16x1_q8_0_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2581 | const int qk = QK8_032; |
| 2582 | const int nb = n / qk; |
| 2583 | const int ncols_interleaved = 16; |
| 2584 | const int blocklen = 1; |
| 2585 | |
| 2586 | assert(n % qk == 0)(static_cast <bool> (n % qk == 0) ? void (0) : __assert_fail ("n % qk == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2587 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2588 | assert(nc % ncols_interleaved == 0)(static_cast <bool> (nc % ncols_interleaved == 0) ? void (0) : __assert_fail ("nc % ncols_interleaved == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2589 | |
| 2590 | float sumf[4][16]; |
| 2591 | int sumi; |
| 2592 | |
| 2593 | for (int y = 0; y < nr / 4; y++) { |
| 2594 | const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb); |
| 2595 | for (int x = 0; x < nc / ncols_interleaved; x++) { |
| 2596 | const block_q8_0x16 * b_ptr = (const block_q8_0x16 *) vx + (x * nb); |
| 2597 | for (int m = 0; m < 4; m++) { |
| 2598 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2599 | sumf[m][j] = 0.0; |
| 2600 | } |
| 2601 | } |
| 2602 | for (int l = 0; l < nb; l++) { |
| 2603 | for (int k = 0; k < (qk / blocklen); k++) { |
| 2604 | for (int m = 0; m < 4; m++) { |
| 2605 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2606 | sumi = 0; |
| 2607 | for (int i = 0; i < blocklen; ++i) { |
| 2608 | const int v0 = b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i]; |
| 2609 | sumi += v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]; |
| 2610 | } |
| 2611 | sumf[m][j] += |
| 2612 | sumi * GGML_CPU_FP16_TO_FP32(b_ptr[l].d[j])ggml_lookup_fp16_to_fp32(b_ptr[l].d[j]) * GGML_CPU_FP16_TO_FP32(a_ptr[l].d[m])ggml_lookup_fp16_to_fp32(a_ptr[l].d[m]); |
| 2613 | } |
| 2614 | } |
| 2615 | } |
| 2616 | } |
| 2617 | for (int m = 0; m < 4; m++) { |
| 2618 | for (int j = 0; j < ncols_interleaved; j++) { |
| 2619 | s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j]; |
| 2620 | } |
| 2621 | } |
| 2622 | } |
| 2623 | } |
| 2624 | } |
| 2625 | |
| 2626 | |
| 2627 | void ggml_gemm_q2_K_16x1_q8_K_generic(int n, float * GGML_RESTRICT__restrict__ s, size_t bs, const void * GGML_RESTRICT__restrict__ vx, const void * GGML_RESTRICT__restrict__ vy, int nr, int nc) { |
| 2628 | assert(n % QK_K == 0)(static_cast <bool> (n % 256 == 0) ? void (0) : __assert_fail ("n % QK_K == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2629 | assert(nr % 4 == 0)(static_cast <bool> (nr % 4 == 0) ? void (0) : __assert_fail ("nr % 4 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2630 | assert(nc % 16 == 0)(static_cast <bool> (nc % 16 == 0) ? void (0) : __assert_fail ("nc % 16 == 0", __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__)); |
| 2631 | const int nb = n / QK_K256; |
| 2632 | const block_q2_Kx16 * x = (const block_q2_Kx16 *)vx; |
| 2633 | const block_q8_Kx4 * y = (const block_q8_Kx4 *)vy; |
| 2634 | |
| 2635 | const int sb_perm[16] = { |
| 2636 | 0, 4, 1, 5, 2, 6, 3, 7, |
| 2637 | 8, 12, 9, 13, 10, 14, 11, 15 |
| 2638 | }; |
| 2639 | |
| 2640 | // Iterate Rows in tiles of 4 |
| 2641 | for (int row_tile = 0; row_tile < nr; row_tile += 4) { |
| 2642 | // Iterate Columns in tiles of 16 |
| 2643 | for (int col_tile = 0; col_tile < nc; col_tile += 16) { |
| 2644 | |
| 2645 | const block_q2_Kx16 * x_ptr = x + (col_tile / 16) * nb; |
| 2646 | const block_q8_Kx4 * y_ptr = y + (row_tile / 4) * nb; |
| 2647 | |
| 2648 | float sumf[4][16]; |
| 2649 | memset(sumf, 0, sizeof(sumf)); |
| 2650 | |
| 2651 | for (int k_block = 0; k_block < nb; ++k_block) { |
| 2652 | int32_t isum[4][16]; |
| 2653 | int32_t summs[4][16]; |
| 2654 | memset(isum, 0, sizeof(isum)); |
| 2655 | memset(summs, 0, sizeof(summs)); |
| 2656 | |
| 2657 | const uint8_t * qs_rhs = x_ptr[k_block].qs; |
| 2658 | const uint8_t * sc_rhs = x_ptr[k_block].scales; |
| 2659 | const int8_t * qs_lhs = y_ptr[k_block].qs; |
| 2660 | const int16_t * bs_lhs = y_ptr[k_block].bsums; |
| 2661 | |
| 2662 | for (int sb = 0; sb < 16; ++sb) { |
| 2663 | int scale_offset = sb_perm[sb] * 16; |
| 2664 | |
| 2665 | int byte_base; |
| 2666 | if (sb < 8) byte_base = (sb % 2 == 0) ? 0 : 16; |
| 2667 | else byte_base = (sb % 2 == 0) ? 32 : 48; |
| 2668 | int shift = ((sb / 2) % 4) * 2; |
| 2669 | |
| 2670 | for (int col = 0; col < 16; ++col) { |
| 2671 | uint8_t sc_val = sc_rhs[scale_offset + col]; |
| 2672 | int32_t d_sb = sc_val & 0xF; |
| 2673 | int32_t m_sb = sc_val >> 4; |
| 2674 | |
| 2675 | // Correction Term |
| 2676 | for (int r = 0; r < 4; ++r) { |
| 2677 | int bsum_idx = (sb / 4) * 16 + r * 4 + (sb % 4); |
| 2678 | summs[r][col] += bs_lhs[bsum_idx] * m_sb; |
| 2679 | } |
| 2680 | |
| 2681 | // Main Dot Product |
| 2682 | for (int l = 0; l < 16; ++l) { |
| 2683 | int qs_idx = (byte_base + l) * 16 + col; |
| 2684 | uint8_t q2_val = (qs_rhs[qs_idx] >> shift) & 3; |
| 2685 | |
| 2686 | // Calculate Q8 index for this specific k and row |
| 2687 | int k = sb * 16 + l; |
| 2688 | int q8_idx = (k / 4) * 16 + (k % 4); |
| 2689 | |
| 2690 | for (int r = 0; r < 4; ++r) { |
| 2691 | // Add r*4 to jump to the correct row within the 4x4 chunk |
| 2692 | int8_t q8_val = qs_lhs[q8_idx + r * 4]; |
| 2693 | isum[r][col] += q8_val * q2_val * d_sb; |
| 2694 | } |
| 2695 | } |
| 2696 | } |
| 2697 | } |
| 2698 | |
| 2699 | // Finalize K-Block |
| 2700 | for (int col = 0; col < 16; ++col) { |
| 2701 | float d_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].d[col])ggml_compute_fp16_to_fp32(x_ptr[k_block].d[col]); |
| 2702 | float dm_rhs = GGML_FP16_TO_FP32(x_ptr[k_block].dmin[col])ggml_compute_fp16_to_fp32(x_ptr[k_block].dmin[col]); |
| 2703 | |
| 2704 | for (int r = 0; r < 4; ++r) { |
| 2705 | float d_lhs = y_ptr[k_block].d[r]; |
| 2706 | float d_all = d_lhs * d_rhs; |
| 2707 | float d_min = d_lhs * dm_rhs; |
| 2708 | sumf[r][col] += (isum[r][col] * d_all) - (summs[r][col] * d_min); |
| 2709 | } |
| 2710 | } |
| 2711 | } |
| 2712 | |
| 2713 | for (int r = 0; r < 4; ++r) { |
| 2714 | for (int col = 0; col < 16; ++col) { |
| 2715 | s[(row_tile + r) * bs + (col_tile + col)] = sumf[r][col]; |
| 2716 | } |
| 2717 | } |
| 2718 | } |
| 2719 | } |
| 2720 | } |
| 2721 | #endif |
| 2722 | |
| 2723 | } // extern "C" |
| 2724 | |
| 2725 | static block_q8_0x4 make_block_q8_0x4(block_q8_0 * in, unsigned int blck_size_interleave) { |
| 2726 | block_q8_0x4 out; |
| 2727 | |
| 2728 | for (int i = 0; i < 4; i++) { |
| 2729 | out.d[i] = in[i].d; |
| 2730 | } |
| 2731 | |
| 2732 | const int end = QK8_032 * 4 / blck_size_interleave; |
| 2733 | for (int i = 0; i < end; ++i) { |
| 2734 | int src_id = i % 4; |
| 2735 | int src_offset = (i / 4) * blck_size_interleave; |
| 2736 | int dst_offset = i * blck_size_interleave; |
| 2737 | memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave); |
| 2738 | } |
| 2739 | return out; |
| 2740 | } |
| 2741 | |
| 2742 | static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave) { |
| 2743 | block_q4_0x4 out; |
| 2744 | |
| 2745 | for (int i = 0; i < 4; i++) { |
| 2746 | out.d[i] = in[i].d; |
| 2747 | } |
| 2748 | |
| 2749 | const int end = QK4_032 * 2 / blck_size_interleave; |
| 2750 | |
| 2751 | if (blck_size_interleave == 8) { |
| 2752 | const uint64_t xor_mask = 0x8888888888888888ULL; |
| 2753 | for (int i = 0; i < end; ++i) { |
| 2754 | int src_id = i % 4; |
| 2755 | int src_offset = (i / 4) * blck_size_interleave; |
| 2756 | int dst_offset = i * blck_size_interleave; |
| 2757 | |
| 2758 | uint64_t elems; |
| 2759 | // Using memcpy to avoid unaligned memory accesses |
| 2760 | memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t)); |
| 2761 | elems ^= xor_mask; |
| 2762 | memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t)); |
| 2763 | } |
| 2764 | } else if (blck_size_interleave == 4) { |
| 2765 | const uint32_t xor_mask = 0x88888888; |
| 2766 | for (int i = 0; i < end; ++i) { |
| 2767 | int src_id = i % 4; |
| 2768 | int src_offset = (i / 4) * blck_size_interleave; |
| 2769 | int dst_offset = i * blck_size_interleave; |
| 2770 | |
| 2771 | uint32_t elems; |
| 2772 | memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint32_t)); |
| 2773 | elems ^= xor_mask; |
| 2774 | memcpy(&out.qs[dst_offset], &elems, sizeof(uint32_t)); |
| 2775 | } |
| 2776 | } else { |
| 2777 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 2777, "GGML_ASSERT(%s) failed", "false"); |
| 2778 | } |
| 2779 | |
| 2780 | return out; |
| 2781 | } |
| 2782 | |
| 2783 | // interleave 8 block_q4_0s in blocks of blck_size_interleave |
| 2784 | // returns an interleaved block_q4_0x8 |
| 2785 | // in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks |
| 2786 | // first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave |
| 2787 | static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave) { |
| 2788 | block_q4_0x8 out; |
| 2789 | |
| 2790 | for (int i = 0; i < 8; i++) { |
| 2791 | out.d[i] = in[i].d; |
| 2792 | } |
| 2793 | |
| 2794 | const int end = QK4_032 * 4 / blck_size_interleave; |
| 2795 | const uint64_t xor_mask = 0x8888888888888888ULL; |
| 2796 | |
| 2797 | for (int i = 0; i < end; ++i) { |
| 2798 | int src_id = i % 8; |
| 2799 | int src_offset = (i / 8) * blck_size_interleave; |
| 2800 | int dst_offset = i * blck_size_interleave; |
| 2801 | |
| 2802 | uint64_t elems; |
| 2803 | memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t)); |
| 2804 | elems ^= xor_mask; |
| 2805 | memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t)); |
| 2806 | } |
| 2807 | |
| 2808 | return out; |
| 2809 | } |
| 2810 | |
| 2811 | static block_q4_0x16 make_block_q4_0x16(block_q4_0 * in, unsigned int blck_size_interleave) { |
| 2812 | block_q4_0x16 out; |
| 2813 | |
| 2814 | for (int i = 0; i < 16; i++) { |
| 2815 | out.d[i] = in[i].d; |
| 2816 | } |
| 2817 | |
| 2818 | const int end = QK4_032 * 8 / blck_size_interleave; |
| 2819 | |
| 2820 | if (blck_size_interleave == 1) { |
| 2821 | const uint8_t xor_mask = 0x88; |
| 2822 | for (int i = 0; i < end; ++i) { |
| 2823 | int src_id = i % 16; |
| 2824 | int src_offset = i / 16; |
| 2825 | int dst_offset = i; |
| 2826 | |
| 2827 | out.qs[dst_offset] = in[src_id].qs[src_offset] ^ xor_mask; |
| 2828 | } |
| 2829 | } else { |
| 2830 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 2830, "GGML_ASSERT(%s) failed", "false"); |
| 2831 | } |
| 2832 | |
| 2833 | return out; |
| 2834 | } |
| 2835 | |
| 2836 | static block_q4_Kx8 make_block_q4_Kx8(block_q4_K * in, unsigned int blck_size_interleave) { |
| 2837 | block_q4_Kx8 out; |
| 2838 | //Delta(scale) and dmin values of the eight Q4_K structures are copied onto the output interleaved structure |
| 2839 | for (int i = 0; i < 8; i++) { |
| 2840 | out.d[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.d; |
| 2841 | } |
| 2842 | |
| 2843 | for (int i = 0; i < 8; i++) { |
| 2844 | out.dmin[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.dmin; |
| 2845 | } |
| 2846 | |
| 2847 | const int end = QK_K256 * 4 / blck_size_interleave; |
| 2848 | |
| 2849 | // Interleave Q4_K quants by taking 8 bytes at a time |
| 2850 | for (int i = 0; i < end; ++i) { |
| 2851 | int src_id = i % 8; |
| 2852 | int src_offset = (i / 8) * blck_size_interleave; |
| 2853 | int dst_offset = i * blck_size_interleave; |
| 2854 | |
| 2855 | // buffer large enough for the max interleave block size (8 bytes) |
| 2856 | uint64_t elems; |
| 2857 | memcpy(&elems, &in[src_id].qs[src_offset], blck_size_interleave); |
| 2858 | memcpy(&out.qs[dst_offset], &elems, blck_size_interleave); |
| 2859 | } |
| 2860 | |
| 2861 | // The below logic is designed so as to unpack and rearrange scales and mins values in Q4_K |
| 2862 | // Currently the Q4_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value) |
| 2863 | // The output Q4_Kx8 structure has 96 bytes |
| 2864 | // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q4_K structure |
| 2865 | // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q4_K structures |
| 2866 | uint8_t s[8], m[8]; |
| 2867 | |
| 2868 | for (int i = 0; i < 4; i++) { |
| 2869 | for (int j = 0; j < 8; j++) { |
| 2870 | s[j] = in[j].scales[i] & 63; |
| 2871 | m[j] = in[j].scales[i + 4] & 63; |
| 2872 | } |
| 2873 | |
| 2874 | out.scales[i * 12] = (s[0] & 63) + ((s[4] & 48) << 2); |
| 2875 | out.scales[i * 12 + 1] = (s[1] & 63) + ((s[5] & 48) << 2); |
| 2876 | out.scales[i * 12 + 2] = (s[2] & 63) + ((s[6] & 48) << 2); |
| 2877 | out.scales[i * 12 + 3] = (s[3] & 63) + ((s[7] & 48) << 2); |
| 2878 | out.scales[i * 12 + 4] = (m[0] & 63) + ((m[4] & 48) << 2); |
| 2879 | out.scales[i * 12 + 5] = (m[1] & 63) + ((m[5] & 48) << 2); |
| 2880 | out.scales[i * 12 + 6] = (m[2] & 63) + ((m[6] & 48) << 2); |
| 2881 | out.scales[i * 12 + 7] = (m[3] & 63) + ((m[7] & 48) << 2); |
| 2882 | out.scales[i * 12 + 8] = (s[4] & 15) + ((m[4] & 15) << 4); |
| 2883 | out.scales[i * 12 + 9] = (s[5] & 15) + ((m[5] & 15) << 4); |
| 2884 | out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4); |
| 2885 | out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4); |
| 2886 | |
| 2887 | } |
| 2888 | |
| 2889 | for (int i = 0; i < 4; i++) { |
| 2890 | for (int j = 0; j < 8; j++) { |
| 2891 | s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15); |
| 2892 | m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4); |
| 2893 | } |
| 2894 | |
| 2895 | out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2); |
| 2896 | out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2); |
| 2897 | out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2); |
| 2898 | out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2); |
| 2899 | out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2); |
| 2900 | out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2); |
| 2901 | out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2); |
| 2902 | out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2); |
| 2903 | out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4); |
| 2904 | out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4); |
| 2905 | out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4); |
| 2906 | out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4); |
| 2907 | |
| 2908 | } |
| 2909 | |
| 2910 | return out; |
| 2911 | } |
| 2912 | |
| 2913 | static block_q4_Kx16 make_block_q4_Kx16(block_q4_K * in, unsigned int blck_size_interleave) { |
| 2914 | block_q4_Kx16 out; |
| 2915 | //Delta(scale) and dmin values of the 16 Q4_K structures are copied onto the output interleaved structure |
| 2916 | for (int i = 0; i < 16; i++) { |
| 2917 | out.d[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.d; |
| 2918 | } |
| 2919 | |
| 2920 | for (int i = 0; i < 16; i++) { |
| 2921 | out.dmin[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.dmin; |
| 2922 | } |
| 2923 | |
| 2924 | const int end = QK_K256 * 8 / blck_size_interleave; |
| 2925 | |
| 2926 | if (blck_size_interleave == 1) { |
| 2927 | for (int i = 0; i < end; ++i) { |
| 2928 | int src_id = i % 16; |
| 2929 | int src_offset = i / 16; |
| 2930 | int dst_offset = i; |
| 2931 | |
| 2932 | out.qs[dst_offset] = in[src_id].qs[src_offset]; |
| 2933 | } |
| 2934 | |
| 2935 | // RVV repacking. |
| 2936 | // |
| 2937 | // Extract sums and mins for all 8 sub-blocks for each block of Q4_K. |
| 2938 | uint8_t s[128], m[128]; |
| 2939 | for (int i = 0; i < 4; i++) { |
| 2940 | for (int j = 0; j < 16; j++) { |
| 2941 | s[i * 16 + j] = in[j].scales[i] & 63; |
| 2942 | m[i * 16 + j] = in[j].scales[i + 4] & 63; |
| 2943 | } |
| 2944 | } |
| 2945 | for (int i = 0; i < 4; i++) { |
| 2946 | for (int j = 0; j < 16; j++) { |
| 2947 | s[64 + i * 16 + j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i+8] & 15); |
| 2948 | m[64 + i * 16 + j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i+8] & 240) >> 4); |
| 2949 | } |
| 2950 | } |
| 2951 | |
| 2952 | for (int i = 0; i < 128; i++) { |
| 2953 | out.scales[i] = (s[i] & 15) | ((m[i] & 15) << 4); |
| 2954 | } |
| 2955 | for (int i = 0; i < 64; i++) { |
| 2956 | out.scales[128 + i] = ((s[i] & 48) >> 4) | ((m[i] & 48) >> 2) | (s[64 + i] & 48) | ((m[64 + i] & 48) << 2); |
| 2957 | } |
| 2958 | } else { |
| 2959 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 2959, "GGML_ASSERT(%s) failed", "false"); |
| 2960 | } |
| 2961 | |
| 2962 | return out; |
| 2963 | } |
| 2964 | |
| 2965 | static block_q2_Kx8 make_block_q2_Kx8(block_q2_K * in, unsigned int blck_size_interleave) { |
| 2966 | block_q2_Kx8 out; |
| 2967 | |
| 2968 | // Delta(scale) and dmin values of the eight Q2_K structures are copied onto the output interleaved structure |
| 2969 | for (int i = 0; i < 8; i++) { |
| 2970 | out.d[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.d; |
| 2971 | } |
| 2972 | |
| 2973 | for (int i = 0; i < 8; i++) { |
| 2974 | out.dmin[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.dmin; |
| 2975 | } |
| 2976 | |
| 2977 | const int end = QK_K256 * 2 / blck_size_interleave; |
| 2978 | |
| 2979 | // Interleave Q2_K quants by taking 8 bytes at a time |
| 2980 | for (int i = 0; i < end; ++i) { |
| 2981 | int src_id = i % 8; |
| 2982 | int src_offset = (i / 8) * blck_size_interleave; |
| 2983 | int dst_offset = i * blck_size_interleave; |
| 2984 | |
| 2985 | uint64_t elems; |
| 2986 | memcpy(&elems, &in[src_id].qs[src_offset], sizeof(uint64_t)); |
| 2987 | memcpy(&out.qs[dst_offset], &elems, sizeof(uint64_t)); |
| 2988 | } |
| 2989 | |
| 2990 | // The below logic is designed so as to unpack and rearrange scales and mins values in Q2_K |
| 2991 | // Currently the Q2_K structure has 16 scales and 16 mins packed in 16 bytes ( 4 bits for each value) |
| 2992 | // The output Q2_Kx8 structure has 128 bytes for storing scales and mins |
| 2993 | // Every 16 byte is packed such that it contains scales and mins for corresponding sub blocks from Q2_K structure |
| 2994 | // For eg - First 16 bytes contains 16 scales and 16 mins - each of first and second sub blocks from different Q2_K structures |
| 2995 | |
| 2996 | for (int i = 0; i < 128; i++) { |
| 2997 | // Index for selecting which q2k super block |
| 2998 | int src1 = (i % 16) / 2; |
| 2999 | // Index for selecting scale |
| 3000 | int src2 = ((i / 16) * 2) + (i % 2); |
| 3001 | |
| 3002 | out.scales[i] = in[src1].scales[src2]; |
| 3003 | } |
| 3004 | return out; |
| 3005 | } |
| 3006 | |
| 3007 | static block_q5_Kx8 make_block_q5_Kx8(block_q5_K * in, unsigned int blck_size_interleave) { |
| 3008 | block_q5_Kx8 out; |
| 3009 | //Delta(scale) and dmin values of the eight Q5_K structures are copied onto the output interleaved structure |
| 3010 | for (int i = 0; i < 8; i++) { |
| 3011 | out.d[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.d; |
| 3012 | } |
| 3013 | |
| 3014 | for (int i = 0; i < 8; i++) { |
| 3015 | out.dmin[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.dmin; |
| 3016 | } |
| 3017 | |
| 3018 | const int end = QK_K256 * 4 / blck_size_interleave; |
| 3019 | |
| 3020 | // Interleave Q5_K quants by taking blck_size_interleave bytes at a time |
| 3021 | for (int i = 0; i < end; ++i) { |
| 3022 | int src_id = i % 8; |
| 3023 | int src_offset = (i / 8) * blck_size_interleave; |
| 3024 | int dst_offset = i * blck_size_interleave; |
| 3025 | |
| 3026 | memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], blck_size_interleave); |
| 3027 | } |
| 3028 | |
| 3029 | // Repeat for high bits with the same chunk size, since |
| 3030 | // the high bits are interleaved in Q5_K and the index is |
| 3031 | // qh_idx = (qs_idx % 32); |
| 3032 | // qh_val = qh[qh_idx] >> (qs_idx / 32); |
| 3033 | for (int i = 0; i < end / 4; ++i) { |
| 3034 | int src_id = i % 8; |
| 3035 | int src_offset = (i / 8) * blck_size_interleave; |
| 3036 | int dst_offset = i * blck_size_interleave; |
| 3037 | |
| 3038 | memcpy(&out.qh[dst_offset], &in[src_id].qh[src_offset], blck_size_interleave); |
| 3039 | } |
| 3040 | |
| 3041 | // The below logic is copied over from Q4_K |
| 3042 | // The point is to unpack all the scales and mins for each sub block every time we load 12 bytes. |
| 3043 | // Currently the Q5_K structure has 8 scales and 8 mins packed in 12 bytes ( 6 bits for each value) |
| 3044 | // The output Q5_Kx8 structure has 96 bytes |
| 3045 | // Every 12 byte is packed such that it contains scales and mins for corresponding sub blocks from Q5_K structure |
| 3046 | // For eg - First 12 bytes contains 8 scales and 8 mins - each of first sub block from different Q5_K structures |
| 3047 | uint8_t s[8], m[8]; |
| 3048 | |
| 3049 | for (int i = 0; i < 4; i++) { |
| 3050 | for (int j = 0; j < 8; j++) { |
| 3051 | s[j] = in[j].scales[i] & 63; |
| 3052 | m[j] = in[j].scales[i + 4] & 63; |
| 3053 | } |
| 3054 | |
| 3055 | out.scales[i * 12] = (s[0] & 63) + ((s[4] & 48) << 2); |
| 3056 | out.scales[i * 12 + 1] = (s[1] & 63) + ((s[5] & 48) << 2); |
| 3057 | out.scales[i * 12 + 2] = (s[2] & 63) + ((s[6] & 48) << 2); |
| 3058 | out.scales[i * 12 + 3] = (s[3] & 63) + ((s[7] & 48) << 2); |
| 3059 | out.scales[i * 12 + 4] = (m[0] & 63) + ((m[4] & 48) << 2); |
| 3060 | out.scales[i * 12 + 5] = (m[1] & 63) + ((m[5] & 48) << 2); |
| 3061 | out.scales[i * 12 + 6] = (m[2] & 63) + ((m[6] & 48) << 2); |
| 3062 | out.scales[i * 12 + 7] = (m[3] & 63) + ((m[7] & 48) << 2); |
| 3063 | out.scales[i * 12 + 8] = (s[4] & 15) + ((m[4] & 15) << 4); |
| 3064 | out.scales[i * 12 + 9] = (s[5] & 15) + ((m[5] & 15) << 4); |
| 3065 | out.scales[i * 12 + 10] = (s[6] & 15) + ((m[6] & 15) << 4); |
| 3066 | out.scales[i * 12 + 11] = (s[7] & 15) + ((m[7] & 15) << 4); |
| 3067 | } |
| 3068 | |
| 3069 | for (int i = 0; i < 4; i++) { |
| 3070 | for (int j = 0; j < 8; j++) { |
| 3071 | s[j] = ((in[j].scales[i] & 192) >> 2) | (in[j].scales[i + 8] & 15); |
| 3072 | m[j] = ((in[j].scales[i + 4] & 192) >> 2) | ((in[j].scales[i + 8] & 240) >> 4); |
| 3073 | } |
| 3074 | |
| 3075 | out.scales[i * 12 + 48] = (s[0] & 63) + ((s[4] & 48) << 2); |
| 3076 | out.scales[i * 12 + 49] = (s[1] & 63) + ((s[5] & 48) << 2); |
| 3077 | out.scales[i * 12 + 50] = (s[2] & 63) + ((s[6] & 48) << 2); |
| 3078 | out.scales[i * 12 + 51] = (s[3] & 63) + ((s[7] & 48) << 2); |
| 3079 | out.scales[i * 12 + 52] = (m[0] & 63) + ((m[4] & 48) << 2); |
| 3080 | out.scales[i * 12 + 53] = (m[1] & 63) + ((m[5] & 48) << 2); |
| 3081 | out.scales[i * 12 + 54] = (m[2] & 63) + ((m[6] & 48) << 2); |
| 3082 | out.scales[i * 12 + 55] = (m[3] & 63) + ((m[7] & 48) << 2); |
| 3083 | out.scales[i * 12 + 56] = (s[4] & 15) + ((m[4] & 15) << 4); |
| 3084 | out.scales[i * 12 + 57] = (s[5] & 15) + ((m[5] & 15) << 4); |
| 3085 | out.scales[i * 12 + 58] = (s[6] & 15) + ((m[6] & 15) << 4); |
| 3086 | out.scales[i * 12 + 59] = (s[7] & 15) + ((m[7] & 15) << 4); |
| 3087 | } |
| 3088 | |
| 3089 | return out; |
| 3090 | } |
| 3091 | |
| 3092 | static block_q6_Kx8 make_block_q6_Kx8(block_q6_K * in, unsigned int blck_size_interleave) { |
| 3093 | block_q6_Kx8 out; |
| 3094 | constexpr int n_blocks = 8; // Kx8 |
| 3095 | for (int i = 0; i < n_blocks; i++) { |
| 3096 | out.d[i] = in[i].d; |
| 3097 | } |
| 3098 | |
| 3099 | const int end_ls = QK_K256 * 4 / blck_size_interleave; |
| 3100 | // Interleave Q6_K quants by taking blck_size_interleave bytes at a time |
| 3101 | for (int i = 0; i < end_ls; ++i) { |
| 3102 | int src_id = i % n_blocks; |
| 3103 | int src_offset = (i / n_blocks) * blck_size_interleave; |
| 3104 | int dst_offset = i * blck_size_interleave; |
| 3105 | |
| 3106 | uint64_t elem_ls; |
| 3107 | memcpy(&elem_ls, &in[src_id].ql[src_offset], blck_size_interleave); |
| 3108 | memcpy(&out.ql[dst_offset], &elem_ls, blck_size_interleave); |
| 3109 | } |
| 3110 | |
| 3111 | // Interleave high bits using same chunk size as low bits |
| 3112 | const int end_hs = end_ls / 2; |
| 3113 | for (int i = 0; i < end_hs; ++i) { |
| 3114 | int src_id = i % n_blocks; |
| 3115 | int src_offset = (i / n_blocks) * blck_size_interleave; |
| 3116 | int dst_offset = i * blck_size_interleave; |
| 3117 | |
| 3118 | uint64_t elem_hs; |
| 3119 | memcpy(&elem_hs, &in[src_id].qh[src_offset], blck_size_interleave); |
| 3120 | memcpy(&out.qh[dst_offset], &elem_hs, blck_size_interleave); |
| 3121 | } |
| 3122 | |
| 3123 | // The below logic is designed so as to unpack and rearrange scales in Q6_K |
| 3124 | // The output Q6_Kx8 structure interleaves the 8 bit scales in the same fashion as the quants |
| 3125 | // Q6_K structure has an 8-bit scale per 16 elements -> 16 scales |
| 3126 | // scales: [0 bl0 0 bl1 ... 0 bl7][1 bl0 ... 1 bl7] ... [15 bl0 ... 15 bl7] (bl = block) |
| 3127 | constexpr int n_scales = QK_K256 / 16; |
| 3128 | |
| 3129 | for (int i = 0; i < n_blocks; i++) { |
| 3130 | for (int j = 0; j < n_scales; j++) { |
| 3131 | out.scales[j * n_blocks + i] = in[i].scales[j]; |
| 3132 | } |
| 3133 | } |
| 3134 | |
| 3135 | return out; |
| 3136 | } |
| 3137 | |
| 3138 | static block_q2_Kx16 make_block_q2_Kx16(const block_q2_K * in, unsigned int blck_size_interleave) { |
| 3139 | block_q2_Kx16 out; |
| 3140 | constexpr int N_COLS = 16; |
| 3141 | |
| 3142 | // 1. Copy Super-Scales (d) and Super-Mins (dmin) |
| 3143 | for (int i = 0; i < N_COLS; i++) { |
| 3144 | out.d[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.d; |
| 3145 | out.dmin[i] = in[i].GGML_COMMON_AGGR_Udata.GGML_COMMON_AGGR_Sdata.dmin; |
| 3146 | } |
| 3147 | |
| 3148 | // 2. Interleave Q2_K Data |
| 3149 | const int bytes_per_col = 64; |
| 3150 | const int total_bytes = N_COLS * bytes_per_col; |
| 3151 | const int end = total_bytes / blck_size_interleave; |
| 3152 | |
| 3153 | for (int i = 0; i < end; ++i) { |
| 3154 | int src_col_id = i % N_COLS; |
| 3155 | int src_offset = (i / N_COLS) * blck_size_interleave; |
| 3156 | int dst_offset = i * blck_size_interleave; |
| 3157 | memcpy(&out.qs[dst_offset], &in[src_col_id].qs[src_offset], blck_size_interleave); |
| 3158 | } |
| 3159 | |
| 3160 | // 3. Repack Scales into the Optimized "Sequential-Parallel" Layout |
| 3161 | int out_idx = 0; |
| 3162 | |
| 3163 | // Arrays define the sub-block order for each group |
| 3164 | const int even_low_sbs[] = {0, 2, 4, 6}; |
| 3165 | const int odd_low_sbs[] = {1, 3, 5, 7}; |
| 3166 | const int even_high_sbs[] = {8, 10, 12, 14}; |
| 3167 | const int odd_high_sbs[] = {9, 11, 13, 15}; |
| 3168 | |
| 3169 | // Pack Group 1: Even-Low |
| 3170 | for (int sb : even_low_sbs) { |
| 3171 | for (int col = 0; col < N_COLS; col++) { |
| 3172 | out.scales[out_idx++] = in[col].scales[sb]; |
| 3173 | } |
| 3174 | } |
| 3175 | |
| 3176 | // Pack Group 2: Odd-Low |
| 3177 | for (int sb : odd_low_sbs) { |
| 3178 | for (int col = 0; col < N_COLS; col++) { |
| 3179 | out.scales[out_idx++] = in[col].scales[sb]; |
| 3180 | } |
| 3181 | } |
| 3182 | |
| 3183 | // Pack Group 3: Even-High |
| 3184 | for (int sb : even_high_sbs) { |
| 3185 | for (int col = 0; col < N_COLS; col++) { |
| 3186 | out.scales[out_idx++] = in[col].scales[sb]; |
| 3187 | } |
| 3188 | } |
| 3189 | |
| 3190 | // Pack Group 4: Odd-High |
| 3191 | for (int sb : odd_high_sbs) { |
| 3192 | for (int col = 0; col < N_COLS; col++) { |
| 3193 | out.scales[out_idx++] = in[col].scales[sb]; |
| 3194 | } |
| 3195 | } |
| 3196 | |
| 3197 | return out; |
| 3198 | } |
| 3199 | |
| 3200 | static int repack_q4_0_to_q4_0_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3201 | GGML_ASSERT(t->type == GGML_TYPE_Q4_0)if (!(t->type == GGML_TYPE_Q4_0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3201, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q4_0" ); |
| 3202 | GGML_ASSERT(interleave_block == 4 || interleave_block == 8)if (!(interleave_block == 4 || interleave_block == 8)) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3202, "GGML_ASSERT(%s) failed", "interleave_block == 4 || interleave_block == 8" ); |
| 3203 | constexpr int nrows_interleaved = 4; |
| 3204 | |
| 3205 | block_q4_0x4 * dst = (block_q4_0x4 *)t->data; |
| 3206 | const block_q4_0 * src = (const block_q4_0 *)data; |
| 3207 | block_q4_0 dst_tmp[4]; |
| 3208 | int nrow = ggml_nrows(t); |
| 3209 | int nblocks = t->ne[0] / QK4_032; |
| 3210 | |
| 3211 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0))if (!(data_size == nrow * nblocks * sizeof(block_q4_0))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3211, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q4_0)" ); |
| 3212 | |
| 3213 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3214 | return -1; |
| 3215 | } |
| 3216 | |
| 3217 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3218 | for (int64_t x = 0; x < nblocks; x++) { |
| 3219 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3220 | dst_tmp[i] = src[x + i * nblocks]; |
| 3221 | } |
| 3222 | *dst++ = make_block_q4_0x4(dst_tmp, interleave_block); |
| 3223 | } |
| 3224 | src += nrows_interleaved * nblocks; |
| 3225 | } |
| 3226 | return 0; |
| 3227 | |
| 3228 | GGML_UNUSED(data_size)(void)(data_size); |
| 3229 | } |
| 3230 | |
| 3231 | static int repack_q4_K_to_q4_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3232 | GGML_ASSERT(t->type == GGML_TYPE_Q4_K)if (!(t->type == GGML_TYPE_Q4_K)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3232, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q4_K" ); |
| 3233 | GGML_ASSERT(interleave_block == 8 || interleave_block == 4)if (!(interleave_block == 8 || interleave_block == 4)) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3233, "GGML_ASSERT(%s) failed", "interleave_block == 8 || interleave_block == 4" ); |
| 3234 | constexpr int nrows_interleaved = 8; |
| 3235 | |
| 3236 | block_q4_Kx8 * dst = (block_q4_Kx8*)t->data; |
| 3237 | const block_q4_K * src = (const block_q4_K*) data; |
| 3238 | block_q4_K dst_tmp[8]; |
| 3239 | int nrow = ggml_nrows(t); |
| 3240 | int nblocks = t->ne[0] / QK_K256; |
| 3241 | |
| 3242 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_K))if (!(data_size == nrow * nblocks * sizeof(block_q4_K))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3242, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q4_K)" ); |
| 3243 | |
| 3244 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3245 | return -1; |
| 3246 | } |
| 3247 | |
| 3248 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3249 | for (int64_t x = 0; x < nblocks; x++) { |
| 3250 | for (int i = 0; i < nrows_interleaved; i++ ) { |
| 3251 | dst_tmp[i] = src[x + i * nblocks]; |
| 3252 | } |
| 3253 | *dst++ = make_block_q4_Kx8(dst_tmp, interleave_block); |
| 3254 | } |
| 3255 | src += nrows_interleaved * nblocks; |
| 3256 | } |
| 3257 | return 0; |
| 3258 | |
| 3259 | GGML_UNUSED(data_size)(void)(data_size); |
| 3260 | } |
| 3261 | |
| 3262 | static int repack_q4_K_to_q4_K_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3263 | GGML_ASSERT(t->type == GGML_TYPE_Q4_K)if (!(t->type == GGML_TYPE_Q4_K)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3263, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q4_K" ); |
| 3264 | constexpr int nrows_interleaved = 16; |
| 3265 | |
| 3266 | block_q4_Kx16 * dst = (block_q4_Kx16*)t->data; |
| 3267 | const block_q4_K * src = (const block_q4_K*) data; |
| 3268 | block_q4_K dst_tmp[16]; |
| 3269 | int nrow = ggml_nrows(t); |
| 3270 | int nblocks = t->ne[0] / QK_K256; |
| 3271 | |
| 3272 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_K))if (!(data_size == nrow * nblocks * sizeof(block_q4_K))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3272, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q4_K)" ); |
| 3273 | |
| 3274 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3275 | return -1; |
| 3276 | } |
| 3277 | |
| 3278 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3279 | for (int64_t x = 0; x < nblocks; x++) { |
| 3280 | for (int i = 0; i < nrows_interleaved; i++ ) { |
| 3281 | dst_tmp[i] = src[x + i * nblocks]; |
| 3282 | } |
| 3283 | *dst++ = make_block_q4_Kx16(dst_tmp, interleave_block); |
| 3284 | } |
| 3285 | src += nrows_interleaved * nblocks; |
| 3286 | } |
| 3287 | return 0; |
| 3288 | |
| 3289 | GGML_UNUSED(data_size)(void)(data_size); |
| 3290 | } |
| 3291 | |
| 3292 | static int repack_q2_K_to_q2_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3293 | GGML_ASSERT(t->type == GGML_TYPE_Q2_K)if (!(t->type == GGML_TYPE_Q2_K)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3293, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q2_K" ); |
| 3294 | GGML_ASSERT(interleave_block == 8)if (!(interleave_block == 8)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3294, "GGML_ASSERT(%s) failed", "interleave_block == 8"); |
| 3295 | constexpr int nrows_interleaved = 8; |
| 3296 | |
| 3297 | block_q2_Kx8 * dst = (block_q2_Kx8*)t->data; |
| 3298 | const block_q2_K * src = (const block_q2_K*) data; |
| 3299 | block_q2_K dst_tmp[8]; |
| 3300 | int nrow = ggml_nrows(t); |
| 3301 | int nblocks = t->ne[0] / QK_K256; |
| 3302 | |
| 3303 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K))if (!(data_size == nrow * nblocks * sizeof(block_q2_K))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3303, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q2_K)" ); |
| 3304 | |
| 3305 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3306 | return -1; |
| 3307 | } |
| 3308 | |
| 3309 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3310 | for (int64_t x = 0; x < nblocks; x++) { |
| 3311 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3312 | dst_tmp[i] = src[x + i * nblocks]; |
| 3313 | } |
| 3314 | *dst++ = make_block_q2_Kx8(dst_tmp, interleave_block); |
| 3315 | } |
| 3316 | src += nrows_interleaved * nblocks; |
| 3317 | } |
| 3318 | return 0; |
| 3319 | |
| 3320 | GGML_UNUSED(data_size)(void)(data_size); |
| 3321 | } |
| 3322 | |
| 3323 | static int repack_q2_K_to_q2_K_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3324 | GGML_ASSERT(t->type == GGML_TYPE_Q2_K)if (!(t->type == GGML_TYPE_Q2_K)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3324, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q2_K" ); |
| 3325 | constexpr int nrows_interleaved = 16; |
| 3326 | |
| 3327 | block_q2_Kx16 * dst = (block_q2_Kx16*)t->data; |
| 3328 | const block_q2_K * src = (const block_q2_K*) data; |
| 3329 | |
| 3330 | block_q2_K dst_tmp[nrows_interleaved]; |
| 3331 | |
| 3332 | int nrow = ggml_nrows(t); |
| 3333 | int nblocks = t->ne[0] / QK_K256; |
| 3334 | |
| 3335 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q2_K))if (!(data_size == nrow * nblocks * sizeof(block_q2_K))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3335, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q2_K)" ); |
| 3336 | |
| 3337 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3338 | return -1; |
| 3339 | } |
| 3340 | |
| 3341 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3342 | for (int64_t x = 0; x < nblocks; x++) { |
| 3343 | // This loop gathers 16 separate blocks (one from each column) |
| 3344 | // that correspond to the same K-dimension chunk. |
| 3345 | for (int i = 0; i < nrows_interleaved; i++ ) { |
| 3346 | dst_tmp[i] = src[x + i * nblocks]; |
| 3347 | } |
| 3348 | |
| 3349 | *dst++ = make_block_q2_Kx16(dst_tmp, interleave_block); |
| 3350 | } |
| 3351 | src += nrows_interleaved * nblocks; |
| 3352 | } |
| 3353 | return 0; |
| 3354 | |
| 3355 | GGML_UNUSED(data_size)(void)(data_size); |
| 3356 | } |
| 3357 | |
| 3358 | static int repack_q4_0_to_q4_0_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3359 | GGML_ASSERT(t->type == GGML_TYPE_Q4_0)if (!(t->type == GGML_TYPE_Q4_0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3359, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q4_0" ); |
| 3360 | constexpr int nrows_interleaved = 16; |
| 3361 | |
| 3362 | block_q4_0x16 * dst = (block_q4_0x16*)t->data; |
| 3363 | const block_q4_0 * src = (const block_q4_0*) data; |
| 3364 | block_q4_0 dst_tmp[16]; |
| 3365 | int nrow = ggml_nrows(t); |
| 3366 | int nblocks = t->ne[0] / QK4_032; |
| 3367 | |
| 3368 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0))if (!(data_size == nrow * nblocks * sizeof(block_q4_0))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3368, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q4_0)" ); |
| 3369 | |
| 3370 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3371 | return -1; |
| 3372 | } |
| 3373 | |
| 3374 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3375 | for (int64_t x = 0; x < nblocks; x++) { |
| 3376 | for (int i = 0; i < nrows_interleaved; i++ ) { |
| 3377 | dst_tmp[i] = src[x + i * nblocks]; |
| 3378 | } |
| 3379 | *dst++ = make_block_q4_0x16(dst_tmp, interleave_block); |
| 3380 | } |
| 3381 | src += nrows_interleaved * nblocks; |
| 3382 | } |
| 3383 | return 0; |
| 3384 | |
| 3385 | GGML_UNUSED(data_size)(void)(data_size); |
| 3386 | } |
| 3387 | |
| 3388 | static int repack_q5_K_to_q5_K_8_bl(struct ggml_tensor * t, |
| 3389 | int interleave_block, |
| 3390 | const void * GGML_RESTRICT__restrict__ data, |
| 3391 | size_t data_size) { |
| 3392 | GGML_ASSERT(t->type == GGML_TYPE_Q5_K)if (!(t->type == GGML_TYPE_Q5_K)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3392, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q5_K" ); |
| 3393 | GGML_ASSERT(interleave_block == 4 || interleave_block == 8)if (!(interleave_block == 4 || interleave_block == 8)) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3393, "GGML_ASSERT(%s) failed", "interleave_block == 4 || interleave_block == 8" ); |
| 3394 | constexpr int nrows_interleaved = 8; |
| 3395 | |
| 3396 | block_q5_Kx8 * dst = (block_q5_Kx8 *) t->data; |
| 3397 | const block_q5_K * src = (const block_q5_K *) data; |
| 3398 | block_q5_K dst_tmp[8]; |
| 3399 | int nrow = ggml_nrows(t); |
| 3400 | int nblocks = t->ne[0] / QK_K256; |
| 3401 | |
| 3402 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q5_K))if (!(data_size == nrow * nblocks * sizeof(block_q5_K))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3402, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q5_K)" ); |
| 3403 | |
| 3404 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3405 | return -1; |
| 3406 | } |
| 3407 | |
| 3408 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3409 | for (int64_t x = 0; x < nblocks; x++) { |
| 3410 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3411 | dst_tmp[i] = src[x + i * nblocks]; |
| 3412 | } |
| 3413 | *dst++ = make_block_q5_Kx8(dst_tmp, interleave_block); |
| 3414 | } |
| 3415 | src += nrows_interleaved * nblocks; |
| 3416 | } |
| 3417 | return 0; |
| 3418 | } |
| 3419 | |
| 3420 | static int repack_q6_K_to_q6_K_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3421 | GGML_ASSERT(t->type == GGML_TYPE_Q6_K)if (!(t->type == GGML_TYPE_Q6_K)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3421, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q6_K" ); |
| 3422 | GGML_ASSERT(interleave_block == 4 || interleave_block == 8)if (!(interleave_block == 4 || interleave_block == 8)) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3422, "GGML_ASSERT(%s) failed", "interleave_block == 4 || interleave_block == 8" ); |
| 3423 | constexpr int nrows_interleaved = 8; |
| 3424 | |
| 3425 | block_q6_Kx8 * dst = (block_q6_Kx8 *)t->data; |
| 3426 | const block_q6_K * src = (const block_q6_K *) data; |
| 3427 | block_q6_K dst_tmp[8]; |
| 3428 | int nrow = ggml_nrows(t); |
| 3429 | int nblocks = t->ne[0] / QK_K256; |
| 3430 | |
| 3431 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q6_K))if (!(data_size == nrow * nblocks * sizeof(block_q6_K))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3431, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q6_K)" ); |
| 3432 | |
| 3433 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3434 | return -1; |
| 3435 | } |
| 3436 | |
| 3437 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3438 | for (int64_t x = 0; x < nblocks; x++) { |
| 3439 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3440 | dst_tmp[i] = src[x + i * nblocks]; |
| 3441 | } |
| 3442 | *dst++ = make_block_q6_Kx8(dst_tmp, interleave_block); |
| 3443 | } |
| 3444 | src += nrows_interleaved * nblocks; |
| 3445 | } |
| 3446 | return 0; |
| 3447 | } |
| 3448 | |
| 3449 | static int repack_q4_0_to_q4_0_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3450 | GGML_ASSERT(t->type == GGML_TYPE_Q4_0)if (!(t->type == GGML_TYPE_Q4_0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3450, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q4_0" ); |
| 3451 | GGML_ASSERT(interleave_block == 8)if (!(interleave_block == 8)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3451, "GGML_ASSERT(%s) failed", "interleave_block == 8"); |
| 3452 | constexpr int nrows_interleaved = 8; |
| 3453 | |
| 3454 | block_q4_0x8 * dst = (block_q4_0x8*)t->data; |
| 3455 | const block_q4_0 * src = (const block_q4_0*) data; |
| 3456 | block_q4_0 dst_tmp[8]; |
| 3457 | int nrow = ggml_nrows(t); |
| 3458 | int nblocks = t->ne[0] / QK4_032; |
| 3459 | |
| 3460 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q4_0))if (!(data_size == nrow * nblocks * sizeof(block_q4_0))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3460, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q4_0)" ); |
| 3461 | |
| 3462 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3463 | return -1; |
| 3464 | } |
| 3465 | |
| 3466 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3467 | for (int64_t x = 0; x < nblocks; x++) { |
| 3468 | for (int i = 0; i < nrows_interleaved; i++ ) { |
| 3469 | dst_tmp[i] = src[x + i * nblocks]; |
| 3470 | } |
| 3471 | *dst++ = make_block_q4_0x8(dst_tmp, interleave_block); |
| 3472 | } |
| 3473 | src += nrows_interleaved * nblocks; |
| 3474 | } |
| 3475 | return 0; |
| 3476 | |
| 3477 | GGML_UNUSED(data_size)(void)(data_size); |
| 3478 | } |
| 3479 | |
| 3480 | static int repack_q8_0_to_q8_0_4_bl(struct ggml_tensor * t, |
| 3481 | int interleave_block, |
| 3482 | const void * GGML_RESTRICT__restrict__ data, |
| 3483 | size_t data_size) { |
| 3484 | GGML_ASSERT(t->type == GGML_TYPE_Q8_0)if (!(t->type == GGML_TYPE_Q8_0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3484, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q8_0" ); |
| 3485 | GGML_ASSERT(interleave_block == 4 || interleave_block == 8)if (!(interleave_block == 4 || interleave_block == 8)) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3485, "GGML_ASSERT(%s) failed", "interleave_block == 4 || interleave_block == 8" ); |
| 3486 | constexpr int nrows_interleaved = 4; |
| 3487 | |
| 3488 | block_q8_0x4 * dst = (block_q8_0x4 *) t->data; |
| 3489 | const block_q8_0 * src = (const block_q8_0 *) data; |
| 3490 | block_q8_0 dst_tmp[4]; |
| 3491 | int nrow = ggml_nrows(t); |
| 3492 | int nblocks = t->ne[0] / QK8_032; |
| 3493 | |
| 3494 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0))if (!(data_size == nrow * nblocks * sizeof(block_q8_0))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3494, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q8_0)" ); |
| 3495 | |
| 3496 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3497 | return -1; |
| 3498 | } |
| 3499 | |
| 3500 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3501 | for (int64_t x = 0; x < nblocks; x++) { |
| 3502 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3503 | dst_tmp[i] = src[x + i * nblocks]; |
| 3504 | } |
| 3505 | *dst++ = make_block_q8_0x4(dst_tmp, interleave_block); |
| 3506 | } |
| 3507 | src += nrows_interleaved * nblocks; |
| 3508 | } |
| 3509 | return 0; |
| 3510 | } |
| 3511 | |
| 3512 | static block_q8_0x16 make_block_q8_0x16(block_q8_0 * in, unsigned int blck_size_interleave) { |
| 3513 | block_q8_0x16 out; |
| 3514 | |
| 3515 | for (int i = 0; i < 16; i++) { |
| 3516 | out.d[i] = in[i].d; |
| 3517 | } |
| 3518 | |
| 3519 | const int end = QK8_032 * 16 / blck_size_interleave; |
| 3520 | |
| 3521 | if (blck_size_interleave == 1) { |
| 3522 | for (int i = 0; i < end; ++i) { |
| 3523 | int src_id = i % 16; |
| 3524 | int src_offset = i / 16; |
| 3525 | int dst_offset = i; |
| 3526 | out.qs[dst_offset] = in[src_id].qs[src_offset]; |
| 3527 | } |
| 3528 | } else { |
| 3529 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3529, "GGML_ASSERT(%s) failed", "false"); |
| 3530 | } |
| 3531 | |
| 3532 | return out; |
| 3533 | } |
| 3534 | |
| 3535 | static int repack_q8_0_to_q8_0_16_bl(struct ggml_tensor * t, |
| 3536 | int interleave_block, |
| 3537 | const void * GGML_RESTRICT__restrict__ data, |
| 3538 | size_t data_size) { |
| 3539 | GGML_ASSERT(t->type == GGML_TYPE_Q8_0)if (!(t->type == GGML_TYPE_Q8_0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3539, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_Q8_0" ); |
| 3540 | constexpr int nrows_interleaved = 16; |
| 3541 | |
| 3542 | block_q8_0x16 * dst = (block_q8_0x16 *) t->data; |
| 3543 | const block_q8_0 * src = (const block_q8_0 *) data; |
| 3544 | block_q8_0 dst_tmp[16]; |
| 3545 | int nrow = ggml_nrows(t); |
| 3546 | int nblocks = t->ne[0] / QK8_032; |
| 3547 | |
| 3548 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_q8_0))if (!(data_size == nrow * nblocks * sizeof(block_q8_0))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3548, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_q8_0)" ); |
| 3549 | |
| 3550 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3551 | return -1; |
| 3552 | } |
| 3553 | |
| 3554 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3555 | for (int64_t x = 0; x < nblocks; x++) { |
| 3556 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3557 | dst_tmp[i] = src[x + i * nblocks]; |
| 3558 | } |
| 3559 | *dst++ = make_block_q8_0x16(dst_tmp, interleave_block); |
| 3560 | } |
| 3561 | src += nrows_interleaved * nblocks; |
| 3562 | } |
| 3563 | return 0; |
| 3564 | } |
| 3565 | |
| 3566 | static block_iq4_nlx4 make_block_iq4_nlx4(block_iq4_nl * in, unsigned int blck_size_interleave) { |
| 3567 | block_iq4_nlx4 out; |
| 3568 | |
| 3569 | for (int i = 0; i < 4; i++) { |
| 3570 | out.d[i] = in[i].d; |
| 3571 | } |
| 3572 | |
| 3573 | const int end = QK4_NL32 * 2 / blck_size_interleave; |
| 3574 | |
| 3575 | // TODO: this branch seems wrong |
| 3576 | //if (blck_size_interleave == 8) { |
| 3577 | // for (int i = 0; i < end; ++i) { |
| 3578 | // int src_id = i % 4; |
| 3579 | // int src_offset = (i / 4) * blck_size_interleave; |
| 3580 | // int dst_offset = i * blck_size_interleave; |
| 3581 | |
| 3582 | // // Using memcpy to avoid unaligned memory accesses |
| 3583 | // memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); |
| 3584 | // } |
| 3585 | //} else |
| 3586 | if (blck_size_interleave == 4) { |
| 3587 | for (int i = 0; i < end; ++i) { |
| 3588 | int src_id = i % 4; |
| 3589 | int src_offset = (i / 4) * blck_size_interleave; |
| 3590 | int dst_offset = i * blck_size_interleave; |
| 3591 | |
| 3592 | memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t)); |
| 3593 | } |
| 3594 | } else { |
| 3595 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3595, "GGML_ASSERT(%s) failed", "false"); |
| 3596 | } |
| 3597 | |
| 3598 | return out; |
| 3599 | } |
| 3600 | |
| 3601 | static int repack_iq4_nl_to_iq4_nl_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3602 | GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL)if (!(t->type == GGML_TYPE_IQ4_NL)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3602, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_IQ4_NL" ); |
| 3603 | GGML_ASSERT(interleave_block == 4)if (!(interleave_block == 4)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3603, "GGML_ASSERT(%s) failed", "interleave_block == 4"); |
| 3604 | |
| 3605 | const block_iq4_nl * src = (const block_iq4_nl *)data; |
| 3606 | block_iq4_nlx4 * dst = ( block_iq4_nlx4 *)t->data; |
| 3607 | |
| 3608 | block_iq4_nl dst_tmp[4]; |
| 3609 | |
| 3610 | int nrow = ggml_nrows(t); |
| 3611 | int nrows_interleaved = 4; |
| 3612 | int nblocks = t->ne[0] / QK4_NL32; |
| 3613 | |
| 3614 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl))if (!(data_size == nrow * nblocks * sizeof(block_iq4_nl))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3614, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_iq4_nl)" ); |
| 3615 | |
| 3616 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3617 | return -1; |
| 3618 | } |
| 3619 | |
| 3620 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3621 | for (int64_t x = 0; x < nblocks; x++) { |
| 3622 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3623 | dst_tmp[i] = src[x + i * nblocks]; |
| 3624 | } |
| 3625 | *dst++ = make_block_iq4_nlx4(dst_tmp, interleave_block); |
| 3626 | } |
| 3627 | src += nrows_interleaved * nblocks; |
| 3628 | } |
| 3629 | return 0; |
| 3630 | |
| 3631 | GGML_UNUSED(data_size)(void)(data_size); |
| 3632 | } |
| 3633 | |
| 3634 | static block_iq4_nlx8 make_block_iq4_nlx8(block_iq4_nl * in, unsigned int blck_size_interleave) { |
| 3635 | block_iq4_nlx8 out; |
| 3636 | |
| 3637 | for (int i = 0; i < 8; i++) { |
| 3638 | out.d[i] = in[i].d; |
| 3639 | } |
| 3640 | |
| 3641 | const int end = QK4_NL32 * 4 / blck_size_interleave; |
| 3642 | |
| 3643 | if (blck_size_interleave == 8) { |
| 3644 | for (int i = 0; i < end; ++i) { |
| 3645 | int src_id = i % 8; |
| 3646 | int src_offset = (i / 8) * blck_size_interleave; |
| 3647 | int dst_offset = i * blck_size_interleave; |
| 3648 | |
| 3649 | memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); |
| 3650 | } |
| 3651 | } else { |
| 3652 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3652, "GGML_ASSERT(%s) failed", "false"); |
| 3653 | } |
| 3654 | |
| 3655 | return out; |
| 3656 | } |
| 3657 | |
| 3658 | static int repack_iq4_nl_to_iq4_nl_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3659 | GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL)if (!(t->type == GGML_TYPE_IQ4_NL)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3659, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_IQ4_NL" ); |
| 3660 | GGML_ASSERT(interleave_block == 8)if (!(interleave_block == 8)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3660, "GGML_ASSERT(%s) failed", "interleave_block == 8"); |
| 3661 | |
| 3662 | const block_iq4_nl * src = (const block_iq4_nl *)data; |
| 3663 | block_iq4_nlx8 * dst = ( block_iq4_nlx8 *)t->data; |
| 3664 | |
| 3665 | block_iq4_nl dst_tmp[8]; |
| 3666 | |
| 3667 | int nrow = ggml_nrows(t); |
| 3668 | int nrows_interleaved = 8; |
| 3669 | int nblocks = t->ne[0] / QK4_NL32; |
| 3670 | |
| 3671 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl))if (!(data_size == nrow * nblocks * sizeof(block_iq4_nl))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3671, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_iq4_nl)" ); |
| 3672 | |
| 3673 | if (t->ne[1] % nrows_interleaved != 0) { |
| 3674 | return -1; |
| 3675 | } |
| 3676 | |
| 3677 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3678 | for (int64_t x = 0; x < nblocks; x++) { |
| 3679 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3680 | dst_tmp[i] = src[x + i * nblocks]; |
| 3681 | } |
| 3682 | *dst++ = make_block_iq4_nlx8(dst_tmp, interleave_block); |
| 3683 | } |
| 3684 | src += nrows_interleaved * nblocks; |
| 3685 | } |
| 3686 | return 0; |
| 3687 | |
| 3688 | GGML_UNUSED(data_size)(void)(data_size); |
| 3689 | } |
| 3690 | |
| 3691 | static block_iq4_nlx16 make_block_iq4_nlx16(block_iq4_nl * in, unsigned int blck_size_interleave) { |
| 3692 | block_iq4_nlx16 out; |
| 3693 | |
| 3694 | for (int i = 0; i < 16; i++) { |
| 3695 | out.d[i] = in[i].d; |
| 3696 | } |
| 3697 | |
| 3698 | const int end = QK4_NL32 * 8 / blck_size_interleave; |
| 3699 | |
| 3700 | if (blck_size_interleave == 1) { |
| 3701 | for (int i = 0; i < end; ++i) { |
| 3702 | int src_id = i % 16; |
| 3703 | int src_offset = i / 16; |
| 3704 | int dst_offset = i; |
| 3705 | |
| 3706 | out.qs[dst_offset] = in[src_id].qs[src_offset]; |
| 3707 | } |
| 3708 | } else { |
| 3709 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3709, "GGML_ASSERT(%s) failed", "false"); |
| 3710 | } |
| 3711 | |
| 3712 | return out; |
| 3713 | } |
| 3714 | |
| 3715 | static int repack_iq4_nl_to_iq4_nl_16_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3716 | GGML_ASSERT(t->type == GGML_TYPE_IQ4_NL)if (!(t->type == GGML_TYPE_IQ4_NL)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3716, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_IQ4_NL" ); |
| 3717 | GGML_ASSERT(interleave_block == 1)if (!(interleave_block == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3717, "GGML_ASSERT(%s) failed", "interleave_block == 1"); |
| 3718 | |
| 3719 | const block_iq4_nl * src = (const block_iq4_nl *)data; |
| 3720 | block_iq4_nlx16 * dst = ( block_iq4_nlx16 *)t->data; |
| 3721 | |
| 3722 | block_iq4_nl dst_tmp[16]; |
| 3723 | |
| 3724 | int nrow = ggml_nrows(t); |
| 3725 | int nrows_interleaved = 16; |
| 3726 | int nblocks = t->ne[0] / QK4_NL32; |
| 3727 | |
| 3728 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_iq4_nl))if (!(data_size == nrow * nblocks * sizeof(block_iq4_nl))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3728, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_iq4_nl)" ); |
| 3729 | |
| 3730 | if (t->ne[1] % nrows_interleaved != 0) { |
| 3731 | return -1; |
| 3732 | } |
| 3733 | |
| 3734 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3735 | for (int64_t x = 0; x < nblocks; x++) { |
| 3736 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3737 | dst_tmp[i] = src[x + i * nblocks]; |
| 3738 | } |
| 3739 | *dst++ = make_block_iq4_nlx16(dst_tmp, interleave_block); |
| 3740 | } |
| 3741 | src += nrows_interleaved * nblocks; |
| 3742 | } |
| 3743 | return 0; |
| 3744 | |
| 3745 | GGML_UNUSED(data_size)(void)(data_size); |
| 3746 | } |
| 3747 | |
| 3748 | static block_mxfp4x4 make_block_mxfp4x4(block_mxfp4 * in, unsigned int blck_size_interleave) { |
| 3749 | block_mxfp4x4 out; |
| 3750 | |
| 3751 | for (int i = 0; i < 4; i++) { |
| 3752 | out.e[i] = in[i].e; |
| 3753 | } |
| 3754 | |
| 3755 | const int end = QK_MXFP432 * 2 / blck_size_interleave; |
| 3756 | |
| 3757 | if (blck_size_interleave == 4) { |
| 3758 | for (int i = 0; i < end; ++i) { |
| 3759 | int src_id = i % 4; |
| 3760 | int src_offset = (i / 4) * blck_size_interleave; |
| 3761 | int dst_offset = i * blck_size_interleave; |
| 3762 | |
| 3763 | memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint32_t)); |
| 3764 | } |
| 3765 | } else { |
| 3766 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3766, "GGML_ASSERT(%s) failed", "false"); |
| 3767 | } |
| 3768 | |
| 3769 | return out; |
| 3770 | } |
| 3771 | |
| 3772 | static int repack_mxfp4_to_mxfp4_4_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3773 | GGML_ASSERT(t->type == GGML_TYPE_MXFP4)if (!(t->type == GGML_TYPE_MXFP4)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3773, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_MXFP4" ); |
| 3774 | GGML_ASSERT(interleave_block == 4)if (!(interleave_block == 4)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3774, "GGML_ASSERT(%s) failed", "interleave_block == 4"); |
| 3775 | |
| 3776 | const block_mxfp4 * src = (const block_mxfp4 *)data; |
| 3777 | block_mxfp4x4 * dst = ( block_mxfp4x4 *)t->data; |
| 3778 | |
| 3779 | block_mxfp4 dst_tmp[4]; |
| 3780 | |
| 3781 | int nrow = ggml_nrows(t); |
| 3782 | int nrows_interleaved = 4; |
| 3783 | int nblocks = t->ne[0] / QK_MXFP432; |
| 3784 | |
| 3785 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4))if (!(data_size == nrow * nblocks * sizeof(block_mxfp4))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3785, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_mxfp4)" ); |
| 3786 | |
| 3787 | if (t->ne[1] % nrows_interleaved != 0 || t->ne[0] % 8 != 0) { |
| 3788 | return -1; |
| 3789 | } |
| 3790 | |
| 3791 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3792 | for (int64_t x = 0; x < nblocks; x++) { |
| 3793 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3794 | dst_tmp[i] = src[x + i * nblocks]; |
| 3795 | } |
| 3796 | *dst++ = make_block_mxfp4x4(dst_tmp, interleave_block); |
| 3797 | } |
| 3798 | src += nrows_interleaved * nblocks; |
| 3799 | } |
| 3800 | return 0; |
| 3801 | |
| 3802 | GGML_UNUSED(data_size)(void)(data_size); |
| 3803 | } |
| 3804 | |
| 3805 | static block_mxfp4x8 make_block_mxfp4x8(block_mxfp4 * in, unsigned int blck_size_interleave) { |
| 3806 | block_mxfp4x8 out; |
| 3807 | |
| 3808 | for (int i = 0; i < 8; i++) { |
| 3809 | out.e[i] = in[i].e; |
| 3810 | } |
| 3811 | |
| 3812 | const int end = QK_MXFP432 * 4 / blck_size_interleave; |
| 3813 | |
| 3814 | if (blck_size_interleave == 8) { |
| 3815 | for (int i = 0; i < end; ++i) { |
| 3816 | int src_id = i % 8; |
| 3817 | int src_offset = (i / 8) * blck_size_interleave; |
| 3818 | int dst_offset = i * blck_size_interleave; |
| 3819 | |
| 3820 | memcpy(&out.qs[dst_offset], &in[src_id].qs[src_offset], sizeof(uint64_t)); |
| 3821 | } |
| 3822 | } else { |
| 3823 | GGML_ASSERT(false)if (!(false)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3823, "GGML_ASSERT(%s) failed", "false"); |
| 3824 | } |
| 3825 | |
| 3826 | return out; |
| 3827 | } |
| 3828 | |
| 3829 | static int repack_mxfp4_to_mxfp4_8_bl(struct ggml_tensor * t, int interleave_block, const void * GGML_RESTRICT__restrict__ data, size_t data_size) { |
| 3830 | GGML_ASSERT(t->type == GGML_TYPE_MXFP4)if (!(t->type == GGML_TYPE_MXFP4)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3830, "GGML_ASSERT(%s) failed", "t->type == GGML_TYPE_MXFP4" ); |
| 3831 | GGML_ASSERT(interleave_block == 8)if (!(interleave_block == 8)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3831, "GGML_ASSERT(%s) failed", "interleave_block == 8"); |
| 3832 | |
| 3833 | const block_mxfp4 * src = (const block_mxfp4 *)data; |
| 3834 | block_mxfp4x8 * dst = ( block_mxfp4x8 *)t->data; |
| 3835 | |
| 3836 | block_mxfp4 dst_tmp[8]; |
| 3837 | |
| 3838 | int nrow = ggml_nrows(t); |
| 3839 | int nrows_interleaved = 8; |
| 3840 | int nblocks = t->ne[0] / QK_MXFP432; |
| 3841 | |
| 3842 | GGML_ASSERT(data_size == nrow * nblocks * sizeof(block_mxfp4))if (!(data_size == nrow * nblocks * sizeof(block_mxfp4))) ggml_abort ("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 3842, "GGML_ASSERT(%s) failed", "data_size == nrow * nblocks * sizeof(block_mxfp4)" ); |
| 3843 | |
| 3844 | if (t->ne[1] % nrows_interleaved != 0) { |
| 3845 | return -1; |
| 3846 | } |
| 3847 | |
| 3848 | for (int b = 0; b < nrow; b += nrows_interleaved) { |
| 3849 | for (int64_t x = 0; x < nblocks; x++) { |
| 3850 | for (int i = 0; i < nrows_interleaved; i++) { |
| 3851 | dst_tmp[i] = src[x + i * nblocks]; |
| 3852 | } |
| 3853 | *dst++ = make_block_mxfp4x8(dst_tmp, interleave_block); |
| 3854 | } |
| 3855 | src += nrows_interleaved * nblocks; |
| 3856 | } |
| 3857 | return 0; |
| 3858 | |
| 3859 | GGML_UNUSED(data_size)(void)(data_size); |
| 3860 | } |
| 3861 | |
| 3862 | namespace ggml::cpu::repack { |
| 3863 | // repack |
| 3864 | template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS> |
| 3865 | int repack(struct ggml_tensor *, const void *, size_t); |
| 3866 | |
| 3867 | // TODO: generalise. |
| 3868 | template <> int repack<block_q4_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3869 | return repack_q4_0_to_q4_0_4_bl(t, 4, data, data_size); |
| 3870 | } |
| 3871 | |
| 3872 | template <> int repack<block_q4_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3873 | return repack_q4_0_to_q4_0_4_bl(t, 8, data, data_size); |
| 3874 | } |
| 3875 | |
| 3876 | template <> int repack<block_q4_0, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3877 | return repack_q4_0_to_q4_0_8_bl(t, 8, data, data_size); |
| 3878 | } |
| 3879 | |
| 3880 | template <> int repack<block_q4_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3881 | return repack_q4_K_to_q4_K_8_bl(t, 8, data, data_size); |
| 3882 | } |
| 3883 | |
| 3884 | template <> int repack<block_q4_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3885 | return repack_q4_K_to_q4_K_8_bl(t, 4, data, data_size); |
| 3886 | } |
| 3887 | |
| 3888 | template <> int repack<block_q2_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3889 | return repack_q2_K_to_q2_K_8_bl(t, 8, data, data_size); |
| 3890 | } |
| 3891 | |
| 3892 | template <> int repack<block_q5_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3893 | return repack_q5_K_to_q5_K_8_bl(t, 4, data, data_size); |
| 3894 | } |
| 3895 | |
| 3896 | template <> int repack<block_q5_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3897 | return repack_q5_K_to_q5_K_8_bl(t, 8, data, data_size); |
| 3898 | } |
| 3899 | |
| 3900 | template <> int repack<block_q6_K, 4, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3901 | return repack_q6_K_to_q6_K_8_bl(t, 4, data, data_size); |
| 3902 | } |
| 3903 | |
| 3904 | template <> int repack<block_q6_K, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3905 | return repack_q6_K_to_q6_K_8_bl(t, 8, data, data_size); |
| 3906 | } |
| 3907 | |
| 3908 | template <> int repack<block_iq4_nl, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3909 | return repack_iq4_nl_to_iq4_nl_4_bl(t, 4, data, data_size); |
| 3910 | } |
| 3911 | |
| 3912 | // TODO: needs to be revisited |
| 3913 | //template <> int repack<block_iq4_nl, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3914 | // return repack_iq4_nl_to_iq4_nl_4_bl(t, 8, data, data_size); |
| 3915 | //} |
| 3916 | |
| 3917 | template <> int repack<block_iq4_nl, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3918 | return repack_iq4_nl_to_iq4_nl_8_bl(t, 8, data, data_size); |
| 3919 | } |
| 3920 | |
| 3921 | template <> int repack<block_mxfp4, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3922 | return repack_mxfp4_to_mxfp4_4_bl(t, 4, data, data_size); |
| 3923 | } |
| 3924 | |
| 3925 | template <> int repack<block_mxfp4, 8, 8>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3926 | return repack_mxfp4_to_mxfp4_8_bl(t, 8, data, data_size); |
| 3927 | } |
| 3928 | |
| 3929 | template <> int repack<block_q8_0, 4, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3930 | return repack_q8_0_to_q8_0_4_bl(t, 4, data, data_size); |
| 3931 | } |
| 3932 | |
| 3933 | template <> int repack<block_q8_0, 8, 4>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3934 | return repack_q8_0_to_q8_0_4_bl(t, 8, data, data_size); |
| 3935 | } |
| 3936 | |
| 3937 | #if defined __riscv_zvfh |
| 3938 | template <> int repack<block_q4_0, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3939 | return repack_q4_0_to_q4_0_16_bl(t, 1, data, data_size); |
| 3940 | } |
| 3941 | |
| 3942 | template <> int repack<block_q4_K, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3943 | return repack_q4_K_to_q4_K_16_bl(t, 1, data, data_size); |
| 3944 | } |
| 3945 | |
| 3946 | template <> int repack<block_iq4_nl, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3947 | return repack_iq4_nl_to_iq4_nl_16_bl(t, 1, data, data_size); |
| 3948 | } |
| 3949 | |
| 3950 | template <> int repack<block_q8_0, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3951 | return repack_q8_0_to_q8_0_16_bl(t, 1, data, data_size); |
| 3952 | } |
| 3953 | |
| 3954 | template <> int repack<block_q2_K, 1, 16>(struct ggml_tensor * t, const void * data, size_t data_size) { |
| 3955 | return repack_q2_K_to_q2_K_16_bl(t, 1, data, data_size); |
| 3956 | } |
| 3957 | #endif |
| 3958 | |
| 3959 | // gemv |
| 3960 | template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE> |
| 3961 | void gemv(int, float *, size_t, const void *, const void *, int, int); |
| 3962 | |
| 3963 | template <> void gemv<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3964 | ggml_gemv_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 3965 | } |
| 3966 | |
| 3967 | template <> void gemv<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3968 | ggml_gemv_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 3969 | } |
| 3970 | |
| 3971 | template <> void gemv<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3972 | ggml_gemv_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 3973 | } |
| 3974 | |
| 3975 | template <> |
| 3976 | void gemv<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, |
| 3977 | float * s, |
| 3978 | size_t bs, |
| 3979 | const void * vx, |
| 3980 | const void * vy, |
| 3981 | int nr, |
| 3982 | int nc) { |
| 3983 | ggml_gemv_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 3984 | } |
| 3985 | |
| 3986 | template <> void gemv<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3987 | ggml_gemv_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc); |
| 3988 | } |
| 3989 | |
| 3990 | template <> void gemv<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3991 | ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 3992 | } |
| 3993 | |
| 3994 | template <> void gemv<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3995 | ggml_gemv_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc); |
| 3996 | } |
| 3997 | |
| 3998 | template <> void gemv<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 3999 | ggml_gemv_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4000 | } |
| 4001 | |
| 4002 | template <> void gemv<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4003 | ggml_gemv_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4004 | } |
| 4005 | |
| 4006 | template <> void gemv<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4007 | ggml_gemv_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4008 | } |
| 4009 | |
| 4010 | template <> void gemv<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4011 | ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4012 | } |
| 4013 | |
| 4014 | template <> void gemv<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4015 | ggml_gemv_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4016 | } |
| 4017 | |
| 4018 | template <> void gemv<block_mxfp4, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4019 | ggml_gemv_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4020 | } |
| 4021 | |
| 4022 | template <> void gemv<block_mxfp4, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4023 | ggml_gemv_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4024 | } |
| 4025 | |
| 4026 | template <> void gemv<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4027 | ggml_gemv_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4028 | } |
| 4029 | |
| 4030 | template <> void gemv<block_q8_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4031 | ggml_gemv_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4032 | } |
| 4033 | |
| 4034 | #if defined __riscv_zvfh |
| 4035 | template <> void gemv<block_q4_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4036 | ggml_gemv_q4_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4037 | } |
| 4038 | |
| 4039 | template <> void gemv<block_q4_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4040 | ggml_gemv_q4_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4041 | } |
| 4042 | |
| 4043 | template <> void gemv<block_iq4_nl, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4044 | ggml_gemv_iq4_nl_16x1_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4045 | } |
| 4046 | |
| 4047 | template <> void gemv<block_q8_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4048 | ggml_gemv_q8_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4049 | } |
| 4050 | |
| 4051 | template <> void gemv<block_q2_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4052 | ggml_gemv_q2_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4053 | } |
| 4054 | #endif |
| 4055 | |
| 4056 | // gemm |
| 4057 | template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE> |
| 4058 | void gemm(int, float *, size_t, const void *, const void *, int, int); |
| 4059 | |
| 4060 | template <> void gemm<block_q4_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4061 | ggml_gemm_q4_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4062 | } |
| 4063 | |
| 4064 | template <> void gemm<block_q4_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4065 | ggml_gemm_q4_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4066 | } |
| 4067 | |
| 4068 | template <> |
| 4069 | void gemm<block_q4_0, 8, 8, GGML_TYPE_Q8_0>(int n, |
| 4070 | float * s, |
| 4071 | size_t bs, |
| 4072 | const void * vx, |
| 4073 | const void * vy, |
| 4074 | int nr, |
| 4075 | int nc) { |
| 4076 | ggml_gemm_q4_0_8x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4077 | } |
| 4078 | |
| 4079 | template <> void gemm<block_q2_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4080 | ggml_gemm_q2_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4081 | } |
| 4082 | |
| 4083 | template <> void gemm<block_q4_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4084 | ggml_gemm_q4_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4085 | } |
| 4086 | |
| 4087 | template <> void gemm<block_q4_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4088 | ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4089 | } |
| 4090 | |
| 4091 | template <> void gemm<block_q5_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4092 | ggml_gemm_q5_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4093 | } |
| 4094 | |
| 4095 | template <> void gemm<block_q5_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4096 | ggml_gemm_q5_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4097 | } |
| 4098 | |
| 4099 | template <> void gemm<block_q6_K, 4, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4100 | ggml_gemm_q6_K_8x4_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4101 | } |
| 4102 | |
| 4103 | template <> void gemm<block_q6_K, 8, 8, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4104 | ggml_gemm_q6_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4105 | } |
| 4106 | |
| 4107 | template <> void gemm<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4108 | ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4109 | } |
| 4110 | |
| 4111 | template <> void gemm<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4112 | ggml_gemm_iq4_nl_8x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4113 | } |
| 4114 | |
| 4115 | template <> void gemm<block_mxfp4, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4116 | ggml_gemm_mxfp4_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4117 | } |
| 4118 | |
| 4119 | template <> void gemm<block_mxfp4, 8, 8, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4120 | ggml_gemm_mxfp4_8x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4121 | } |
| 4122 | |
| 4123 | template <> void gemm<block_q8_0, 4, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4124 | ggml_gemm_q8_0_4x4_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4125 | } |
| 4126 | |
| 4127 | template <> void gemm<block_q8_0, 8, 4, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4128 | ggml_gemm_q8_0_4x8_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4129 | } |
| 4130 | |
| 4131 | #if defined __riscv_zvfh |
| 4132 | template <> void gemm<block_q4_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4133 | ggml_gemm_q4_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4134 | } |
| 4135 | |
| 4136 | template <> void gemm<block_q4_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4137 | ggml_gemm_q4_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4138 | } |
| 4139 | |
| 4140 | template <> void gemm<block_iq4_nl, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4141 | ggml_gemm_iq4_nl_16x1_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4142 | } |
| 4143 | |
| 4144 | template <> void gemm<block_q8_0, 1, 16, GGML_TYPE_Q8_0>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4145 | ggml_gemm_q8_0_16x1_q8_0(n, s, bs, vx, vy, nr, nc); |
| 4146 | } |
| 4147 | |
| 4148 | template <> void gemm<block_q2_K, 1, 16, GGML_TYPE_Q8_K>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) { |
| 4149 | ggml_gemm_q2_K_16x1_q8_K(n, s, bs, vx, vy, nr, nc); |
| 4150 | } |
| 4151 | #endif |
| 4152 | |
| 4153 | class tensor_traits_base : public ggml::cpu::tensor_traits { |
| 4154 | public: |
| 4155 | virtual int repack(struct ggml_tensor * t, const void * data, size_t data_size) = 0; |
| 4156 | }; |
| 4157 | |
| 4158 | template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PARAM_TYPE> class tensor_traits : public tensor_traits_base { |
| 4159 | |
| 4160 | bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { |
| 4161 | // not realy a GGML_TYPE_Q8_0 but same size. |
| 4162 | switch (op->op) { |
| 4163 | case GGML_OP_MUL_MAT: |
| 4164 | { |
| 4165 | size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); |
| 4166 | return true; |
| 4167 | } |
| 4168 | case GGML_OP_MUL_MAT_ID: |
| 4169 | { |
| 4170 | size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); |
| 4171 | size = GGML_PAD(size, sizeof(int64_t))(((size) + (sizeof(int64_t)) - 1) & ~((sizeof(int64_t)) - 1)); // + padding for next block. |
| 4172 | |
| 4173 | const int64_t ne02 = op->src[0]->ne[2]; // n_as, n_expert |
| 4174 | const int64_t ne12 = op->src[1]->ne[2]; // n_tokens |
| 4175 | |
| 4176 | const size_t sizeof_mmid_row_mapping = sizeof(int64_t); |
| 4177 | |
| 4178 | size += sizeof_mmid_row_mapping*ne02*(ne12 + 1); |
| 4179 | |
| 4180 | return true; |
| 4181 | } |
| 4182 | default: |
| 4183 | // GGML_ABORT("fatal error"); |
| 4184 | break; |
| 4185 | } |
| 4186 | return false; |
| 4187 | } |
| 4188 | |
| 4189 | bool compute_forward(struct ggml_compute_params * params, struct ggml_tensor * op) override { |
| 4190 | switch (op->op) { |
| 4191 | case GGML_OP_MUL_MAT: |
| 4192 | forward_mul_mat(params, op); |
| 4193 | return true; |
| 4194 | case GGML_OP_MUL_MAT_ID: |
| 4195 | forward_mul_mat_id(params, op); |
| 4196 | return true; |
| 4197 | default: |
| 4198 | // GGML_ABORT("fatal error"); |
| 4199 | break; |
| 4200 | } |
| 4201 | return false; |
| 4202 | } |
| 4203 | |
| 4204 | void forward_mul_mat_one_chunk(ggml_compute_params * params, |
| 4205 | ggml_tensor * op, |
| 4206 | int64_t src0_start, |
| 4207 | int64_t src0_end, |
| 4208 | int64_t src1_start, |
| 4209 | int64_t src1_end) { |
| 4210 | const ggml_tensor * src0 = op->src[0]; |
| 4211 | const ggml_tensor * src1 = op->src[1]; |
| 4212 | ggml_tensor * dst = op; |
| 4213 | |
| 4214 | GGML_TENSOR_BINARY_OP_LOCALSconst int64_t ne00 = (src0) ? (src0)->ne[0] : 0; (void)(ne00 ); const int64_t ne01 = (src0) ? (src0)->ne[1] : 0; (void) (ne01); const int64_t ne02 = (src0) ? (src0)->ne[2] : 0; ( void)(ne02); const int64_t ne03 = (src0) ? (src0)->ne[3] : 0; (void)(ne03); const size_t nb00 = (src0) ? (src0)->nb[ 0] : 0; (void)(nb00); const size_t nb01 = (src0) ? (src0)-> nb[1] : 0; (void)(nb01); const size_t nb02 = (src0) ? (src0)-> nb[2] : 0; (void)(nb02); const size_t nb03 = (src0) ? (src0)-> nb[3] : 0; (void)(nb03); const int64_t ne10 = (src1) ? (src1) ->ne[0] : 0; (void)(ne10); const int64_t ne11 = (src1) ? ( src1)->ne[1] : 0; (void)(ne11); const int64_t ne12 = (src1 ) ? (src1)->ne[2] : 0; (void)(ne12); const int64_t ne13 = ( src1) ? (src1)->ne[3] : 0; (void)(ne13); const size_t nb10 = (src1) ? (src1)->nb[0] : 0; (void)(nb10); const size_t nb11 = (src1) ? (src1)->nb[1] : 0; (void)(nb11); const size_t nb12 = (src1) ? (src1)->nb[2] : 0; (void)(nb12); const size_t nb13 = (src1) ? (src1)->nb[3] : 0; (void)(nb13); const int64_t ne0 = (dst) ? (dst)->ne[0] : 0; (void)(ne0); const int64_t ne1 = (dst) ? (dst)->ne[1] : 0; (void)(ne1); const int64_t ne2 = (dst) ? (dst)->ne[2] : 0; (void)(ne2); const int64_t ne3 = (dst) ? (dst)->ne[3] : 0; (void)(ne3); const size_t nb0 = (dst) ? (dst)->nb[0] : 0; (void)(nb0); const size_t nb1 = (dst) ? (dst)->nb[1] : 0; (void)(nb1); const size_t nb2 = (dst) ? (dst)->nb[2] : 0; (void)(nb2); const size_t nb3 = (dst) ? (dst)->nb[3] : 0; (void)(nb3); |
| 4215 | |
| 4216 | const size_t src1_col_stride = ggml_row_size(PARAM_TYPE, ne10); |
| 4217 | |
| 4218 | GGML_ASSERT(ne03 == 1 && ne13 == 1)if (!(ne03 == 1 && ne13 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4218, "GGML_ASSERT(%s) failed", "ne03 == 1 && ne13 == 1" ); |
| 4219 | GGML_ASSERT(ne12 % ne02 == 0)if (!(ne12 % ne02 == 0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4219, "GGML_ASSERT(%s) failed", "ne12 % ne02 == 0"); |
| 4220 | const int64_t r2 = ne12 / ne02; |
| 4221 | |
| 4222 | const int64_t i12 = src1_start / ne1; |
| 4223 | const int64_t i11 = src1_start - i12 * ne1; |
| 4224 | |
| 4225 | // Determine batch index |
| 4226 | const int64_t i02 = i12 / r2; |
| 4227 | |
| 4228 | const int64_t i1 = i11; |
| 4229 | const int64_t i2 = i12; |
| 4230 | |
| 4231 | const char * src0_ptr = (const char *) src0->data + i02 * nb02; |
| 4232 | const char * src1_ptr = (const char *) params->wdata + (i11 + i12 * ne11) * src1_col_stride; |
| 4233 | char * dst_ptr = ((char *) dst->data + (i1 * nb1 + i2 * nb2)); |
| 4234 | |
| 4235 | const int64_t nrows = src1_end - src1_start; |
| 4236 | const int64_t ncols = src0_end - src0_start; |
| 4237 | |
| 4238 | GGML_ASSERT(src1_ptr + src1_col_stride * nrows <= (const char *) params->wdata + params->wsize)if (!(src1_ptr + src1_col_stride * nrows <= (const char *) params->wdata + params->wsize)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4238, "GGML_ASSERT(%s) failed", "src1_ptr + src1_col_stride * nrows <= (const char *) params->wdata + params->wsize" ); |
| 4239 | |
| 4240 | // If there are more than three rows in src1, use gemm; otherwise, use gemv. |
| 4241 | if (nrows > 3) { |
| 4242 | gemm<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00, (float *) (dst_ptr) + src0_start, nb1 / nb0, |
| 4243 | src0_ptr + src0_start * nb01, src1_ptr, |
| 4244 | nrows - (nrows % 4), ncols); |
| 4245 | } |
| 4246 | for (int iter = nrows - (nrows % 4); iter < nrows; iter++) { |
| 4247 | gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>(ne00, (float *) (dst_ptr + (iter * nb1)) + src0_start, |
| 4248 | ne01, src0_ptr + src0_start * nb01, |
| 4249 | src1_ptr + (src1_col_stride * iter), 1 /* nrows */, ncols); |
| 4250 | } |
| 4251 | } |
| 4252 | |
| 4253 | void forward_mul_mat(ggml_compute_params * params, ggml_tensor * op) { |
| 4254 | const ggml_tensor * src0 = op->src[0]; |
| 4255 | const ggml_tensor * src1 = op->src[1]; |
| 4256 | ggml_tensor * dst = op; |
| 4257 | |
| 4258 | GGML_TENSOR_BINARY_OP_LOCALSconst int64_t ne00 = (src0) ? (src0)->ne[0] : 0; (void)(ne00 ); const int64_t ne01 = (src0) ? (src0)->ne[1] : 0; (void) (ne01); const int64_t ne02 = (src0) ? (src0)->ne[2] : 0; ( void)(ne02); const int64_t ne03 = (src0) ? (src0)->ne[3] : 0; (void)(ne03); const size_t nb00 = (src0) ? (src0)->nb[ 0] : 0; (void)(nb00); const size_t nb01 = (src0) ? (src0)-> nb[1] : 0; (void)(nb01); const size_t nb02 = (src0) ? (src0)-> nb[2] : 0; (void)(nb02); const size_t nb03 = (src0) ? (src0)-> nb[3] : 0; (void)(nb03); const int64_t ne10 = (src1) ? (src1) ->ne[0] : 0; (void)(ne10); const int64_t ne11 = (src1) ? ( src1)->ne[1] : 0; (void)(ne11); const int64_t ne12 = (src1 ) ? (src1)->ne[2] : 0; (void)(ne12); const int64_t ne13 = ( src1) ? (src1)->ne[3] : 0; (void)(ne13); const size_t nb10 = (src1) ? (src1)->nb[0] : 0; (void)(nb10); const size_t nb11 = (src1) ? (src1)->nb[1] : 0; (void)(nb11); const size_t nb12 = (src1) ? (src1)->nb[2] : 0; (void)(nb12); const size_t nb13 = (src1) ? (src1)->nb[3] : 0; (void)(nb13); const int64_t ne0 = (dst) ? (dst)->ne[0] : 0; (void)(ne0); const int64_t ne1 = (dst) ? (dst)->ne[1] : 0; (void)(ne1); const int64_t ne2 = (dst) ? (dst)->ne[2] : 0; (void)(ne2); const int64_t ne3 = (dst) ? (dst)->ne[3] : 0; (void)(ne3); const size_t nb0 = (dst) ? (dst)->nb[0] : 0; (void)(nb0); const size_t nb1 = (dst) ? (dst)->nb[1] : 0; (void)(nb1); const size_t nb2 = (dst) ? (dst)->nb[2] : 0; (void)(nb2); const size_t nb3 = (dst) ? (dst)->nb[3] : 0; (void)(nb3); |
| 4259 | |
| 4260 | const int ith = params->ith; |
| 4261 | const int nth = params->nth; |
| 4262 | |
| 4263 | GGML_ASSERT(ne0 == ne01)if (!(ne0 == ne01)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4263, "GGML_ASSERT(%s) failed", "ne0 == ne01"); |
| 4264 | GGML_ASSERT(ne1 == ne11)if (!(ne1 == ne11)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4264, "GGML_ASSERT(%s) failed", "ne1 == ne11"); |
| 4265 | GGML_ASSERT(ne2 == ne12)if (!(ne2 == ne12)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4265, "GGML_ASSERT(%s) failed", "ne2 == ne12"); |
| 4266 | GGML_ASSERT(ne3 == ne13)if (!(ne3 == ne13)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4266, "GGML_ASSERT(%s) failed", "ne3 == ne13"); |
| 4267 | |
| 4268 | // dst cannot be transposed or permuted |
| 4269 | GGML_ASSERT(nb0 == sizeof(float))if (!(nb0 == sizeof(float))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4269, "GGML_ASSERT(%s) failed", "nb0 == sizeof(float)"); |
| 4270 | GGML_ASSERT(nb0 <= nb1)if (!(nb0 <= nb1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4270, "GGML_ASSERT(%s) failed", "nb0 <= nb1"); |
| 4271 | GGML_ASSERT(nb1 <= nb2)if (!(nb1 <= nb2)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4271, "GGML_ASSERT(%s) failed", "nb1 <= nb2"); |
| 4272 | GGML_ASSERT(nb2 <= nb3)if (!(nb2 <= nb3)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4272, "GGML_ASSERT(%s) failed", "nb2 <= nb3"); |
| 4273 | |
| 4274 | // TODO: General batched mul mat for 4D tensors |
| 4275 | // Currently only supports 3D tensors |
| 4276 | GGML_ASSERT(ne03 == 1)if (!(ne03 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4276, "GGML_ASSERT(%s) failed", "ne03 == 1"); |
| 4277 | GGML_ASSERT(ne13 == 1)if (!(ne13 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4277, "GGML_ASSERT(%s) failed", "ne13 == 1"); |
| 4278 | GGML_ASSERT(ne3 == 1)if (!(ne3 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4278, "GGML_ASSERT(%s) failed", "ne3 == 1"); |
| 4279 | |
| 4280 | GGML_ASSERT(src1->type == GGML_TYPE_F32)if (!(src1->type == GGML_TYPE_F32)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4280, "GGML_ASSERT(%s) failed", "src1->type == GGML_TYPE_F32" ); |
| 4281 | |
| 4282 | GGML_ASSERT(ggml_n_dims(op->src[0]) == 2)if (!(ggml_n_dims(op->src[0]) == 2)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4282, "GGML_ASSERT(%s) failed", "ggml_n_dims(op->src[0]) == 2" ); |
| 4283 | // GGML_ASSERT(ggml_n_dims(op->src[1]) == 2); |
| 4284 | |
| 4285 | char * wdata = static_cast<char *>(params->wdata); |
| 4286 | const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10); |
| 4287 | const size_t nbw2 = nbw1 * ne11; |
| 4288 | |
| 4289 | assert(params->wsize >= nbw2 * ne12)(static_cast <bool> (params->wsize >= nbw2 * ne12 ) ? void (0) : __assert_fail ("params->wsize >= nbw2 * ne12" , __builtin_FILE (), __builtin_LINE (), __extension__ __PRETTY_FUNCTION__ )); |
| 4290 | |
| 4291 | const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float; |
| 4292 | |
| 4293 | // INFO: Quantization is done in planes to avoid extra complexity in chunking. |
| 4294 | // Flattening dimensions not multiple of INTER_SIZE would require extra handling depending on how |
| 4295 | // the planes are broadcast. |
| 4296 | for (int64_t i12 = 0; i12 < ne12; i12++) { |
| 4297 | char * data_ptr = (char *) src1->data + i12 * nb12; |
| 4298 | char * wdata_ptr = wdata + i12 * nbw2; |
| 4299 | |
| 4300 | for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { |
| 4301 | ggml_quantize_mat_t<INTER_SIZE, PARAM_TYPE>((float *) (data_ptr + i11 * nb11), |
| 4302 | (void *) (wdata_ptr + i11 * nbw1), 4, ne10); |
| 4303 | } |
| 4304 | |
| 4305 | const int64_t i11_processed = ne11 - ne11 % 4; |
| 4306 | for (int64_t i11 = i11_processed + ith; i11 < ne11; i11 += nth) { |
| 4307 | from_float((float *) (data_ptr + i11 * nb11), (void *) (wdata_ptr + i11 * nbw1), ne10); |
| 4308 | } |
| 4309 | } |
| 4310 | |
| 4311 | // disable for NUMA |
| 4312 | const bool disable_chunking = ggml_is_numa(); |
| 4313 | |
| 4314 | // 4x chunks per thread |
| 4315 | const int64_t nr0 = ggml_nrows(op->src[0]); |
| 4316 | |
| 4317 | int nth_scaled = nth * 4; |
| 4318 | int64_t chunk_size0 = (nr0 + nth_scaled - 1) / nth_scaled; |
| 4319 | int64_t nchunk0 = (nr0 + chunk_size0 - 1) / chunk_size0; |
| 4320 | |
| 4321 | // src1 is chunked only by full planes. |
| 4322 | // When we flatten we need to address dimensions not multiple of the q8 INTER_SIZE |
| 4323 | // to route them thorugh GEMV. |
| 4324 | // nchunk1 = ne12 also avoids messing the chunking for models with no 3d tensors |
| 4325 | // to avoid affecting their performance |
| 4326 | int64_t nchunk1 = ne12; |
| 4327 | |
| 4328 | // Ensure minimum chunk size to avoid alignment issues with high thread counts |
| 4329 | // Minimum chunk size should be at least NB_COLS to prevent overlapping chunks after alignment |
| 4330 | const int64_t min_chunk_size = NB_COLS; |
| 4331 | if (nchunk0 > 0 && (nr0 / nchunk0) < min_chunk_size && nr0 >= min_chunk_size) { |
| 4332 | nchunk0 = (nr0 + min_chunk_size - 1) / min_chunk_size; |
| 4333 | } |
| 4334 | |
| 4335 | int64_t dr0 = (nr0 + nchunk0 - 1) / nchunk0; |
| 4336 | // Only increase nchunk0 to nth if it won't make chunks too small |
| 4337 | if (nth == 1 || ((nchunk0 < nth || disable_chunking) && (nr0 + nth - 1) / nth >= min_chunk_size)) { |
| 4338 | nchunk0 = nth; |
| 4339 | dr0 = (nr0 + nchunk0 - 1) / nchunk0; |
| 4340 | } |
| 4341 | |
| 4342 | // Ensure nchunk doesn't exceed the number of rows divided by minimum chunk size |
| 4343 | // This prevents creating too many tiny chunks that could overlap after alignment |
| 4344 | const int64_t max_nchunk = (nr0 + min_chunk_size - 1) / min_chunk_size; |
| 4345 | nchunk0 = MIN(nchunk0, max_nchunk)((nchunk0) < (max_nchunk) ? (nchunk0) : (max_nchunk)); |
| 4346 | |
| 4347 | if (ith == 0) { |
| 4348 | // Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start. |
| 4349 | ggml_threadpool_chunk_set(params->threadpool, nth); |
| 4350 | } |
| 4351 | |
| 4352 | ggml_barrier(params->threadpool); |
| 4353 | |
| 4354 | // The first chunk comes from our thread_id, the rest will get auto-assigned. |
| 4355 | int current_chunk = ith; |
| 4356 | |
| 4357 | while (current_chunk < nchunk0 * nchunk1) { |
| 4358 | const int64_t ith0 = current_chunk % nchunk0; |
| 4359 | const int64_t ith1 = current_chunk / nchunk0; |
| 4360 | |
| 4361 | int64_t src0_start = dr0 * ith0; |
| 4362 | int64_t src0_end = MIN(src0_start + dr0, nr0)((src0_start + dr0) < (nr0) ? (src0_start + dr0) : (nr0)); |
| 4363 | |
| 4364 | // full-plane range for src1 |
| 4365 | int64_t src1_start = ith1 * ne11; |
| 4366 | int64_t src1_end = (ith1 + 1) * ne11; |
| 4367 | |
| 4368 | // Align boundaries to NB_COLS - round up to ensure all data is included |
| 4369 | // The chunk size limiting above ensures chunks are large enough to prevent overlaps |
| 4370 | src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start; |
| 4371 | src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end; |
| 4372 | src0_end = MIN(src0_end, ne01)((src0_end) < (ne01) ? (src0_end) : (ne01)); |
| 4373 | |
| 4374 | // Make sure current plane is the last one before exiting |
| 4375 | if (src0_start >= src0_end) { |
| 4376 | current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1); |
| 4377 | continue; |
| 4378 | } |
| 4379 | |
| 4380 | forward_mul_mat_one_chunk(params, dst, src0_start, src0_end, src1_start, src1_end); |
| 4381 | |
| 4382 | current_chunk = ggml_threadpool_chunk_add(params->threadpool, 1); |
| 4383 | } |
| 4384 | } |
| 4385 | |
| 4386 | void forward_mul_mat_id(ggml_compute_params * params, ggml_tensor * op) { |
| 4387 | const ggml_tensor * src0 = op->src[0]; |
| 4388 | const ggml_tensor * src1 = op->src[1]; |
| 4389 | const ggml_tensor * ids = op->src[2]; |
| 4390 | ggml_tensor * dst = op; |
| 4391 | |
| 4392 | GGML_TENSOR_BINARY_OP_LOCALSconst int64_t ne00 = (src0) ? (src0)->ne[0] : 0; (void)(ne00 ); const int64_t ne01 = (src0) ? (src0)->ne[1] : 0; (void) (ne01); const int64_t ne02 = (src0) ? (src0)->ne[2] : 0; ( void)(ne02); const int64_t ne03 = (src0) ? (src0)->ne[3] : 0; (void)(ne03); const size_t nb00 = (src0) ? (src0)->nb[ 0] : 0; (void)(nb00); const size_t nb01 = (src0) ? (src0)-> nb[1] : 0; (void)(nb01); const size_t nb02 = (src0) ? (src0)-> nb[2] : 0; (void)(nb02); const size_t nb03 = (src0) ? (src0)-> nb[3] : 0; (void)(nb03); const int64_t ne10 = (src1) ? (src1) ->ne[0] : 0; (void)(ne10); const int64_t ne11 = (src1) ? ( src1)->ne[1] : 0; (void)(ne11); const int64_t ne12 = (src1 ) ? (src1)->ne[2] : 0; (void)(ne12); const int64_t ne13 = ( src1) ? (src1)->ne[3] : 0; (void)(ne13); const size_t nb10 = (src1) ? (src1)->nb[0] : 0; (void)(nb10); const size_t nb11 = (src1) ? (src1)->nb[1] : 0; (void)(nb11); const size_t nb12 = (src1) ? (src1)->nb[2] : 0; (void)(nb12); const size_t nb13 = (src1) ? (src1)->nb[3] : 0; (void)(nb13); const int64_t ne0 = (dst) ? (dst)->ne[0] : 0; (void)(ne0); const int64_t ne1 = (dst) ? (dst)->ne[1] : 0; (void)(ne1); const int64_t ne2 = (dst) ? (dst)->ne[2] : 0; (void)(ne2); const int64_t ne3 = (dst) ? (dst)->ne[3] : 0; (void)(ne3); const size_t nb0 = (dst) ? (dst)->nb[0] : 0; (void)(nb0); const size_t nb1 = (dst) ? (dst)->nb[1] : 0; (void)(nb1); const size_t nb2 = (dst) ? (dst)->nb[2] : 0; (void)(nb2); const size_t nb3 = (dst) ? (dst)->nb[3] : 0; (void)(nb3); |
| 4393 | |
| 4394 | const int ith = params->ith; |
| 4395 | const int nth = params->nth; |
| 4396 | |
| 4397 | const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float; |
| 4398 | |
| 4399 | // we don't support permuted src0 or src1 |
| 4400 | GGML_ASSERT(nb00 == ggml_type_size(src0->type))if (!(nb00 == ggml_type_size(src0->type))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4400, "GGML_ASSERT(%s) failed", "nb00 == ggml_type_size(src0->type)" ); |
| 4401 | GGML_ASSERT(nb10 == ggml_type_size(src1->type))if (!(nb10 == ggml_type_size(src1->type))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4401, "GGML_ASSERT(%s) failed", "nb10 == ggml_type_size(src1->type)" ); |
| 4402 | |
| 4403 | // dst cannot be transposed or permuted |
| 4404 | GGML_ASSERT(nb0 == sizeof(float))if (!(nb0 == sizeof(float))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4404, "GGML_ASSERT(%s) failed", "nb0 == sizeof(float)"); |
| 4405 | GGML_ASSERT(nb0 <= nb1)if (!(nb0 <= nb1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4405, "GGML_ASSERT(%s) failed", "nb0 <= nb1"); |
| 4406 | GGML_ASSERT(nb1 <= nb2)if (!(nb1 <= nb2)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4406, "GGML_ASSERT(%s) failed", "nb1 <= nb2"); |
| 4407 | GGML_ASSERT(nb2 <= nb3)if (!(nb2 <= nb3)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4407, "GGML_ASSERT(%s) failed", "nb2 <= nb3"); |
| 4408 | |
| 4409 | GGML_ASSERT(ne03 == 1)if (!(ne03 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4409, "GGML_ASSERT(%s) failed", "ne03 == 1"); |
| 4410 | GGML_ASSERT(ne13 == 1)if (!(ne13 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4410, "GGML_ASSERT(%s) failed", "ne13 == 1"); |
| 4411 | GGML_ASSERT(ne3 == 1)if (!(ne3 == 1)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4411, "GGML_ASSERT(%s) failed", "ne3 == 1"); |
| 4412 | |
| 4413 | GGML_ASSERT(src1->type == GGML_TYPE_F32)if (!(src1->type == GGML_TYPE_F32)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4413, "GGML_ASSERT(%s) failed", "src1->type == GGML_TYPE_F32" ); |
| 4414 | |
| 4415 | // row groups |
| 4416 | const int n_ids = ids->ne[0]; // n_expert_used |
| 4417 | const int n_as = ne02; // n_expert |
| 4418 | |
| 4419 | const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10); |
| 4420 | const size_t nbw2 = nbw1*ne11; |
| 4421 | const size_t nbw3 = nbw2*ne12; |
| 4422 | |
| 4423 | struct mmid_row_mapping { |
| 4424 | int32_t i1; |
| 4425 | int32_t i2; |
| 4426 | }; |
| 4427 | |
| 4428 | GGML_ASSERT(params->wsize >=if (!(params->wsize >= ((((nbw3) + (sizeof(int64_t)) - 1 ) & ~((sizeof(int64_t)) - 1)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping )))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4431, "GGML_ASSERT(%s) failed", "params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping))" ) |
| 4429 | (GGML_PAD(nbw3, sizeof(int64_t)) +if (!(params->wsize >= ((((nbw3) + (sizeof(int64_t)) - 1 ) & ~((sizeof(int64_t)) - 1)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping )))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4431, "GGML_ASSERT(%s) failed", "params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping))" ) |
| 4430 | n_as*(ne12 + 1)*sizeof(mmid_row_mapping))if (!(params->wsize >= ((((nbw3) + (sizeof(int64_t)) - 1 ) & ~((sizeof(int64_t)) - 1)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping )))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4431, "GGML_ASSERT(%s) failed", "params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping))" ) |
| 4431 | )if (!(params->wsize >= ((((nbw3) + (sizeof(int64_t)) - 1 ) & ~((sizeof(int64_t)) - 1)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping )))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4431, "GGML_ASSERT(%s) failed", "params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as*(ne12 + 1)*sizeof(mmid_row_mapping))" ); |
| 4432 | |
| 4433 | auto * wdata = (char *)params->wdata; |
| 4434 | auto * wdata_src1_end = (char *)wdata + GGML_PAD(nbw3, sizeof(int64_t))(((nbw3) + (sizeof(int64_t)) - 1) & ~((sizeof(int64_t)) - 1)); |
| 4435 | |
| 4436 | // total of [n_as][ne12 + 1] elements of type mmid_row_mapping (2*int32_t = int64_t) |
| 4437 | auto * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] |
| 4438 | struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] |
| 4439 | |
| 4440 | // src1: float32 => param type |
| 4441 | for (int64_t i12 = 0; i12 < ne12; ++i12) { |
| 4442 | for (int64_t i11 = ith; i11 < ne11; i11 += nth) { |
| 4443 | from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11), |
| 4444 | (void *) (wdata + i12 * nbw2 + i11 * nbw1), |
| 4445 | ne10); |
| 4446 | } |
| 4447 | } |
| 4448 | |
| 4449 | #define MMID_MATRIX_ROW(row_id, i1) matrix_rows[(row_id) * ne12 + (i1)] |
| 4450 | |
| 4451 | if (ith == 0) { |
| 4452 | // initialize matrix_row_counts |
| 4453 | memset(matrix_row_counts, 0, n_as * sizeof(int64_t)); |
| 4454 | |
| 4455 | // group rows by src0 matrix |
| 4456 | for (int32_t iid1 = 0; iid1 < ids->ne[1]; ++iid1) { |
| 4457 | for (int32_t id = 0; id < n_ids; ++id) { |
| 4458 | const int32_t i02 = |
| 4459 | *(const int32_t *) ((const char *) ids->data + iid1 * ids->nb[1] + id * ids->nb[0]); |
| 4460 | |
| 4461 | GGML_ASSERT(i02 >= 0 && i02 < n_as)if (!(i02 >= 0 && i02 < n_as)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4461, "GGML_ASSERT(%s) failed", "i02 >= 0 && i02 < n_as" ); |
| 4462 | |
| 4463 | MMID_MATRIX_ROW(i02, matrix_row_counts[i02]) = { id, iid1 }; |
| 4464 | matrix_row_counts[i02] += 1; |
| 4465 | } |
| 4466 | } |
| 4467 | } |
| 4468 | |
| 4469 | ggml_barrier(params->threadpool); |
| 4470 | |
| 4471 | // compute each matrix multiplication in sequence |
| 4472 | for (int cur_a = 0; cur_a < n_as; ++cur_a) { |
| 4473 | const int64_t cne1 = matrix_row_counts[cur_a]; |
| 4474 | |
| 4475 | if (cne1 == 0) { |
| 4476 | continue; |
| 4477 | } |
| 4478 | |
| 4479 | const auto * src0_cur = (const char *) src0->data + cur_a*nb02; |
| 4480 | |
| 4481 | //const int64_t nr0 = ne01; // src0 rows |
| 4482 | const int64_t nr1 = cne1; // src1 rows |
| 4483 | |
| 4484 | int64_t src0_cur_start = (ith * ne01) / nth; |
| 4485 | int64_t src0_cur_end = ((ith + 1) * ne01) / nth; |
| 4486 | |
| 4487 | // Align boundaries to NB_COLS - round up to ensure all data is included |
| 4488 | src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start; |
| 4489 | src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end; |
| 4490 | if (src0_cur_end > ne01) { |
| 4491 | src0_cur_end = ne01; |
| 4492 | } |
| 4493 | |
| 4494 | if (src0_cur_start >= src0_cur_end) { |
| 4495 | return; |
| 4496 | } |
| 4497 | |
| 4498 | for (int ir1 = 0; ir1 < nr1; ir1++) { |
| 4499 | struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); |
| 4500 | |
| 4501 | const int id = row_mapping.i1; // selected expert index |
| 4502 | |
| 4503 | const int64_t i11 = id % ne11; |
| 4504 | const int64_t i12 = row_mapping.i2; // row index in src1 |
| 4505 | |
| 4506 | const int64_t i1 = id; // selected expert index |
| 4507 | const int64_t i2 = i12; // row |
| 4508 | |
| 4509 | const auto * src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2); |
| 4510 | |
| 4511 | gemv<BLOC_TYPE, INTER_SIZE, NB_COLS, PARAM_TYPE>( |
| 4512 | ne00, (float *) ((char *) dst->data + (i1 * nb1 + i2 * nb2)) + src0_cur_start, ne01, |
| 4513 | src0_cur + src0_cur_start * nb01, src1_col, 1, src0_cur_end - src0_cur_start); |
| 4514 | } |
| 4515 | } |
| 4516 | #undef MMID_MATRIX_ROW |
| 4517 | } |
| 4518 | |
| 4519 | int repack(struct ggml_tensor * t, const void * data, size_t data_size) override { |
| 4520 | GGML_LOG_DEBUG("%s: repack tensor %s with %s_%dx%d\n", __func__, t->name, ggml_type_name(t->type),ggml_log_internal(GGML_LOG_LEVEL_DEBUG, "%s: repack tensor %s with %s_%dx%d\n" , __func__, t->name, ggml_type_name(t->type), (int) NB_COLS , (int) INTER_SIZE) |
| 4521 | (int) NB_COLS, (int) INTER_SIZE)ggml_log_internal(GGML_LOG_LEVEL_DEBUG, "%s: repack tensor %s with %s_%dx%d\n" , __func__, t->name, ggml_type_name(t->type), (int) NB_COLS , (int) INTER_SIZE); |
| 4522 | return ggml::cpu::repack::repack<BLOC_TYPE, INTER_SIZE, NB_COLS>(t, data, data_size); |
| 4523 | } |
| 4524 | }; |
| 4525 | |
| 4526 | } // namespace ggml::cpu::repack |
| 4527 | |
| 4528 | static const ggml::cpu::tensor_traits * ggml_repack_get_optimal_repack_type(const struct ggml_tensor * cur) { |
| 4529 | // instance for Q4 |
| 4530 | static const ggml::cpu::repack::tensor_traits<block_q4_0, 4, 4, GGML_TYPE_Q8_0> q4_0_4x4_q8_0; |
| 4531 | static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 4, GGML_TYPE_Q8_0> q4_0_4x8_q8_0; |
| 4532 | static const ggml::cpu::repack::tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0; |
| 4533 | |
| 4534 | // instance for Q4_K |
| 4535 | static const ggml::cpu::repack::tensor_traits<block_q4_K, 4, 8, GGML_TYPE_Q8_K> q4_K_8x4_q8_K; |
| 4536 | static const ggml::cpu::repack::tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K; |
| 4537 | |
| 4538 | // instance for Q5_K |
| 4539 | static const ggml::cpu::repack::tensor_traits<block_q5_K, 4, 8, GGML_TYPE_Q8_K> q5_K_8x4_q8_K; |
| 4540 | static const ggml::cpu::repack::tensor_traits<block_q5_K, 8, 8, GGML_TYPE_Q8_K> q5_K_8x8_q8_K; |
| 4541 | |
| 4542 | // instance for Q6_K |
| 4543 | static const ggml::cpu::repack::tensor_traits<block_q6_K, 4, 8, GGML_TYPE_Q8_K> q6_K_8x4_q8_K; |
| 4544 | static const ggml::cpu::repack::tensor_traits<block_q6_K, 8, 8, GGML_TYPE_Q8_K> q6_K_8x8_q8_K; |
| 4545 | |
| 4546 | // instance for Q2 |
| 4547 | static const ggml::cpu::repack::tensor_traits<block_q2_K, 8, 8, GGML_TYPE_Q8_K> q2_K_8x8_q8_K; |
| 4548 | |
| 4549 | // instance for IQ4 |
| 4550 | static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0; |
| 4551 | static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 8, 8, GGML_TYPE_Q8_0> iq4_nl_8x8_q8_0; |
| 4552 | |
| 4553 | // instance for MXFP4 |
| 4554 | static const ggml::cpu::repack::tensor_traits<block_mxfp4, 4, 4, GGML_TYPE_Q8_0> mxfp4_4x4_q8_0; |
| 4555 | static const ggml::cpu::repack::tensor_traits<block_mxfp4, 8, 8, GGML_TYPE_Q8_0> mxfp4_8x8_q8_0; |
| 4556 | |
| 4557 | // instance for Q8_0 |
| 4558 | static const ggml::cpu::repack::tensor_traits<block_q8_0, 4, 4, GGML_TYPE_Q8_0> q8_0_4x4_q8_0; |
| 4559 | static const ggml::cpu::repack::tensor_traits<block_q8_0, 8, 4, GGML_TYPE_Q8_0> q8_0_4x8_q8_0; |
| 4560 | |
| 4561 | // instances for RISC-V |
| 4562 | // |
| 4563 | // These implement outer-product style matrix multiplication kernels with |
| 4564 | // an interleave of 1. |
| 4565 | #if defined __riscv_zvfh |
| 4566 | static const ggml::cpu::repack::tensor_traits<block_q4_0, 1, 16, GGML_TYPE_Q8_0> q4_0_16x1_q8_0; |
| 4567 | static const ggml::cpu::repack::tensor_traits<block_q4_K, 1, 16, GGML_TYPE_Q8_K> q4_K_16x1_q8_K; |
| 4568 | static const ggml::cpu::repack::tensor_traits<block_iq4_nl, 1, 16, GGML_TYPE_Q8_0> iq4_nl_16x1_q8_0; |
| 4569 | static const ggml::cpu::repack::tensor_traits<block_q8_0, 1, 16, GGML_TYPE_Q8_0> q8_0_16x1_q8_0; |
| 4570 | static const ggml::cpu::repack::tensor_traits<block_q2_K, 1, 16, GGML_TYPE_Q8_K> q2_K_16x1_q8_K; |
| 4571 | #endif |
| 4572 | |
| 4573 | if (cur->type == GGML_TYPE_Q4_0) { |
| 4574 | if (ggml_cpu_has_avx2() || (ggml_cpu_has_sve() && ggml_cpu_has_matmul_int8() && ggml_cpu_get_sve_cnt() == QK8_032)) { |
| 4575 | if (cur->ne[1] % 8 == 0) { |
| 4576 | return &q4_0_8x8_q8_0; |
| 4577 | } |
| 4578 | } |
| 4579 | if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { |
| 4580 | if (cur->ne[1] % 4 == 0) { |
| 4581 | return &q4_0_4x8_q8_0; |
| 4582 | } |
| 4583 | } |
| 4584 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4585 | if (cur->ne[1] % 4 == 0) { |
| 4586 | return &q4_0_4x4_q8_0; |
| 4587 | } |
| 4588 | } |
| 4589 | if (ggml_cpu_has_riscv_v()) { |
| 4590 | #if defined __riscv_zvfh |
| 4591 | switch (__riscv_vlenb() * 8) { |
| 4592 | case 128: { break; } // TODO |
| 4593 | case 256: { if (cur->ne[1] % 16 == 0) { return &q4_0_16x1_q8_0; } break; } |
| 4594 | case 512: { break; } // TODO |
| 4595 | case 1024: { break; } // TODO |
| 4596 | default: { return nullptr; } |
| 4597 | } |
| 4598 | #endif |
| 4599 | } |
| 4600 | } else if (cur->type == GGML_TYPE_Q4_K) { |
| 4601 | if (ggml_cpu_has_avx2()) { |
| 4602 | if (cur->ne[1] % 8 == 0) { |
| 4603 | return &q4_K_8x8_q8_K; |
| 4604 | } |
| 4605 | } |
| 4606 | if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { |
| 4607 | if (cur->ne[1] % 8 == 0) { |
| 4608 | return &q4_K_8x8_q8_K; |
| 4609 | } |
| 4610 | } |
| 4611 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4612 | if (cur->ne[1] % 8 == 0) { |
| 4613 | return &q4_K_8x4_q8_K; |
| 4614 | } |
| 4615 | } |
| 4616 | if (ggml_cpu_has_riscv_v()) { |
| 4617 | #if defined __riscv_zvfh |
| 4618 | switch (__riscv_vlenb() * 8) { |
| 4619 | case 128: { break; } // TODO |
| 4620 | case 256: { if (cur->ne[1] % 16 == 0) { return &q4_K_16x1_q8_K; } break; } |
| 4621 | case 512: { break; } // TODO |
| 4622 | case 1024: { break; } // TODO |
| 4623 | default: { return nullptr; } |
| 4624 | } |
| 4625 | #endif |
| 4626 | } |
| 4627 | } else if (cur->type == GGML_TYPE_Q2_K) { |
| 4628 | if (ggml_cpu_has_avx512()) { |
| 4629 | if (cur->ne[1] % 8 == 0) { |
| 4630 | return &q2_K_8x8_q8_K; |
| 4631 | } |
| 4632 | } |
| 4633 | if (ggml_cpu_has_riscv_v()) { |
| 4634 | #if defined __riscv_zvfh |
| 4635 | switch (__riscv_vlenb() * 8) { |
| 4636 | case 128: { break; } // TODO |
| 4637 | case 256: { if (cur->ne[1] % 16 == 0) { return &q2_K_16x1_q8_K; } break; } |
| 4638 | case 512: { break; } // TODO |
| 4639 | case 1024: { break; } // TODO |
| 4640 | default: { return nullptr; } |
| 4641 | } |
| 4642 | #endif |
| 4643 | } |
| 4644 | } else if (cur->type == GGML_TYPE_Q5_K) { |
| 4645 | if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { |
| 4646 | if (cur->ne[1] % 8 == 0) { |
| 4647 | return &q5_K_8x8_q8_K; |
| 4648 | } |
| 4649 | } |
| 4650 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4651 | if (cur->ne[1] % 8 == 0) { |
| 4652 | return &q5_K_8x4_q8_K; |
| 4653 | } |
| 4654 | } |
| 4655 | } else if (cur->type == GGML_TYPE_Q6_K) { |
| 4656 | if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { |
| 4657 | if (cur->ne[1] % 8 == 0) { |
| 4658 | return &q6_K_8x8_q8_K; |
| 4659 | } |
| 4660 | } |
| 4661 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4662 | if (cur->ne[1] % 8 == 0) { |
| 4663 | return &q6_K_8x4_q8_K; |
| 4664 | } |
| 4665 | } |
| 4666 | } else if (cur->type == GGML_TYPE_IQ4_NL) { |
| 4667 | if (ggml_cpu_has_avx2()) { |
| 4668 | if (cur->ne[1] % 8 == 0) { |
| 4669 | return &iq4_nl_8x8_q8_0; |
| 4670 | } |
| 4671 | } |
| 4672 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4673 | if (cur->ne[1] % 4 == 0) { |
| 4674 | return &iq4_nl_4x4_q8_0; |
| 4675 | } |
| 4676 | } |
| 4677 | if (ggml_cpu_has_riscv_v()) { |
| 4678 | #if defined __riscv_zvfh |
| 4679 | switch (__riscv_vlenb() * 8) { |
| 4680 | case 128: { break; } // TODO |
| 4681 | case 256: { if (cur->ne[1] % 16 == 0) { return &iq4_nl_16x1_q8_0; } break; } |
| 4682 | case 512: { break; } // TODO |
| 4683 | case 1024: { break; } // TODO |
| 4684 | default: { return nullptr; } |
| 4685 | } |
| 4686 | #endif |
| 4687 | } |
| 4688 | } else if (cur->type == GGML_TYPE_MXFP4) { |
| 4689 | if (ggml_cpu_has_avx2()) { |
| 4690 | if (cur->ne[1] % 8 == 0) { |
| 4691 | return &mxfp4_8x8_q8_0; |
| 4692 | } |
| 4693 | } |
| 4694 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4695 | if (cur->ne[1] % 4 == 0) { |
| 4696 | return &mxfp4_4x4_q8_0; |
| 4697 | } |
| 4698 | } |
| 4699 | } else if (cur->type == GGML_TYPE_Q8_0) { |
| 4700 | if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) { |
| 4701 | if (cur->ne[1] % 4 == 0) { |
| 4702 | return &q8_0_4x8_q8_0; |
| 4703 | } |
| 4704 | } |
| 4705 | if (ggml_cpu_has_neon() && ggml_cpu_has_dotprod()) { |
| 4706 | if (cur->ne[1] % 4 == 0) { |
| 4707 | return &q8_0_4x4_q8_0; |
| 4708 | } |
| 4709 | } |
| 4710 | if (ggml_cpu_has_riscv_v()) { |
| 4711 | #if defined __riscv_zvfh |
| 4712 | switch (__riscv_vlenb() * 8) { |
| 4713 | case 128: { break; } // TODO |
| 4714 | case 256: { if (cur->ne[1] % 16 == 0) { return &q8_0_16x1_q8_0; } break; } |
| 4715 | case 512: { break; } // TODO |
| 4716 | case 1024: { break; } // TODO |
| 4717 | default: { return nullptr; } |
| 4718 | } |
| 4719 | #endif |
| 4720 | } |
| 4721 | } |
| 4722 | |
| 4723 | return nullptr; |
| 4724 | } |
| 4725 | |
| 4726 | static enum ggml_status ggml_backend_cpu_repack_buffer_init_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) { |
| 4727 | tensor->extra = (void *) const_cast<ggml::cpu::tensor_traits *>(ggml_repack_get_optimal_repack_type(tensor)); |
| 4728 | |
| 4729 | GGML_UNUSED(buffer)(void)(buffer); |
| 4730 | return GGML_STATUS_SUCCESS; |
| 4731 | } |
| 4732 | |
| 4733 | static void ggml_backend_cpu_repack_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, |
| 4734 | const void * data, size_t offset, size_t size) { |
| 4735 | GGML_ASSERT(offset == 0)if (!(offset == 0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4735, "GGML_ASSERT(%s) failed", "offset == 0"); |
| 4736 | GGML_ASSERT(size == ggml_nbytes(tensor))if (!(size == ggml_nbytes(tensor))) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4736, "GGML_ASSERT(%s) failed", "size == ggml_nbytes(tensor)" ); |
| 4737 | |
| 4738 | auto tensor_traits = (ggml::cpu::repack::tensor_traits_base *) tensor->extra; |
| 4739 | auto OK = tensor_traits->repack(tensor, data, size); |
| 4740 | |
| 4741 | GGML_ASSERT(OK == 0)if (!(OK == 0)) ggml_abort("/root/firefox-clang/third_party/llama.cpp/ggml/src/ggml-cpu/repack.cpp" , 4741, "GGML_ASSERT(%s) failed", "OK == 0"); |
| 4742 | GGML_UNUSED(buffer)(void)(buffer); |
| 4743 | } |
| 4744 | |
| 4745 | static const char * ggml_backend_cpu_repack_buffer_type_get_name(ggml_backend_buffer_type_t buft) { |
| 4746 | return "CPU_REPACK"; |
| 4747 | |
| 4748 | GGML_UNUSED(buft)(void)(buft); |
| 4749 | } |
| 4750 | |
| 4751 | static ggml_backend_buffer_t ggml_backend_cpu_repack_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) { |
| 4752 | ggml_backend_buffer_t buffer = ggml_backend_buft_alloc_buffer(ggml_backend_cpu_buffer_type(), size); |
| 4753 | |
| 4754 | if (buffer == nullptr) { |
| 4755 | return nullptr; |
| 4756 | } |
| 4757 | |
| 4758 | buffer->buft = buft; |
| 4759 | buffer->iface.init_tensor = ggml_backend_cpu_repack_buffer_init_tensor; |
| 4760 | buffer->iface.set_tensor = ggml_backend_cpu_repack_buffer_set_tensor; |
| 4761 | buffer->iface.get_tensor = nullptr; |
| 4762 | buffer->iface.cpy_tensor = nullptr; |
| 4763 | return buffer; |
| 4764 | } |
| 4765 | |
| 4766 | static size_t ggml_backend_cpu_repack_buffer_type_get_alignment(ggml_backend_buffer_type_t buft) { |
| 4767 | return TENSOR_ALIGNMENT32; |
| 4768 | |
| 4769 | GGML_UNUSED(buft)(void)(buft); |
| 4770 | } |
| 4771 | |
| 4772 | namespace ggml::cpu::repack { |
| 4773 | class extra_buffer_type : ggml::cpu::extra_buffer_type { |
| 4774 | bool supports_op(ggml_backend_dev_t, const struct ggml_tensor * op) override { |
| 4775 | if ( op->op == GGML_OP_MUL_MAT && |
| 4776 | op->src[0]->buffer && |
| 4777 | (ggml_n_dims(op->src[0]) == 2) && |
| 4778 | op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type() && |
| 4779 | ggml_repack_get_optimal_repack_type(op->src[0]) |
| 4780 | ) { |
| 4781 | if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { |
| 4782 | return false; |
| 4783 | } |
| 4784 | if (op->src[1]->type == GGML_TYPE_F32) { |
| 4785 | return true; |
| 4786 | } |
| 4787 | //if (op->src[1]->type == GGML_TYPE_Q8_0) { |
| 4788 | // return true; |
| 4789 | //} |
| 4790 | // may be possible if Q8_0 packed... |
| 4791 | } else if (op->op == GGML_OP_MUL_MAT_ID |
| 4792 | && op->src[0]->buffer |
| 4793 | && (ggml_n_dims(op->src[0]) == 3) |
| 4794 | && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type() |
| 4795 | && ggml_repack_get_optimal_repack_type(op->src[0]) |
| 4796 | ) { |
| 4797 | if (op->src[1]->buffer && !ggml_backend_buft_is_host(op->src[1]->buffer->buft)) { |
| 4798 | return false; |
| 4799 | } |
| 4800 | if (op->src[1]->type == GGML_TYPE_F32) { |
| 4801 | return true; |
| 4802 | } |
| 4803 | //if (op->src[1]->type == GGML_TYPE_Q8_0) { |
| 4804 | // return true; |
| 4805 | //} |
| 4806 | } |
| 4807 | return false; |
| 4808 | } |
| 4809 | |
| 4810 | ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override { |
| 4811 | if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_MUL_MAT_ID) { |
| 4812 | if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_repack_buffer_type()) { |
| 4813 | return (ggml::cpu::tensor_traits *) op->src[0]->extra; |
| 4814 | } |
| 4815 | } |
| 4816 | return nullptr; |
| 4817 | } |
| 4818 | }; |
| 4819 | } // namespace ggml::cpu::repack |
| 4820 | |
| 4821 | ggml_backend_buffer_type_t ggml_backend_cpu_repack_buffer_type(void) { |
| 4822 | static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_repack = { |
| 4823 | /* .iface = */ { |
| 4824 | /* .get_name = */ ggml_backend_cpu_repack_buffer_type_get_name, |
| 4825 | /* .alloc_buffer = */ ggml_backend_cpu_repack_buffer_type_alloc_buffer, |
| 4826 | /* .get_alignment = */ ggml_backend_cpu_repack_buffer_type_get_alignment, |
| 4827 | /* .get_max_size = */ nullptr, // defaults to SIZE_MAX |
| 4828 | /* .get_alloc_size = */ nullptr, // defaults to ggml_nbytes |
| 4829 | /* .is_host = */ nullptr, |
| 4830 | }, |
| 4831 | /* .device = */ ggml_backend_reg_dev_get(ggml_backend_cpu_reg(), 0), |
| 4832 | /* .context = */ new ggml::cpu::repack::extra_buffer_type(), |
| 4833 | }; |
| 4834 | |
| 4835 | return &ggml_backend_cpu_buffer_type_repack; |
| 4836 | } |