Spaces:
Sleeping
Sleeping
File size: 93,272 Bytes
613af8d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642 1643 1644 1645 1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695 1696 1697 1698 1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715 1716 1717 1718 1719 1720 1721 1722 1723 1724 1725 1726 1727 1728 1729 1730 1731 1732 1733 1734 1735 1736 1737 1738 1739 1740 1741 1742 1743 1744 1745 1746 1747 1748 1749 1750 1751 1752 1753 1754 1755 1756 1757 1758 1759 1760 1761 1762 1763 1764 1765 1766 1767 1768 1769 1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780 1781 1782 1783 1784 1785 1786 1787 1788 1789 1790 1791 1792 1793 1794 1795 1796 1797 1798 1799 1800 1801 1802 1803 1804 1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822 1823 1824 1825 1826 1827 1828 1829 1830 1831 1832 1833 1834 1835 1836 1837 1838 1839 1840 1841 1842 1843 1844 1845 1846 1847 1848 1849 1850 1851 1852 1853 1854 1855 1856 1857 1858 1859 1860 1861 1862 1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878 1879 1880 1881 1882 1883 1884 1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898 1899 1900 1901 1902 1903 1904 1905 1906 1907 1908 1909 1910 1911 1912 1913 1914 1915 1916 1917 1918 1919 1920 1921 1922 1923 1924 1925 1926 1927 1928 1929 1930 1931 1932 1933 1934 1935 1936 1937 1938 1939 1940 1941 1942 1943 1944 1945 1946 1947 1948 1949 1950 1951 1952 1953 1954 1955 1956 1957 1958 1959 1960 1961 1962 1963 1964 1965 1966 1967 1968 1969 1970 1971 1972 1973 1974 1975 1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024 2025 2026 2027 2028 2029 2030 2031 2032 2033 2034 2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048 2049 2050 2051 2052 2053 2054 2055 2056 2057 2058 2059 2060 2061 2062 2063 2064 2065 2066 2067 2068 2069 2070 2071 2072 2073 2074 2075 2076 2077 2078 2079 2080 2081 2082 2083 2084 2085 2086 2087 2088 2089 2090 2091 2092 2093 2094 2095 2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106 2107 2108 2109 2110 2111 2112 2113 2114 2115 2116 2117 2118 2119 2120 2121 2122 2123 2124 2125 2126 2127 2128 2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155 2156 2157 2158 2159 2160 2161 2162 2163 2164 2165 2166 2167 2168 2169 2170 2171 2172 2173 2174 2175 2176 2177 2178 2179 2180 2181 |
// SPDX-FileCopyrightText: Copyright 2024 Arm Ltd.
#define GGML_COMMON_IMPL_C
#include "ggml-common.h"
#include "ggml-quants.h"
#include "ggml-impl.h"
#include <math.h>
#include <string.h>
#include <assert.h>
#include <float.h>
#include <stdlib.h> // for qsort
#include <stdio.h> // for GGML_ASSERT
#include "ggml-aarch64.h"
#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Woverlength-strings"
#elif defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
#endif
#define UNUSED GGML_UNUSED
// Functions to create the interleaved data layout formats
// interleave 4 block_q4_0s in blocks of blck_size_interleave
// returns an interleaved block_q4_0x4
// in the interleaved block_q4_0x4, place deltas for 4 block_q4_0 blocks
// first, then interleave quants from 4 block_q4_0s in blocks of blck_size_interleave
//
// - in : an array of block_q4_0 pointers
// - blck_size_interleave : the block_q4_0 quants bytes are interleaved in blocks of
// blck_size_interleave bytes
// - xor_mask : the mask to convert the nibbles in block_q4_0 quants bytes
// from bias offset form to pure sign form (this saves subtract
// operations durin unpacking)
//
static block_q4_0x4 make_block_q4_0x4(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
block_q4_0x4 out;
for (int i = 0; i < 4; i++) {
out.d[i] = in[i].d;
}
for (int i = 0; i < QK4_0 * 2; i++) {
int src_offset = (i / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % blck_size_interleave);
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
}
return out;
}
// interleave 8 block_q4_0s in blocks of blck_size_interleave
// returns an interleaved block_q4_0x8
// in the interleaved block_q4_0x8, place deltas for 8 block_q4_0 blocks
// first, then interleave quants from 8 block_q4_0s in blocks of blck_size_interleave
static block_q4_0x8 make_block_q4_0x8(block_q4_0 * in, unsigned int blck_size_interleave, unsigned int xor_mask) {
block_q4_0x8 out;
for (int i = 0; i < 8; i++) {
out.d[i] = in[i].d;
}
for (int i = 0; i < QK4_0 * 4; i++) {
int src_offset = (i / (8 * blck_size_interleave)) * blck_size_interleave;
int src_id = (i % (8 * blck_size_interleave)) / blck_size_interleave;
src_offset += (i % blck_size_interleave);
out.qs[i] = in[src_id].qs[src_offset] ^ xor_mask;
}
return out;
}
void quantize_q8_0_4x4(const float * restrict x, void * restrict vy, int64_t k) {
assert(QK8_0 == 32);
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
#if defined(__ARM_NEON)
float32x4_t srcv[4][8];
float id[4];
for (int i = 0; i < nb; i++) {
float32x4_t asrcv[8];
float32x4_t amaxv[8];
for (int row_iter = 0; row_iter < 4; row_iter++) {
for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
const float amax = vmaxvq_f32(amaxv[0]);
const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
}
for (int j = 0; j < 8; j++) {
float32x4_t v = vmulq_n_f32(srcv[0][j], id[0]);
int32x4_t vi = vcvtnq_s32_f32(v);
y[i].qs[16 * j + 0] = vgetq_lane_s32(vi, 0);
y[i].qs[16 * j + 1] = vgetq_lane_s32(vi, 1);
y[i].qs[16 * j + 2] = vgetq_lane_s32(vi, 2);
y[i].qs[16 * j + 3] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[1][j], id[1]);
vi = vcvtnq_s32_f32(v);
y[i].qs[16 * j + 4] = vgetq_lane_s32(vi, 0);
y[i].qs[16 * j + 5] = vgetq_lane_s32(vi, 1);
y[i].qs[16 * j + 6] = vgetq_lane_s32(vi, 2);
y[i].qs[16 * j + 7] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[2][j], id[2]);
vi = vcvtnq_s32_f32(v);
y[i].qs[16 * j + 8] = vgetq_lane_s32(vi, 0);
y[i].qs[16 * j + 9] = vgetq_lane_s32(vi, 1);
y[i].qs[16 * j + 10] = vgetq_lane_s32(vi, 2);
y[i].qs[16 * j + 11] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[3][j], id[3]);
vi = vcvtnq_s32_f32(v);
y[i].qs[16 * j + 12] = vgetq_lane_s32(vi, 0);
y[i].qs[16 * j + 13] = vgetq_lane_s32(vi, 1);
y[i].qs[16 * j + 14] = vgetq_lane_s32(vi, 2);
y[i].qs[16 * j + 15] = vgetq_lane_s32(vi, 3);
}
}
#else
// scalar
const int blck_size_interleave = 4;
float srcv[4][QK8_0];
float id[4];
for (int i = 0; i < nb; i++) {
for (int row_iter = 0; row_iter < 4; row_iter++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
amax = MAX(amax, fabsf(srcv[row_iter][j]));
}
const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
}
for (int j = 0; j < QK8_0 * 4; j++) {
int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % blck_size_interleave);
float x0 = srcv[src_id][src_offset] * id[src_id];
y[i].qs[j] = roundf(x0);
}
}
#endif
}
void quantize_q8_0_4x8(const float * restrict x, void * restrict vy, int64_t k) {
assert(QK8_0 == 32);
assert(k % QK8_0 == 0);
const int nb = k / QK8_0;
block_q8_0x4 * restrict y = (block_q8_0x4 *) vy;
#if defined(__ARM_NEON)
float32x4_t srcv[4][8];
float id[4];
for (int i = 0; i < nb; i++) {
float32x4_t asrcv[8];
float32x4_t amaxv[8];
for (int row_iter = 0; row_iter < 4; row_iter++) {
for (int j = 0; j < 8; j++) srcv[row_iter][j] = vld1q_f32(x + row_iter * k + i * 32 + 4 * j);
for (int j = 0; j < 8; j++) asrcv[j] = vabsq_f32(srcv[row_iter][j]);
for (int j = 0; j < 4; j++) amaxv[2 * j] = vmaxq_f32(asrcv[2 * j], asrcv[2 * j + 1]);
for (int j = 0; j < 2; j++) amaxv[4 * j] = vmaxq_f32(amaxv[4 * j], amaxv[4 * j + 2]);
for (int j = 0; j < 1; j++) amaxv[8 * j] = vmaxq_f32(amaxv[8 * j], amaxv[8 * j + 4]);
const float amax = vmaxvq_f32(amaxv[0]);
const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
}
for (int j = 0; j < 4; j++) {
float32x4_t v = vmulq_n_f32(srcv[0][2 * j], id[0]);
int32x4_t vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 0] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 1] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 2] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 3] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[0][2 * j + 1], id[0]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 4] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 5] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 6] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 7] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[1][2 * j], id[1]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 8] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 9] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 10] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 11] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[1][2 * j + 1], id[1]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 12] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 13] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 14] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 15] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[2][2 * j], id[2]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 16] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 17] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 18] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 19] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[2][2 * j + 1], id[2]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 20] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 21] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 22] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 23] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[3][2 * j], id[3]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 24] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 25] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 26] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 27] = vgetq_lane_s32(vi, 3);
v = vmulq_n_f32(srcv[3][2 * j + 1], id[3]);
vi = vcvtnq_s32_f32(v);
y[i].qs[32 * j + 28] = vgetq_lane_s32(vi, 0);
y[i].qs[32 * j + 29] = vgetq_lane_s32(vi, 1);
y[i].qs[32 * j + 30] = vgetq_lane_s32(vi, 2);
y[i].qs[32 * j + 31] = vgetq_lane_s32(vi, 3);
}
}
#else
// scalar
const int blck_size_interleave = 8;
float srcv[4][QK8_0];
float id[4];
for (int i = 0; i < nb; i++) {
for (int row_iter = 0; row_iter < 4; row_iter++) {
float amax = 0.0f; // absolute max
for (int j = 0; j < QK8_0; j++) {
srcv[row_iter][j] = x[row_iter * k + i * QK8_0 + j];
amax = MAX(amax, fabsf(srcv[row_iter][j]));
}
const float d = amax / ((1 << 7) - 1);
id[row_iter] = d ? 1.0f / d : 0.0f;
y[i].d[row_iter] = GGML_FP32_TO_FP16(d);
}
for (int j = 0; j < QK8_0 * 4; j++) {
int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % blck_size_interleave);
float x0 = srcv[src_id][src_offset] * id[src_id];
y[i].qs[j] = roundf(x0);
}
}
#endif
}
void quantize_mat_q8_0(const float * restrict x, void * restrict vy, int64_t nrow, int64_t n_per_row, int64_t blck_size_interleave) {
assert(nrow == 4);
UNUSED(nrow);
if (blck_size_interleave == 4) {
quantize_q8_0_4x4(x, vy, n_per_row);
} else if (blck_size_interleave == 8) {
quantize_q8_0_4x8(x, vy, n_per_row);
} else {
assert(false);
}
}
static size_t quantize_q4_0_nr_bl(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, int nrows_interleaved, int blck_size_interleave) {
assert(n_per_row % QK4_0 == 0);
const int nb = n_per_row / QK4_0;
void * out_ptr = NULL;
if (nrows_interleaved == 8) {
out_ptr = (block_q4_0x8 *) dst;
}
else if (nrows_interleaved == 4) {
out_ptr = (block_q4_0x4 *) dst;
}
assert(nrows_interleaved <= 8);
block_q4_0 dst_tmp[8];
for (int b = 0; b < (nrow * n_per_row); b += nrows_interleaved * n_per_row) {
for (int64_t x = 0; x < nb; x++) {
for (int i = 0; i < nrows_interleaved; i++ ) {
quantize_row_q4_0_ref(src + b + i * n_per_row + x * QK4_0, (block_q4_0 *) dst_tmp + i, QK4_0);
}
if (nrows_interleaved == 8) {
*(block_q4_0x8 *) out_ptr = make_block_q4_0x8(dst_tmp, blck_size_interleave, 0x88);
out_ptr = (block_q4_0x8 *) out_ptr + 1;
}
else if (nrows_interleaved == 4) {
*(block_q4_0x4 *) out_ptr = make_block_q4_0x4(dst_tmp, blck_size_interleave, 0x88);
out_ptr = (block_q4_0x4 *) out_ptr + 1;
}
}
}
return ((nrow * n_per_row) / QK4_0 * sizeof(block_q4_0));
}
size_t quantize_q4_0_4x4(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
UNUSED(quant_weights);
return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 4);
}
size_t quantize_q4_0_4x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
UNUSED(quant_weights);
return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 4, 8);
}
size_t quantize_q4_0_8x8(const float * restrict src, void * restrict dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
UNUSED(quant_weights);
return quantize_q4_0_nr_bl(src, dst, nrow, n_per_row, 8, 8);
}
void ggml_gemv_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
const int blocklen = 4;
assert (n % qk == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE)
if (ggml_sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
"__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
__asm__ __volatile__(
"movi v31.16b, #0x4\n"
"movi v30.16b, #0xf0\n"
"add %x[b_ptr], %x[b_ptr], #0x8\n"
"1:" // Column loop
"add x22, %x[a_ptr], #0x2\n"
"movi v29.16b, #0x0\n"
"mov x21, %x[nb]\n"
"2:" // Block loop
"ldr q28, [%x[b_ptr], #0x0]\n"
"ldr q27, [x22, #0x0]\n"
"movi v26.4s, #0x0\n"
"sub x20, x22, #0x2\n"
"ldr q25, [x22, #0x10]\n"
"ldr q24, [%x[b_ptr], #0x10]\n"
"sub x21, x21, #0x1\n"
"add x22, x22, #0x22\n"
"ldr q23, [%x[b_ptr], #0x20]\n"
"ldr q22, [%x[b_ptr], #0x30]\n"
"ld1r { v21.8h }, [x20]\n"
"ldr q20, [%x[b_ptr], #-0x8]\n"
"sshl v16.16b, v28.16b, v31.16b\n"
"and v28.16b, v28.16b, v30.16b\n"
"sshl v19.16b, v24.16b, v31.16b\n"
"and v24.16b, v24.16b, v30.16b\n"
"add %x[b_ptr], %x[b_ptr], #0x48\n"
"sshl v18.16b, v23.16b, v31.16b\n"
"and v23.16b, v23.16b, v30.16b\n"
".inst 0x4f9be21a // sdot v26.4s, v16.16b, v27.4b[0]\n"
"sshl v17.16b, v22.16b, v31.16b\n"
"and v22.16b, v22.16b, v30.16b\n"
"fcvtl v21.4s, v21.4h\n"
"fcvtl v16.4s, v20.4h\n"
".inst 0x4f99e39a // sdot v26.4s, v28.16b, v25.4b[0]\n"
"fmul v16.4s, v16.4s, v21.4s\n"
".inst 0x4fbbe27a // sdot v26.4s, v19.16b, v27.4b[1]\n"
".inst 0x4fb9e31a // sdot v26.4s, v24.16b, v25.4b[1]\n"
".inst 0x4f9bea5a // sdot v26.4s, v18.16b, v27.4b[2]\n"
".inst 0x4f99eafa // sdot v26.4s, v23.16b, v25.4b[2]\n"
".inst 0x4fbbea3a // sdot v26.4s, v17.16b, v27.4b[3]\n"
".inst 0x4fb9eada // sdot v26.4s, v22.16b, v25.4b[3]\n"
"scvtf v26.4s, v26.4s, #0x4\n"
"fmla v29.4s, v26.4s, v16.4s\n"
"cbnz x21, 2b\n"
"sub %x[nc], %x[nc], #0x4\n"
"str q29, [%x[res_ptr], #0x0]\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"cbnz %x[nc], 1b\n"
: [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
: [a_ptr] "r" (a_ptr), [nb] "r" (nb)
: "memory", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22"
);
#else
float sumf[4];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
#endif
}
void ggml_gemv_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
const int blocklen = 8;
assert (n % qk == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE)
if (ggml_sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
__asm__ __volatile__(
"movi v2.16b, #0x4\n"
"movi v1.16b, #0xf0\n"
"add %x[b_ptr], %x[b_ptr], #0x8\n"
"1:" // Column loop
"add x23, %x[a_ptr], #0x2\n"
"movi v0.16b, #0x0\n"
"mov x22, %x[nb]\n"
"2:" // Block loop
"ldr q31, [%x[b_ptr], #0x0]\n"
"ldr q30, [%x[b_ptr], #0x10]\n"
"mov x21, x23\n"
"movi v29.4s, #0x0\n"
"ldr q28, [%x[b_ptr], #0x20]\n"
"ldr q27, [%x[b_ptr], #0x30]\n"
"movi v26.4s, #0x0\n"
"sub x20, x23, #0x2\n"
"ld1r { v25.8h }, [x20]\n"
"ldr q24, [%x[b_ptr], #-0x8]\n"
"sub x22, x22, #0x1\n"
"add x23, x23, #0x22\n"
"ld1r { v23.2d }, [x21], #0x8\n"
"sshl v22.16b, v31.16b, v2.16b\n"
"sshl v16.16b, v30.16b, v2.16b\n"
"add %x[b_ptr], %x[b_ptr], #0x48\n"
"ld1r { v21.2d }, [x21], #0x8\n"
"sshl v20.16b, v28.16b, v2.16b\n"
"sshl v19.16b, v27.16b, v2.16b\n"
"ld1r { v18.2d }, [x21], #0x8\n"
"ld1r { v17.2d }, [x21], #0x8\n"
"and v31.16b, v31.16b, v1.16b\n"
"and v30.16b, v30.16b, v1.16b\n"
".inst 0x4e9796dd // sdot v29.4s, v22.16b, v23.16b\n"
".inst 0x4e97961a // sdot v26.4s, v16.16b, v23.16b\n"
"and v28.16b, v28.16b, v1.16b\n"
"and v27.16b, v27.16b, v1.16b\n"
"fcvtl v25.4s, v25.4h\n"
"fcvtl v16.4s, v24.4h\n"
".inst 0x4e95969d // sdot v29.4s, v20.16b, v21.16b\n"
".inst 0x4e95967a // sdot v26.4s, v19.16b, v21.16b\n"
"fmul v16.4s, v16.4s, v25.4s\n"
".inst 0x4e9297fd // sdot v29.4s, v31.16b, v18.16b\n"
".inst 0x4e9297da // sdot v26.4s, v30.16b, v18.16b\n"
".inst 0x4e91979d // sdot v29.4s, v28.16b, v17.16b\n"
".inst 0x4e91977a // sdot v26.4s, v27.16b, v17.16b\n"
"addp v29.4s, v29.4s, v26.4s\n"
"scvtf v29.4s, v29.4s, #0x4\n"
"fmla v0.4s, v29.4s, v16.4s\n"
"cbnz x22, 2b\n"
"sub %x[nc], %x[nc], #0x4\n"
"str q0, [%x[res_ptr], #0x0]\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"cbnz %x[nc], 1b\n"
: [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
: [a_ptr] "r" (a_ptr), [nb] "r" (nb)
: "memory", "v0", "v1", "v2", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x20", "x21", "x22", "x23"
);
#elif defined(__ARM_NEON) && defined(__aarch64__)
GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
"performance");
#else
float sumf[4];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
#endif
}
void ggml_gemv_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert (n % qk == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
if (ggml_sve_cnt_b == QK8_0) {
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
__asm__ __volatile__(
"ptrue p0.b\n"
"add %x[b_ptr], %x[b_ptr], #0x10\n"
"1:" // Column loop
"add x22, %x[a_ptr], #0x2\n"
"mov z31.b, #0x0\n"
"mov x21, %x[nb]\n"
"2:" // Block loop
"ld1b { z30.b }, p0/Z, [%x[b_ptr]]\n"
"ld1b { z29.b }, p0/Z, [%x[b_ptr], #1, MUL VL]\n"
"mov z28.s, #0x0\n"
"mov z27.s, #0x0\n"
"ld1rd { z26.d }, p0/Z, [x22]\n"
"ld1b { z25.b }, p0/Z, [%x[b_ptr], #2, MUL VL]\n"
"sub x20, x22, #0x2\n"
"sub x21, x21, #0x1\n"
"ld1b { z24.b }, p0/Z, [%x[b_ptr], #3, MUL VL]\n"
"ld1rd { z23.d }, p0/Z, [x22, #8]\n"
"lsl z22.b, z30.b, #0x4\n"
"lsl z16.b, z29.b, #0x4\n"
"and z30.b, z30.b, #0xf0\n"
"and z29.b, z29.b, #0xf0\n"
"ld1rd { z21.d }, p0/Z, [x22, #16]\n"
"ld1rd { z20.d }, p0/Z, [x22, #24]\n"
"lsl z19.b, z25.b, #0x4\n"
"and z25.b, z25.b, #0xf0\n"
"ld1rh { z17.h }, p0/Z, [x20]\n"
"ld1h { z18.s }, p0/Z, [%x[b_ptr], #-1, MUL VL]\n"
"sdot z28.s, z22.b, z26.b\n"
"sdot z27.s, z16.b, z26.b\n"
"lsl z16.b, z24.b, #0x4\n"
"add x22, x22, #0x22\n"
"and z24.b, z24.b, #0xf0\n"
"add %x[b_ptr], %x[b_ptr], #0x90\n"
"fcvt z17.s, p0/m, z17.h\n"
"fcvt z18.s, p0/m, z18.h\n"
"sdot z28.s, z19.b, z23.b\n"
"sdot z27.s, z16.b, z23.b\n"
"fmul z18.s, z18.s, z17.s\n"
"sdot z28.s, z30.b, z21.b\n"
"sdot z27.s, z29.b, z21.b\n"
"sdot z28.s, z25.b, z20.b\n"
"sdot z27.s, z24.b, z20.b\n"
"uzp1 z17.s, z28.s, z27.s\n"
"uzp2 z16.s, z28.s, z27.s\n"
"add z17.s, z17.s, z16.s\n"
"asr z17.s, z17.s, #0x4\n"
"scvtf z17.s, p0/m, z17.s\n"
"fmla z31.s, p0/M, z17.s, z18.s\n"
"cbnz x21, 2b\n"
"sub %x[nc], %x[nc], #0x8\n"
"st1w { z31.s }, p0, [%x[res_ptr]]\n"
"add %x[res_ptr], %x[res_ptr], #0x20\n"
"cbnz %x[nc], 1b\n"
: [b_ptr] "+&r" (b_ptr), [res_ptr] "+&r" (res_ptr), [nc] "+&r" (nc)
: [a_ptr] "r" (a_ptr), [nb] "r" (nb)
: "memory", "p0", "x20", "x21", "x22", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
return;
}
else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
"performance");
}
else if (ggml_cpu_has_neon()) {
GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
"quantization format for optimal performance");
}
#endif
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
GGML_ASSERT(ggml_cpu_has_sve() &&
"__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance");
#elif defined(__ARM_NEON) && defined(__aarch64__)
GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
"performance");
#else
float sumf[8];
int sumi;
const block_q8_0 * a_ptr = (const block_q8_0 *) vy;
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int j = 0; j < ncols_interleaved; j++) sumf[j] = 0.0;
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * blocklen + i]) + (v1 * a_ptr[l].qs[k * blocklen + i + qk / 2])) >> 4;
}
sumf[j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d);
}
}
}
for (int j = 0; j < ncols_interleaved; j++) s[x * ncols_interleaved + j] = sumf[j];
}
#endif
}
void ggml_gemm_q4_0_4x4_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
const int blocklen = 4;
assert (n % qk == 0);
assert (nr % 4 == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
if (ggml_sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
GGML_ASSERT(!(ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) &&
"__ARM_NEON and __ARM_FEATURE_MATMUL_INT8 defined, use the Q4_0_4_8 quantization format for optimal performance");
#elif defined(__ARM_NEON) && defined(__aarch64__) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
size_t res_stride = bs * sizeof(float);
__asm__ __volatile__(
"mov x10, %x[nr]\n"
"mov x9, #0x88\n"
"cmp x10, #0x10\n"
"mul x9, %x[nb], x9\n"
"blt 4f\n"
"1:" // Row loop
"add x28, %x[b_ptr], #0x8\n"
"mov x27, %x[nc]\n"
"add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
"2:" // Column loop
"add x25, %x[a_ptr], #0x8\n"
"movi v15.16b, #0x0\n"
"movi v19.16b, #0x0\n"
"mov x24, %x[nb]\n"
"add x23, x25, x9\n"
"movi v18.16b, #0x0\n"
"movi v14.16b, #0x0\n"
"add x22, x23, x9\n"
"movi v11.16b, #0x0\n"
"movi v13.16b, #0x0\n"
"add x21, x22, x9\n"
"movi v23.16b, #0x0\n"
"movi v16.16b, #0x0\n"
"movi v25.16b, #0x0\n"
"movi v7.16b, #0x0\n"
"movi v0.16b, #0x0\n"
"movi v4.16b, #0x0\n"
"movi v5.16b, #0x0\n"
"movi v21.16b, #0x0\n"
"movi v8.16b, #0x0\n"
"movi v1.16b, #0x0\n"
"3:" // Block loop
"ldr q3, [x28, #0x0]\n"
"ldr q31, [x25, #0x0]\n"
"movi v28.16b, #0x4\n"
"movi v10.4s, #0x0\n"
"ldr q22, [x28, #0x10]\n"
"ldr q6, [x25, #0x10]\n"
"movi v29.4s, #0x0\n"
"movi v9.4s, #0x0\n"
"ldr q27, [x28, #0x20]\n"
"ldr q30, [x28, #0x30]\n"
"movi v20.4s, #0x0\n"
"movi v24.16b, #0xf0\n"
"ldr d2, [x25, #-0x8]\n"
"ldr d26, [x23, #-0x8]\n"
"sshl v12.16b, v3.16b, v28.16b\n"
"sub x20, x28, #0x8\n"
"ldr d17, [x20, #0x0]\n"
"and v3.16b, v3.16b, v24.16b\n"
"subs x24, x24, #0x1\n"
"add x28, x28, #0x48\n"
".inst 0x4f9fe18a // sdot v10.4s, v12.16b, v31.4b[0]\n"
".inst 0x4fbfe19d // sdot v29.4s, v12.16b, v31.4b[1]\n"
".inst 0x4f9fe989 // sdot v9.4s, v12.16b, v31.4b[2]\n"
".inst 0x4fbfe994 // sdot v20.4s, v12.16b, v31.4b[3]\n"
"sshl v31.16b, v22.16b, v28.16b\n"
"and v22.16b, v22.16b, v24.16b\n"
"fcvtl v17.4s, v17.4h\n"
"fcvtl v2.4s, v2.4h\n"
"fcvtl v26.4s, v26.4h\n"
".inst 0x4f86e3ea // sdot v10.4s, v31.16b, v6.4b[0]\n"
".inst 0x4fa6e3fd // sdot v29.4s, v31.16b, v6.4b[1]\n"
".inst 0x4f86ebe9 // sdot v9.4s, v31.16b, v6.4b[2]\n"
".inst 0x4fa6ebf4 // sdot v20.4s, v31.16b, v6.4b[3]\n"
"sshl v6.16b, v27.16b, v28.16b\n"
"sshl v28.16b, v30.16b, v28.16b\n"
"and v27.16b, v27.16b, v24.16b\n"
"and v30.16b, v30.16b, v24.16b\n"
"ldr q24, [x25, #0x20]\n"
".inst 0x4f98e0ca // sdot v10.4s, v6.16b, v24.4b[0]\n"
".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n"
".inst 0x4f98e8c9 // sdot v9.4s, v6.16b, v24.4b[2]\n"
".inst 0x4fb8e8d4 // sdot v20.4s, v6.16b, v24.4b[3]\n"
"ldr q24, [x25, #0x30]\n"
".inst 0x4f98e38a // sdot v10.4s, v28.16b, v24.4b[0]\n"
".inst 0x4fb8e39d // sdot v29.4s, v28.16b, v24.4b[1]\n"
".inst 0x4f98eb89 // sdot v9.4s, v28.16b, v24.4b[2]\n"
".inst 0x4fb8eb94 // sdot v20.4s, v28.16b, v24.4b[3]\n"
"ldr q24, [x25, #0x40]\n"
".inst 0x4f98e06a // sdot v10.4s, v3.16b, v24.4b[0]\n"
".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n"
".inst 0x4f98e869 // sdot v9.4s, v3.16b, v24.4b[2]\n"
".inst 0x4fb8e874 // sdot v20.4s, v3.16b, v24.4b[3]\n"
"ldr q24, [x25, #0x50]\n"
".inst 0x4f98e2ca // sdot v10.4s, v22.16b, v24.4b[0]\n"
".inst 0x4fb8e2dd // sdot v29.4s, v22.16b, v24.4b[1]\n"
".inst 0x4f98eac9 // sdot v9.4s, v22.16b, v24.4b[2]\n"
".inst 0x4fb8ead4 // sdot v20.4s, v22.16b, v24.4b[3]\n"
"ldr q24, [x25, #0x60]\n"
".inst 0x4f98e36a // sdot v10.4s, v27.16b, v24.4b[0]\n"
".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n"
".inst 0x4f98eb69 // sdot v9.4s, v27.16b, v24.4b[2]\n"
".inst 0x4fb8eb74 // sdot v20.4s, v27.16b, v24.4b[3]\n"
"ldr q24, [x25, #0x70]\n"
"add x25, x25, #0x88\n"
".inst 0x4f98e3ca // sdot v10.4s, v30.16b, v24.4b[0]\n"
".inst 0x4fb8e3dd // sdot v29.4s, v30.16b, v24.4b[1]\n"
".inst 0x4f98ebc9 // sdot v9.4s, v30.16b, v24.4b[2]\n"
".inst 0x4fb8ebd4 // sdot v20.4s, v30.16b, v24.4b[3]\n"
"fmul v24.4s, v17.4s, v2.s[0]\n"
"scvtf v10.4s, v10.4s, #0x4\n"
"scvtf v29.4s, v29.4s, #0x4\n"
"scvtf v9.4s, v9.4s, #0x4\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"fmla v15.4s, v10.4s, v24.4s\n"
"ldr q24, [x23, #0x0]\n"
"fmul v10.4s, v17.4s, v2.s[1]\n"
"fmla v19.4s, v29.4s, v10.4s\n"
"ldr q10, [x23, #0x10]\n"
"fmul v29.4s, v17.4s, v2.s[2]\n"
"fmul v2.4s, v17.4s, v2.s[3]\n"
"fmla v18.4s, v9.4s, v29.4s\n"
"movi v9.4s, #0x0\n"
"movi v29.4s, #0x0\n"
".inst 0x4f98e189 // sdot v9.4s, v12.16b, v24.4b[0]\n"
".inst 0x4fb8e19d // sdot v29.4s, v12.16b, v24.4b[1]\n"
"fmla v14.4s, v20.4s, v2.4s\n"
"movi v20.4s, #0x0\n"
"movi v2.4s, #0x0\n"
".inst 0x4f98e994 // sdot v20.4s, v12.16b, v24.4b[2]\n"
".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n"
"ldr q24, [x23, #0x20]\n"
".inst 0x4f8ae3e9 // sdot v9.4s, v31.16b, v10.4b[0]\n"
".inst 0x4faae3fd // sdot v29.4s, v31.16b, v10.4b[1]\n"
".inst 0x4f8aebf4 // sdot v20.4s, v31.16b, v10.4b[2]\n"
".inst 0x4faaebe2 // sdot v2.4s, v31.16b, v10.4b[3]\n"
"ldr q10, [x23, #0x30]\n"
".inst 0x4f98e0c9 // sdot v9.4s, v6.16b, v24.4b[0]\n"
".inst 0x4fb8e0dd // sdot v29.4s, v6.16b, v24.4b[1]\n"
".inst 0x4f98e8d4 // sdot v20.4s, v6.16b, v24.4b[2]\n"
".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n"
"ldr q24, [x23, #0x40]\n"
".inst 0x4f8ae389 // sdot v9.4s, v28.16b, v10.4b[0]\n"
".inst 0x4faae39d // sdot v29.4s, v28.16b, v10.4b[1]\n"
".inst 0x4f8aeb94 // sdot v20.4s, v28.16b, v10.4b[2]\n"
".inst 0x4faaeb82 // sdot v2.4s, v28.16b, v10.4b[3]\n"
"ldr q10, [x23, #0x50]\n"
".inst 0x4f98e069 // sdot v9.4s, v3.16b, v24.4b[0]\n"
".inst 0x4fb8e07d // sdot v29.4s, v3.16b, v24.4b[1]\n"
".inst 0x4f98e874 // sdot v20.4s, v3.16b, v24.4b[2]\n"
".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n"
"ldr q24, [x23, #0x60]\n"
".inst 0x4f8ae2c9 // sdot v9.4s, v22.16b, v10.4b[0]\n"
".inst 0x4faae2dd // sdot v29.4s, v22.16b, v10.4b[1]\n"
".inst 0x4f8aead4 // sdot v20.4s, v22.16b, v10.4b[2]\n"
".inst 0x4faaeac2 // sdot v2.4s, v22.16b, v10.4b[3]\n"
"ldr q10, [x23, #0x70]\n"
"add x23, x23, #0x88\n"
".inst 0x4f98e369 // sdot v9.4s, v27.16b, v24.4b[0]\n"
".inst 0x4fb8e37d // sdot v29.4s, v27.16b, v24.4b[1]\n"
".inst 0x4f98eb74 // sdot v20.4s, v27.16b, v24.4b[2]\n"
".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n"
"ldr q24, [x22, #0x0]\n"
".inst 0x4f8ae3c9 // sdot v9.4s, v30.16b, v10.4b[0]\n"
".inst 0x4faae3dd // sdot v29.4s, v30.16b, v10.4b[1]\n"
".inst 0x4f8aebd4 // sdot v20.4s, v30.16b, v10.4b[2]\n"
".inst 0x4faaebc2 // sdot v2.4s, v30.16b, v10.4b[3]\n"
"fmul v10.4s, v17.4s, v26.s[0]\n"
"scvtf v9.4s, v9.4s, #0x4\n"
"scvtf v29.4s, v29.4s, #0x4\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"scvtf v2.4s, v2.4s, #0x4\n"
"fmla v11.4s, v9.4s, v10.4s\n"
"ldr q9, [x22, #0x10]\n"
"fmul v10.4s, v17.4s, v26.s[1]\n"
"fmla v13.4s, v29.4s, v10.4s\n"
"ldr d29, [x22, #-0x8]\n"
"fmul v10.4s, v17.4s, v26.s[2]\n"
"fmul v26.4s, v17.4s, v26.s[3]\n"
"fcvtl v29.4s, v29.4h\n"
"fmla v23.4s, v20.4s, v10.4s\n"
"movi v20.4s, #0x0\n"
"movi v10.4s, #0x0\n"
"fmla v16.4s, v2.4s, v26.4s\n"
"movi v26.4s, #0x0\n"
"movi v2.4s, #0x0\n"
".inst 0x4f98e194 // sdot v20.4s, v12.16b, v24.4b[0]\n"
".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n"
".inst 0x4f98e99a // sdot v26.4s, v12.16b, v24.4b[2]\n"
".inst 0x4fb8e982 // sdot v2.4s, v12.16b, v24.4b[3]\n"
"ldr q24, [x22, #0x20]\n"
".inst 0x4f89e3f4 // sdot v20.4s, v31.16b, v9.4b[0]\n"
".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n"
".inst 0x4f89ebfa // sdot v26.4s, v31.16b, v9.4b[2]\n"
".inst 0x4fa9ebe2 // sdot v2.4s, v31.16b, v9.4b[3]\n"
"ldr q9, [x22, #0x30]\n"
".inst 0x4f98e0d4 // sdot v20.4s, v6.16b, v24.4b[0]\n"
".inst 0x4fb8e0ca // sdot v10.4s, v6.16b, v24.4b[1]\n"
".inst 0x4f98e8da // sdot v26.4s, v6.16b, v24.4b[2]\n"
".inst 0x4fb8e8c2 // sdot v2.4s, v6.16b, v24.4b[3]\n"
"ldr q24, [x22, #0x40]\n"
".inst 0x4f89e394 // sdot v20.4s, v28.16b, v9.4b[0]\n"
".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n"
".inst 0x4f89eb9a // sdot v26.4s, v28.16b, v9.4b[2]\n"
".inst 0x4fa9eb82 // sdot v2.4s, v28.16b, v9.4b[3]\n"
"ldr q9, [x22, #0x50]\n"
".inst 0x4f98e074 // sdot v20.4s, v3.16b, v24.4b[0]\n"
".inst 0x4fb8e06a // sdot v10.4s, v3.16b, v24.4b[1]\n"
".inst 0x4f98e87a // sdot v26.4s, v3.16b, v24.4b[2]\n"
".inst 0x4fb8e862 // sdot v2.4s, v3.16b, v24.4b[3]\n"
"ldr q24, [x22, #0x60]\n"
".inst 0x4f89e2d4 // sdot v20.4s, v22.16b, v9.4b[0]\n"
".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n"
".inst 0x4f89eada // sdot v26.4s, v22.16b, v9.4b[2]\n"
".inst 0x4fa9eac2 // sdot v2.4s, v22.16b, v9.4b[3]\n"
"ldr q9, [x22, #0x70]\n"
"add x22, x22, #0x88\n"
".inst 0x4f98e374 // sdot v20.4s, v27.16b, v24.4b[0]\n"
".inst 0x4fb8e36a // sdot v10.4s, v27.16b, v24.4b[1]\n"
".inst 0x4f98eb7a // sdot v26.4s, v27.16b, v24.4b[2]\n"
".inst 0x4fb8eb62 // sdot v2.4s, v27.16b, v24.4b[3]\n"
"ldr q24, [x21, #0x0]\n"
".inst 0x4f89e3d4 // sdot v20.4s, v30.16b, v9.4b[0]\n"
".inst 0x4fa9e3ca // sdot v10.4s, v30.16b, v9.4b[1]\n"
".inst 0x4f89ebda // sdot v26.4s, v30.16b, v9.4b[2]\n"
".inst 0x4fa9ebc2 // sdot v2.4s, v30.16b, v9.4b[3]\n"
"fmul v9.4s, v17.4s, v29.s[0]\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"scvtf v10.4s, v10.4s, #0x4\n"
"scvtf v26.4s, v26.4s, #0x4\n"
"scvtf v2.4s, v2.4s, #0x4\n"
"fmla v25.4s, v20.4s, v9.4s\n"
"ldr q9, [x21, #0x10]\n"
"fmul v20.4s, v17.4s, v29.s[1]\n"
"fmla v7.4s, v10.4s, v20.4s\n"
"ldr d20, [x21, #-0x8]\n"
"fmul v10.4s, v17.4s, v29.s[2]\n"
"fmul v29.4s, v17.4s, v29.s[3]\n"
"fcvtl v20.4s, v20.4h\n"
"fmla v0.4s, v26.4s, v10.4s\n"
"movi v26.4s, #0x0\n"
"movi v10.4s, #0x0\n"
"fmla v4.4s, v2.4s, v29.4s\n"
"movi v2.4s, #0x0\n"
"movi v29.4s, #0x0\n"
".inst 0x4f98e19a // sdot v26.4s, v12.16b, v24.4b[0]\n"
".inst 0x4fb8e18a // sdot v10.4s, v12.16b, v24.4b[1]\n"
".inst 0x4f98e982 // sdot v2.4s, v12.16b, v24.4b[2]\n"
".inst 0x4fb8e99d // sdot v29.4s, v12.16b, v24.4b[3]\n"
"ldr q12, [x21, #0x20]\n"
"fmul v24.4s, v17.4s, v20.s[0]\n"
".inst 0x4f89e3fa // sdot v26.4s, v31.16b, v9.4b[0]\n"
".inst 0x4fa9e3ea // sdot v10.4s, v31.16b, v9.4b[1]\n"
".inst 0x4f89ebe2 // sdot v2.4s, v31.16b, v9.4b[2]\n"
".inst 0x4fa9ebfd // sdot v29.4s, v31.16b, v9.4b[3]\n"
"ldr q9, [x21, #0x30]\n"
"fmul v31.4s, v17.4s, v20.s[1]\n"
".inst 0x4f8ce0da // sdot v26.4s, v6.16b, v12.4b[0]\n"
".inst 0x4face0ca // sdot v10.4s, v6.16b, v12.4b[1]\n"
".inst 0x4f8ce8c2 // sdot v2.4s, v6.16b, v12.4b[2]\n"
".inst 0x4face8dd // sdot v29.4s, v6.16b, v12.4b[3]\n"
"ldr q12, [x21, #0x40]\n"
"fmul v6.4s, v17.4s, v20.s[2]\n"
"fmul v20.4s, v17.4s, v20.s[3]\n"
".inst 0x4f89e39a // sdot v26.4s, v28.16b, v9.4b[0]\n"
".inst 0x4fa9e38a // sdot v10.4s, v28.16b, v9.4b[1]\n"
".inst 0x4f89eb82 // sdot v2.4s, v28.16b, v9.4b[2]\n"
".inst 0x4fa9eb9d // sdot v29.4s, v28.16b, v9.4b[3]\n"
"ldr q9, [x21, #0x50]\n"
".inst 0x4f8ce07a // sdot v26.4s, v3.16b, v12.4b[0]\n"
".inst 0x4face06a // sdot v10.4s, v3.16b, v12.4b[1]\n"
".inst 0x4f8ce862 // sdot v2.4s, v3.16b, v12.4b[2]\n"
".inst 0x4face87d // sdot v29.4s, v3.16b, v12.4b[3]\n"
"ldr q12, [x21, #0x60]\n"
".inst 0x4f89e2da // sdot v26.4s, v22.16b, v9.4b[0]\n"
".inst 0x4fa9e2ca // sdot v10.4s, v22.16b, v9.4b[1]\n"
".inst 0x4f89eac2 // sdot v2.4s, v22.16b, v9.4b[2]\n"
".inst 0x4fa9eadd // sdot v29.4s, v22.16b, v9.4b[3]\n"
"ldr q17, [x21, #0x70]\n"
"add x21, x21, #0x88\n"
".inst 0x4f8ce37a // sdot v26.4s, v27.16b, v12.4b[0]\n"
".inst 0x4face36a // sdot v10.4s, v27.16b, v12.4b[1]\n"
".inst 0x4f8ceb62 // sdot v2.4s, v27.16b, v12.4b[2]\n"
".inst 0x4faceb7d // sdot v29.4s, v27.16b, v12.4b[3]\n"
".inst 0x4f91e3da // sdot v26.4s, v30.16b, v17.4b[0]\n"
".inst 0x4fb1e3ca // sdot v10.4s, v30.16b, v17.4b[1]\n"
".inst 0x4f91ebc2 // sdot v2.4s, v30.16b, v17.4b[2]\n"
".inst 0x4fb1ebdd // sdot v29.4s, v30.16b, v17.4b[3]\n"
"scvtf v26.4s, v26.4s, #0x4\n"
"scvtf v10.4s, v10.4s, #0x4\n"
"fmla v5.4s, v26.4s, v24.4s\n"
"scvtf v2.4s, v2.4s, #0x4\n"
"scvtf v29.4s, v29.4s, #0x4\n"
"fmla v21.4s, v10.4s, v31.4s\n"
"fmla v8.4s, v2.4s, v6.4s\n"
"fmla v1.4s, v29.4s, v20.4s\n"
"bgt 3b\n"
"mov x20, %x[res_ptr]\n"
"subs x27, x27, #0x4\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"str q15, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q19, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q18, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q14, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q11, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q13, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q23, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q16, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q25, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q7, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q0, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q4, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q5, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q21, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q8, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q1, [x20, #0x0]\n"
"bne 2b\n"
"mov x20, #0x4\n"
"sub x10, x10, #0x10\n"
"cmp x10, #0x10\n"
"mov %x[res_ptr], x26\n"
"madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
"bge 1b\n"
"4:" // Row loop skip
"cbz x10, 9f\n"
"5:" // Row tail: Row loop
"add x24, %x[b_ptr], #0x8\n"
"mov x23, %x[nc]\n"
"add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
"6:" // Row tail: Column loop
"movi v15.16b, #0x0\n"
"movi v19.16b, #0x0\n"
"add x25, %x[a_ptr], #0x8\n"
"mov x21, %x[nb]\n"
"movi v18.16b, #0x0\n"
"movi v14.16b, #0x0\n"
"7:" // Row tail: Block loop
"ldr q7, [x24, #0x0]\n"
"ldr q5, [x25, #0x0]\n"
"movi v9.16b, #0x4\n"
"movi v4.4s, #0x0\n"
"ldr q3, [x24, #0x10]\n"
"ldr q2, [x25, #0x10]\n"
"movi v1.4s, #0x0\n"
"movi v0.4s, #0x0\n"
"ldr q13, [x24, #0x20]\n"
"ldr q31, [x25, #0x20]\n"
"movi v30.4s, #0x0\n"
"movi v29.16b, #0xf0\n"
"ldr q28, [x24, #0x30]\n"
"ldr q27, [x25, #0x30]\n"
"sshl v20.16b, v7.16b, v9.16b\n"
"sub x20, x24, #0x8\n"
"ldr q26, [x25, #0x40]\n"
"ldr q25, [x25, #0x50]\n"
"sshl v17.16b, v3.16b, v9.16b\n"
"and v7.16b, v7.16b, v29.16b\n"
"ldr q24, [x25, #0x60]\n"
"ldr q16, [x25, #0x70]\n"
"sshl v22.16b, v13.16b, v9.16b\n"
"and v3.16b, v3.16b, v29.16b\n"
"ldr d21, [x20, #0x0]\n"
"ldr d12, [x25, #-0x8]\n"
".inst 0x4f85e284 // sdot v4.4s, v20.16b, v5.4b[0]\n"
".inst 0x4fa5e281 // sdot v1.4s, v20.16b, v5.4b[1]\n"
".inst 0x4f85ea80 // sdot v0.4s, v20.16b, v5.4b[2]\n"
".inst 0x4fa5ea9e // sdot v30.4s, v20.16b, v5.4b[3]\n"
"sshl v9.16b, v28.16b, v9.16b\n"
"subs x21, x21, #0x1\n"
"and v13.16b, v13.16b, v29.16b\n"
"and v28.16b, v28.16b, v29.16b\n"
"add x25, x25, #0x88\n"
"add x24, x24, #0x48\n"
"fcvtl v21.4s, v21.4h\n"
"fcvtl v12.4s, v12.4h\n"
".inst 0x4f82e224 // sdot v4.4s, v17.16b, v2.4b[0]\n"
".inst 0x4fa2e221 // sdot v1.4s, v17.16b, v2.4b[1]\n"
".inst 0x4f82ea20 // sdot v0.4s, v17.16b, v2.4b[2]\n"
".inst 0x4fa2ea3e // sdot v30.4s, v17.16b, v2.4b[3]\n"
"fmul v11.4s, v21.4s, v12.s[0]\n"
"fmul v23.4s, v21.4s, v12.s[1]\n"
"fmul v17.4s, v21.4s, v12.s[2]\n"
".inst 0x4f9fe2c4 // sdot v4.4s, v22.16b, v31.4b[0]\n"
"fmul v6.4s, v21.4s, v12.s[3]\n"
".inst 0x4fbfe2c1 // sdot v1.4s, v22.16b, v31.4b[1]\n"
".inst 0x4f9feac0 // sdot v0.4s, v22.16b, v31.4b[2]\n"
".inst 0x4fbfeade // sdot v30.4s, v22.16b, v31.4b[3]\n"
".inst 0x4f9be124 // sdot v4.4s, v9.16b, v27.4b[0]\n"
".inst 0x4fbbe121 // sdot v1.4s, v9.16b, v27.4b[1]\n"
".inst 0x4f9be920 // sdot v0.4s, v9.16b, v27.4b[2]\n"
".inst 0x4fbbe93e // sdot v30.4s, v9.16b, v27.4b[3]\n"
".inst 0x4f9ae0e4 // sdot v4.4s, v7.16b, v26.4b[0]\n"
".inst 0x4fbae0e1 // sdot v1.4s, v7.16b, v26.4b[1]\n"
".inst 0x4f9ae8e0 // sdot v0.4s, v7.16b, v26.4b[2]\n"
".inst 0x4fbae8fe // sdot v30.4s, v7.16b, v26.4b[3]\n"
".inst 0x4f99e064 // sdot v4.4s, v3.16b, v25.4b[0]\n"
".inst 0x4fb9e061 // sdot v1.4s, v3.16b, v25.4b[1]\n"
".inst 0x4f99e860 // sdot v0.4s, v3.16b, v25.4b[2]\n"
".inst 0x4fb9e87e // sdot v30.4s, v3.16b, v25.4b[3]\n"
".inst 0x4f98e1a4 // sdot v4.4s, v13.16b, v24.4b[0]\n"
".inst 0x4fb8e1a1 // sdot v1.4s, v13.16b, v24.4b[1]\n"
".inst 0x4f98e9a0 // sdot v0.4s, v13.16b, v24.4b[2]\n"
".inst 0x4fb8e9be // sdot v30.4s, v13.16b, v24.4b[3]\n"
".inst 0x4f90e384 // sdot v4.4s, v28.16b, v16.4b[0]\n"
".inst 0x4fb0e381 // sdot v1.4s, v28.16b, v16.4b[1]\n"
".inst 0x4f90eb80 // sdot v0.4s, v28.16b, v16.4b[2]\n"
".inst 0x4fb0eb9e // sdot v30.4s, v28.16b, v16.4b[3]\n"
"scvtf v4.4s, v4.4s, #0x4\n"
"scvtf v1.4s, v1.4s, #0x4\n"
"scvtf v0.4s, v0.4s, #0x4\n"
"fmla v15.4s, v4.4s, v11.4s\n"
"scvtf v30.4s, v30.4s, #0x4\n"
"fmla v19.4s, v1.4s, v23.4s\n"
"fmla v18.4s, v0.4s, v17.4s\n"
"fmla v14.4s, v30.4s, v6.4s\n"
"bgt 7b\n"
"mov x20, %x[res_ptr]\n"
"cmp x10, #0x1\n"
"str q15, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"cmp x10, #0x2\n"
"str q19, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"cmp x10, #0x3\n"
"str q18, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"str q14, [x20, #0x0]\n"
"8:" // Row tail: Accumulator store skip
"subs x23, x23, #0x4\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"bne 6b\n"
"subs x10, x10, #0x4\n"
"add %x[a_ptr], %x[a_ptr], x9\n"
"mov %x[res_ptr], x22\n"
"bgt 5b\n"
"9:" // Row tail: Row loop skip
: [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
: [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
#else
float sumf[4][4];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
#endif
}
void ggml_gemm_q4_0_4x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 4;
const int blocklen = 8;
assert (n % qk == 0);
assert (nr % 4 == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8)
if (ggml_sve_cnt_b == QK8_0) {
GGML_ASSERT(!(ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE defined, use the Q4_0_8_8 quantization format for optimal performance");
}
#endif
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
size_t res_stride = bs * sizeof(float);
__asm__ __volatile__(
"mov x10, %x[nr]\n"
"mov x9, #0x88\n"
"cmp x10, #0x10\n"
"mul x9, %x[nb], x9\n"
"blt 4f\n"
"1:" // Row loop
"add x28, %x[b_ptr], #0x8\n"
"mov x27, %x[nc]\n"
"add x26, %x[res_ptr], %x[res_stride], LSL #4\n"
"2:" // Column loop
"add x25, %x[a_ptr], #0x8\n"
"movi v2.16b, #0x0\n"
"movi v10.16b, #0x0\n"
"mov x24, %x[nb]\n"
"add x23, x25, x9\n"
"movi v12.16b, #0x0\n"
"movi v28.16b, #0x0\n"
"add x22, x23, x9\n"
"movi v11.16b, #0x0\n"
"movi v13.16b, #0x0\n"
"add x21, x22, x9\n"
"movi v22.16b, #0x0\n"
"movi v23.16b, #0x0\n"
"movi v25.16b, #0x0\n"
"movi v5.16b, #0x0\n"
"movi v7.16b, #0x0\n"
"movi v4.16b, #0x0\n"
"movi v6.16b, #0x0\n"
"movi v30.16b, #0x0\n"
"movi v24.16b, #0x0\n"
"movi v14.16b, #0x0\n"
"3:" // Block loop
"ldr q21, [x28, #0x0]\n"
"ldr q16, [x28, #0x10]\n"
"movi v1.16b, #0x4\n"
"movi v19.4s, #0x0\n"
"ldr q27, [x25, #0x0]\n"
"ldr q15, [x25, #0x10]\n"
"movi v26.4s, #0x0\n"
"movi v18.4s, #0x0\n"
"ldr q29, [x28, #0x20]\n"
"ldr q3, [x28, #0x30]\n"
"movi v17.4s, #0x0\n"
"movi v0.16b, #0xf0\n"
"ldr d20, [x25, #-0x8]\n"
"ldr d9, [x23, #-0x8]\n"
"sshl v8.16b, v21.16b, v1.16b\n"
"sshl v31.16b, v16.16b, v1.16b\n"
"and v21.16b, v21.16b, v0.16b\n"
"and v16.16b, v16.16b, v0.16b\n"
"sub x20, x28, #0x8\n"
"subs x24, x24, #0x1\n"
"add x28, x28, #0x48\n"
".inst 0x4e88a773 // smmla v19.4s, v27.16b, v8.16b\n"
".inst 0x4e9fa77a // smmla v26.4s, v27.16b, v31.16b\n"
"ldr q27, [x25, #0x20]\n"
".inst 0x4e88a5f2 // smmla v18.4s, v15.16b, v8.16b\n"
".inst 0x4e9fa5f1 // smmla v17.4s, v15.16b, v31.16b\n"
"sshl v15.16b, v29.16b, v1.16b\n"
"sshl v1.16b, v3.16b, v1.16b\n"
"and v29.16b, v29.16b, v0.16b\n"
"and v3.16b, v3.16b, v0.16b\n"
"ldr q0, [x25, #0x30]\n"
"fcvtl v20.4s, v20.4h\n"
".inst 0x4e8fa773 // smmla v19.4s, v27.16b, v15.16b\n"
"fcvtl v9.4s, v9.4h\n"
".inst 0x4e81a77a // smmla v26.4s, v27.16b, v1.16b\n"
"ldr q27, [x25, #0x40]\n"
".inst 0x4e8fa412 // smmla v18.4s, v0.16b, v15.16b\n"
".inst 0x4e81a411 // smmla v17.4s, v0.16b, v1.16b\n"
"ldr q0, [x25, #0x50]\n"
".inst 0x4e95a773 // smmla v19.4s, v27.16b, v21.16b\n"
".inst 0x4e90a77a // smmla v26.4s, v27.16b, v16.16b\n"
"ldr q27, [x25, #0x60]\n"
".inst 0x4e95a412 // smmla v18.4s, v0.16b, v21.16b\n"
".inst 0x4e90a411 // smmla v17.4s, v0.16b, v16.16b\n"
"ldr q0, [x25, #0x70]\n"
"add x25, x25, #0x88\n"
".inst 0x4e9da773 // smmla v19.4s, v27.16b, v29.16b\n"
".inst 0x4e83a77a // smmla v26.4s, v27.16b, v3.16b\n"
"ldr d27, [x20, #0x0]\n"
".inst 0x4e9da412 // smmla v18.4s, v0.16b, v29.16b\n"
".inst 0x4e83a411 // smmla v17.4s, v0.16b, v3.16b\n"
"fcvtl v27.4s, v27.4h\n"
"uzp1 v0.2d, v19.2d, v26.2d\n"
"uzp2 v26.2d, v19.2d, v26.2d\n"
"fmul v19.4s, v27.4s, v20.s[0]\n"
"scvtf v0.4s, v0.4s, #0x4\n"
"scvtf v26.4s, v26.4s, #0x4\n"
"fmla v2.4s, v0.4s, v19.4s\n"
"ldr q19, [x23, #0x0]\n"
"uzp1 v0.2d, v18.2d, v17.2d\n"
"uzp2 v18.2d, v18.2d, v17.2d\n"
"fmul v17.4s, v27.4s, v20.s[1]\n"
"scvtf v0.4s, v0.4s, #0x4\n"
"scvtf v18.4s, v18.4s, #0x4\n"
"fmla v10.4s, v26.4s, v17.4s\n"
"ldr q17, [x23, #0x10]\n"
"fmul v26.4s, v27.4s, v20.s[2]\n"
"fmul v20.4s, v27.4s, v20.s[3]\n"
"fmla v12.4s, v0.4s, v26.4s\n"
"ldr d0, [x22, #-0x8]\n"
"ldr d26, [x21, #-0x8]\n"
"fcvtl v0.4s, v0.4h\n"
"fmla v28.4s, v18.4s, v20.4s\n"
"movi v20.4s, #0x0\n"
"movi v18.4s, #0x0\n"
".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n"
".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n"
"ldr q19, [x23, #0x20]\n"
"fcvtl v26.4s, v26.4h\n"
".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n"
".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n"
"ldr q19, [x23, #0x40]\n"
".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n"
".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n"
"ldr q19, [x23, #0x60]\n"
".inst 0x4e9da674 // smmla v20.4s, v19.16b, v29.16b\n"
".inst 0x4e83a672 // smmla v18.4s, v19.16b, v3.16b\n"
"uzp1 v19.2d, v20.2d, v18.2d\n"
"scvtf v19.4s, v19.4s, #0x4\n"
"uzp2 v20.2d, v20.2d, v18.2d\n"
"fmul v18.4s, v27.4s, v9.s[0]\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"fmla v11.4s, v19.4s, v18.4s\n"
"ldr q18, [x22, #0x0]\n"
"fmul v19.4s, v27.4s, v9.s[1]\n"
"fmla v13.4s, v20.4s, v19.4s\n"
"movi v19.4s, #0x0\n"
"movi v20.4s, #0x0\n"
".inst 0x4e88a633 // smmla v19.4s, v17.16b, v8.16b\n"
".inst 0x4e9fa634 // smmla v20.4s, v17.16b, v31.16b\n"
"ldr q17, [x23, #0x30]\n"
".inst 0x4e8fa633 // smmla v19.4s, v17.16b, v15.16b\n"
".inst 0x4e81a634 // smmla v20.4s, v17.16b, v1.16b\n"
"ldr q17, [x23, #0x50]\n"
".inst 0x4e95a633 // smmla v19.4s, v17.16b, v21.16b\n"
".inst 0x4e90a634 // smmla v20.4s, v17.16b, v16.16b\n"
"ldr q17, [x23, #0x70]\n"
"add x23, x23, #0x88\n"
".inst 0x4e9da633 // smmla v19.4s, v17.16b, v29.16b\n"
".inst 0x4e83a634 // smmla v20.4s, v17.16b, v3.16b\n"
"uzp1 v17.2d, v19.2d, v20.2d\n"
"scvtf v17.4s, v17.4s, #0x4\n"
"uzp2 v20.2d, v19.2d, v20.2d\n"
"fmul v19.4s, v27.4s, v9.s[2]\n"
"fmul v9.4s, v27.4s, v9.s[3]\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"fmla v22.4s, v17.4s, v19.4s\n"
"ldr q17, [x22, #0x10]\n"
"movi v19.4s, #0x0\n"
".inst 0x4e88a653 // smmla v19.4s, v18.16b, v8.16b\n"
"fmla v23.4s, v20.4s, v9.4s\n"
"movi v20.4s, #0x0\n"
"movi v9.4s, #0x0\n"
".inst 0x4e9fa654 // smmla v20.4s, v18.16b, v31.16b\n"
"ldr q18, [x22, #0x20]\n"
".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n"
".inst 0x4e8fa653 // smmla v19.4s, v18.16b, v15.16b\n"
".inst 0x4e81a654 // smmla v20.4s, v18.16b, v1.16b\n"
"ldr q18, [x22, #0x40]\n"
".inst 0x4e95a653 // smmla v19.4s, v18.16b, v21.16b\n"
".inst 0x4e90a654 // smmla v20.4s, v18.16b, v16.16b\n"
"ldr q18, [x22, #0x60]\n"
".inst 0x4e9da653 // smmla v19.4s, v18.16b, v29.16b\n"
".inst 0x4e83a654 // smmla v20.4s, v18.16b, v3.16b\n"
"movi v18.4s, #0x0\n"
".inst 0x4e9fa632 // smmla v18.4s, v17.16b, v31.16b\n"
"ldr q17, [x22, #0x30]\n"
".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n"
".inst 0x4e81a632 // smmla v18.4s, v17.16b, v1.16b\n"
"ldr q17, [x22, #0x50]\n"
".inst 0x4e95a629 // smmla v9.4s, v17.16b, v21.16b\n"
".inst 0x4e90a632 // smmla v18.4s, v17.16b, v16.16b\n"
"ldr q17, [x22, #0x70]\n"
"add x22, x22, #0x88\n"
".inst 0x4e9da629 // smmla v9.4s, v17.16b, v29.16b\n"
".inst 0x4e83a632 // smmla v18.4s, v17.16b, v3.16b\n"
"uzp1 v17.2d, v19.2d, v20.2d\n"
"uzp2 v20.2d, v19.2d, v20.2d\n"
"fmul v19.4s, v27.4s, v0.s[0]\n"
"scvtf v17.4s, v17.4s, #0x4\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"fmla v25.4s, v17.4s, v19.4s\n"
"ldr q19, [x21, #0x0]\n"
"fmul v17.4s, v27.4s, v0.s[1]\n"
"fmla v5.4s, v20.4s, v17.4s\n"
"ldr q17, [x21, #0x10]\n"
"uzp1 v20.2d, v9.2d, v18.2d\n"
"uzp2 v9.2d, v9.2d, v18.2d\n"
"fmul v18.4s, v27.4s, v0.s[2]\n"
"fmul v0.4s, v27.4s, v0.s[3]\n"
"scvtf v20.4s, v20.4s, #0x4\n"
"scvtf v9.4s, v9.4s, #0x4\n"
"fmla v7.4s, v20.4s, v18.4s\n"
"movi v20.4s, #0x0\n"
"movi v18.4s, #0x0\n"
".inst 0x4e88a674 // smmla v20.4s, v19.16b, v8.16b\n"
".inst 0x4e9fa672 // smmla v18.4s, v19.16b, v31.16b\n"
"ldr q19, [x21, #0x20]\n"
"fmla v4.4s, v9.4s, v0.4s\n"
"movi v9.4s, #0x0\n"
"movi v0.4s, #0x0\n"
".inst 0x4e88a629 // smmla v9.4s, v17.16b, v8.16b\n"
"fmul v8.4s, v27.4s, v26.s[0]\n"
".inst 0x4e9fa620 // smmla v0.4s, v17.16b, v31.16b\n"
"ldr q17, [x21, #0x30]\n"
".inst 0x4e8fa674 // smmla v20.4s, v19.16b, v15.16b\n"
"fmul v31.4s, v27.4s, v26.s[1]\n"
".inst 0x4e81a672 // smmla v18.4s, v19.16b, v1.16b\n"
"ldr q19, [x21, #0x40]\n"
".inst 0x4e8fa629 // smmla v9.4s, v17.16b, v15.16b\n"
"fmul v15.4s, v27.4s, v26.s[2]\n"
"fmul v27.4s, v27.4s, v26.s[3]\n"
".inst 0x4e81a620 // smmla v0.4s, v17.16b, v1.16b\n"
"ldr q1, [x21, #0x50]\n"
".inst 0x4e95a674 // smmla v20.4s, v19.16b, v21.16b\n"
".inst 0x4e90a672 // smmla v18.4s, v19.16b, v16.16b\n"
"ldr q26, [x21, #0x60]\n"
".inst 0x4e95a429 // smmla v9.4s, v1.16b, v21.16b\n"
".inst 0x4e90a420 // smmla v0.4s, v1.16b, v16.16b\n"
"ldr q21, [x21, #0x70]\n"
"add x21, x21, #0x88\n"
".inst 0x4e9da754 // smmla v20.4s, v26.16b, v29.16b\n"
".inst 0x4e83a752 // smmla v18.4s, v26.16b, v3.16b\n"
".inst 0x4e9da6a9 // smmla v9.4s, v21.16b, v29.16b\n"
".inst 0x4e83a6a0 // smmla v0.4s, v21.16b, v3.16b\n"
"uzp1 v29.2d, v20.2d, v18.2d\n"
"uzp2 v21.2d, v20.2d, v18.2d\n"
"scvtf v29.4s, v29.4s, #0x4\n"
"uzp1 v18.2d, v9.2d, v0.2d\n"
"uzp2 v16.2d, v9.2d, v0.2d\n"
"scvtf v21.4s, v21.4s, #0x4\n"
"fmla v6.4s, v29.4s, v8.4s\n"
"scvtf v18.4s, v18.4s, #0x4\n"
"scvtf v16.4s, v16.4s, #0x4\n"
"fmla v30.4s, v21.4s, v31.4s\n"
"fmla v24.4s, v18.4s, v15.4s\n"
"fmla v14.4s, v16.4s, v27.4s\n"
"bgt 3b\n"
"mov x20, %x[res_ptr]\n"
"subs x27, x27, #0x4\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"str q2, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q10, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q12, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q28, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q11, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q13, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q22, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q23, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q25, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q5, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q7, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q4, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q6, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q30, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q24, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"str q14, [x20, #0x0]\n"
"bne 2b\n"
"mov x20, #0x4\n"
"sub x10, x10, #0x10\n"
"cmp x10, #0x10\n"
"mov %x[res_ptr], x26\n"
"madd %x[a_ptr], x20, x9, %x[a_ptr]\n"
"bge 1b\n"
"4:" // Row loop skip
"cbz x10, 9f\n"
"5:" // Row tail: Row loop
"add x24, %x[b_ptr], #0x8\n"
"mov x23, %x[nc]\n"
"add x22, %x[res_ptr], %x[res_stride], LSL #2\n"
"6:" // Row tail: Column loop
"movi v2.16b, #0x0\n"
"movi v10.16b, #0x0\n"
"add x25, %x[a_ptr], #0x8\n"
"mov x21, %x[nb]\n"
"movi v12.16b, #0x0\n"
"movi v28.16b, #0x0\n"
"7:" // Row tail: Block loop
"ldr q6, [x24, #0x0]\n"
"ldr q5, [x24, #0x10]\n"
"movi v17.16b, #0x4\n"
"movi v8.4s, #0x0\n"
"ldr q4, [x25, #0x0]\n"
"ldr q13, [x25, #0x10]\n"
"movi v27.4s, #0x0\n"
"movi v0.4s, #0x0\n"
"ldr q31, [x24, #0x20]\n"
"ldr q14, [x24, #0x30]\n"
"movi v29.4s, #0x0\n"
"movi v22.16b, #0xf0\n"
"ldr q11, [x25, #0x20]\n"
"ldr q23, [x25, #0x30]\n"
"sshl v21.16b, v6.16b, v17.16b\n"
"sshl v16.16b, v5.16b, v17.16b\n"
"ldr q20, [x25, #0x40]\n"
"ldr q26, [x25, #0x50]\n"
"and v6.16b, v6.16b, v22.16b\n"
"and v5.16b, v5.16b, v22.16b\n"
"ldr q25, [x25, #0x60]\n"
"ldr q3, [x25, #0x70]\n"
"sshl v19.16b, v31.16b, v17.16b\n"
"sshl v18.16b, v14.16b, v17.16b\n"
"ldr d17, [x25, #-0x8]\n"
".inst 0x4e95a488 // smmla v8.4s, v4.16b, v21.16b\n"
".inst 0x4e90a49b // smmla v27.4s, v4.16b, v16.16b\n"
"and v31.16b, v31.16b, v22.16b\n"
".inst 0x4e95a5a0 // smmla v0.4s, v13.16b, v21.16b\n"
".inst 0x4e90a5bd // smmla v29.4s, v13.16b, v16.16b\n"
"and v14.16b, v14.16b, v22.16b\n"
"sub x20, x24, #0x8\n"
"ldr d16, [x20, #0x0]\n"
"subs x21, x21, #0x1\n"
"add x25, x25, #0x88\n"
"fcvtl v17.4s, v17.4h\n"
"add x24, x24, #0x48\n"
".inst 0x4e93a568 // smmla v8.4s, v11.16b, v19.16b\n"
".inst 0x4e92a57b // smmla v27.4s, v11.16b, v18.16b\n"
".inst 0x4e93a6e0 // smmla v0.4s, v23.16b, v19.16b\n"
".inst 0x4e92a6fd // smmla v29.4s, v23.16b, v18.16b\n"
"fcvtl v16.4s, v16.4h\n"
".inst 0x4e86a688 // smmla v8.4s, v20.16b, v6.16b\n"
".inst 0x4e85a69b // smmla v27.4s, v20.16b, v5.16b\n"
"fmul v23.4s, v16.4s, v17.s[0]\n"
"fmul v21.4s, v16.4s, v17.s[1]\n"
"fmul v1.4s, v16.4s, v17.s[2]\n"
"fmul v20.4s, v16.4s, v17.s[3]\n"
".inst 0x4e86a740 // smmla v0.4s, v26.16b, v6.16b\n"
".inst 0x4e85a75d // smmla v29.4s, v26.16b, v5.16b\n"
".inst 0x4e9fa728 // smmla v8.4s, v25.16b, v31.16b\n"
".inst 0x4e8ea73b // smmla v27.4s, v25.16b, v14.16b\n"
".inst 0x4e9fa460 // smmla v0.4s, v3.16b, v31.16b\n"
".inst 0x4e8ea47d // smmla v29.4s, v3.16b, v14.16b\n"
"uzp1 v19.2d, v8.2d, v27.2d\n"
"uzp2 v18.2d, v8.2d, v27.2d\n"
"scvtf v19.4s, v19.4s, #0x4\n"
"uzp1 v17.2d, v0.2d, v29.2d\n"
"uzp2 v16.2d, v0.2d, v29.2d\n"
"scvtf v18.4s, v18.4s, #0x4\n"
"fmla v2.4s, v19.4s, v23.4s\n"
"scvtf v17.4s, v17.4s, #0x4\n"
"scvtf v16.4s, v16.4s, #0x4\n"
"fmla v10.4s, v18.4s, v21.4s\n"
"fmla v12.4s, v17.4s, v1.4s\n"
"fmla v28.4s, v16.4s, v20.4s\n"
"bgt 7b\n"
"mov x20, %x[res_ptr]\n"
"cmp x10, #0x1\n"
"str q2, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"cmp x10, #0x2\n"
"str q10, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"cmp x10, #0x3\n"
"str q12, [x20, #0x0]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"str q28, [x20, #0x0]\n"
"8:" // Row tail: Accumulator store skip
"subs x23, x23, #0x4\n"
"add %x[res_ptr], %x[res_ptr], #0x10\n"
"bne 6b\n"
"subs x10, x10, #0x4\n"
"add %x[a_ptr], %x[a_ptr], x9\n"
"mov %x[res_ptr], x22\n"
"bgt 5b\n"
"9:" // Row tail: Row loop skip
: [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
: [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
: "cc", "memory", "v0", "v1", "v2", "v3", "v4", "v5", "v6", "v7", "v8", "v9", "v10", "v11", "v12", "v13", "v14", "v15", "v16", "v17", "v18", "v19", "v20", "v21", "v22", "v23", "v24", "v25", "v26", "v27", "v28", "v29", "v30", "v31", "x9", "x10", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28"
);
#elif defined(__ARM_NEON) && defined(__aarch64__)
GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
"performance");
#else
float sumf[4][4];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x4 * b_ptr = (const block_q4_0x4 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
#endif
}
void ggml_gemm_q4_0_8x8_q8_0(int n, float * restrict s, size_t bs, const void * restrict vx, const void * restrict vy, int nr, int nc) {
const int qk = QK8_0;
const int nb = n / qk;
const int ncols_interleaved = 8;
const int blocklen = 8;
assert (n % qk == 0);
assert (nr % 4 == 0);
assert (nc % ncols_interleaved == 0);
UNUSED(s);
UNUSED(bs);
UNUSED(vx);
UNUSED(vy);
UNUSED(nr);
UNUSED(nc);
UNUSED(nb);
UNUSED(ncols_interleaved);
UNUSED(blocklen);
#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_MATMUL_INT8) && ! ((defined(_MSC_VER)) && ! defined(__clang__))
if (ggml_sve_cnt_b == QK8_0) {
const void * b_ptr = vx;
const void * a_ptr = vy;
float * res_ptr = s;
size_t res_stride = bs * sizeof(float);
__asm__ __volatile__(
"mov x20, #0x4\n"
"mov x13, %x[nr]\n"
"mov z28.s, #-0x4\n"
"mov x12, #0x88\n"
"ptrue p1.b\n"
"whilelt p0.s, XZR, x20\n"
"cmp x13, #0x10\n"
"mul x12, %x[nb], x12\n"
"blt 4f\n"
"1:" // Row loop
"add x11, %x[b_ptr], #0x10\n"
"mov x10, %x[nc]\n"
"add x9, %x[res_ptr], %x[res_stride], LSL #4\n"
"2:" // Column loop
"add x28, %x[a_ptr], #0x8\n"
"mov z24.b, #0x0\n"
"mov z15.b, #0x0\n"
"mov x27, %x[nb]\n"
"add x26, x28, x12\n"
"mov z12.b, #0x0\n"
"mov z0.b, #0x0\n"
"add x25, x26, x12\n"
"mov z13.b, #0x0\n"
"mov z1.b, #0x0\n"
"add x24, x25, x12\n"
"mov z20.b, #0x0\n"
"mov z25.b, #0x0\n"
"mov z11.b, #0x0\n"
"mov z16.b, #0x0\n"
"mov z19.b, #0x0\n"
"mov z26.b, #0x0\n"
"mov z8.b, #0x0\n"
"mov z29.b, #0x0\n"
"mov z27.b, #0x0\n"
"mov z10.b, #0x0\n"
"3:" // Block loop
"ld1b { z30.b }, p1/Z, [x11]\n"
"ld1b { z21.b }, p1/Z, [x11, #1, MUL VL]\n"
"mov z18.s, #0x0\n"
"mov z7.s, #0x0\n"
"ld1rqb { z3.b }, p1/Z, [x28]\n"
"ld1rqb { z5.b }, p1/Z, [x28, #16]\n"
"mov z9.s, #0x0\n"
"mov z22.s, #0x0\n"
"ld1b { z4.b }, p1/Z, [x11, #2, MUL VL]\n"
"ld1b { z17.b }, p1/Z, [x11, #3, MUL VL]\n"
"sub x20, x11, #0x10\n"
"sub x23, x28, #0x8\n"
"lsl z31.b, z30.b, #0x4\n"
"lsl z6.b, z21.b, #0x4\n"
"ld1h { z23.s }, p1/Z, [x20]\n"
"sub x22, x26, #0x8\n"
"and z30.b, z30.b, #0xf0\n"
"and z21.b, z21.b, #0xf0\n"
"sub x21, x25, #0x8\n"
"sub x20, x24, #0x8\n"
"lsl z14.b, z4.b, #0x4\n"
"lsl z2.b, z17.b, #0x4\n"
"subs x27, x27, #0x1\n"
"add x11, x11, #0x90\n"
".inst 0x451f9872 // smmla z18.s, z3.b, z31.b\n"
".inst 0x45069867 // smmla z7.s, z3.b, z6.b\n"
"ld1rqb { z3.b }, p1/Z, [x28, #32]\n"
"and z4.b, z4.b, #0xf0\n"
".inst 0x451f98a9 // smmla z9.s, z5.b, z31.b\n"
".inst 0x450698b6 // smmla z22.s, z5.b, z6.b\n"
"ld1rqb { z5.b }, p1/Z, [x28, #48]\n"
"and z17.b, z17.b, #0xf0\n"
"fcvt z23.s, p1/m, z23.h\n"
".inst 0x450e9872 // smmla z18.s, z3.b, z14.b\n"
".inst 0x45029867 // smmla z7.s, z3.b, z2.b\n"
"ld1rqb { z3.b }, p1/Z, [x28, #64]\n"
".inst 0x450e98a9 // smmla z9.s, z5.b, z14.b\n"
".inst 0x450298b6 // smmla z22.s, z5.b, z2.b\n"
"ld1rqb { z5.b }, p1/Z, [x28, #80]\n"
"fscale z23.s, p1/m, z23.s, z28.s\n"
".inst 0x451e9872 // smmla z18.s, z3.b, z30.b\n"
".inst 0x45159867 // smmla z7.s, z3.b, z21.b\n"
"ld1rqb { z3.b }, p1/Z, [x28, #96]\n"
".inst 0x451e98a9 // smmla z9.s, z5.b, z30.b\n"
".inst 0x451598b6 // smmla z22.s, z5.b, z21.b\n"
"ld1rqb { z5.b }, p1/Z, [x28, #112]\n"
"add x28, x28, #0x88\n"
".inst 0x45049872 // smmla z18.s, z3.b, z4.b\n"
".inst 0x45119867 // smmla z7.s, z3.b, z17.b\n"
"ld1h { z3.s }, p0/Z, [x23]\n"
".inst 0x450498a9 // smmla z9.s, z5.b, z4.b\n"
".inst 0x451198b6 // smmla z22.s, z5.b, z17.b\n"
"fcvt z3.s, p1/m, z3.h\n"
"uzp1 z5.d, z18.d, z7.d\n"
"uzp2 z18.d, z18.d, z7.d\n"
"mov z3.q, z3.q[0]\n"
"uzp1 z7.d, z9.d, z22.d\n"
"uzp2 z22.d, z9.d, z22.d\n"
"fmul z9.s, z23.s, z3.s[0]\n"
"scvtf z5.s, p1/m, z5.s\n"
"scvtf z18.s, p1/m, z18.s\n"
"scvtf z7.s, p1/m, z7.s\n"
"scvtf z22.s, p1/m, z22.s\n"
"fmla z24.s, p1/M, z5.s, z9.s\n"
"ld1rqb { z5.b }, p1/Z, [x26]\n"
"fmul z9.s, z23.s, z3.s[1]\n"
"fmla z15.s, p1/M, z18.s, z9.s\n"
"ld1rqb { z18.b }, p1/Z, [x26, #16]\n"
"fmul z9.s, z23.s, z3.s[2]\n"
"fmul z3.s, z23.s, z3.s[3]\n"
"fmla z12.s, p1/M, z7.s, z9.s\n"
"mov z9.s, #0x0\n"
"ld1h { z7.s }, p0/Z, [x22]\n"
".inst 0x451f98a9 // smmla z9.s, z5.b, z31.b\n"
"fmla z0.s, p1/M, z22.s, z3.s\n"
"mov z22.s, #0x0\n"
"ld1h { z3.s }, p0/Z, [x21]\n"
".inst 0x450698b6 // smmla z22.s, z5.b, z6.b\n"
"ld1rqb { z5.b }, p1/Z, [x26, #32]\n"
"fcvt z7.s, p1/m, z7.h\n"
"fcvt z3.s, p1/m, z3.h\n"
".inst 0x450e98a9 // smmla z9.s, z5.b, z14.b\n"
".inst 0x450298b6 // smmla z22.s, z5.b, z2.b\n"
"ld1rqb { z5.b }, p1/Z, [x26, #64]\n"
"mov z7.q, z7.q[0]\n"
"mov z3.q, z3.q[0]\n"
".inst 0x451e98a9 // smmla z9.s, z5.b, z30.b\n"
".inst 0x451598b6 // smmla z22.s, z5.b, z21.b\n"
"ld1rqb { z5.b }, p1/Z, [x26, #96]\n"
".inst 0x450498a9 // smmla z9.s, z5.b, z4.b\n"
".inst 0x451198b6 // smmla z22.s, z5.b, z17.b\n"
"uzp1 z5.d, z9.d, z22.d\n"
"scvtf z5.s, p1/m, z5.s\n"
"uzp2 z22.d, z9.d, z22.d\n"
"fmul z9.s, z23.s, z7.s[0]\n"
"scvtf z22.s, p1/m, z22.s\n"
"fmla z13.s, p1/M, z5.s, z9.s\n"
"ld1rqb { z9.b }, p1/Z, [x25]\n"
"fmul z5.s, z23.s, z7.s[1]\n"
"fmla z1.s, p1/M, z22.s, z5.s\n"
"mov z5.s, #0x0\n"
"mov z22.s, #0x0\n"
".inst 0x451f9a45 // smmla z5.s, z18.b, z31.b\n"
".inst 0x45069a56 // smmla z22.s, z18.b, z6.b\n"
"ld1rqb { z18.b }, p1/Z, [x26, #48]\n"
".inst 0x450e9a45 // smmla z5.s, z18.b, z14.b\n"
".inst 0x45029a56 // smmla z22.s, z18.b, z2.b\n"
"ld1rqb { z18.b }, p1/Z, [x26, #80]\n"
".inst 0x451e9a45 // smmla z5.s, z18.b, z30.b\n"
".inst 0x45159a56 // smmla z22.s, z18.b, z21.b\n"
"ld1rqb { z18.b }, p1/Z, [x26, #112]\n"
"add x26, x26, #0x88\n"
".inst 0x45049a45 // smmla z5.s, z18.b, z4.b\n"
".inst 0x45119a56 // smmla z22.s, z18.b, z17.b\n"
"uzp1 z18.d, z5.d, z22.d\n"
"scvtf z18.s, p1/m, z18.s\n"
"uzp2 z22.d, z5.d, z22.d\n"
"fmul z5.s, z23.s, z7.s[2]\n"
"fmul z7.s, z23.s, z7.s[3]\n"
"scvtf z22.s, p1/m, z22.s\n"
"fmla z20.s, p1/M, z18.s, z5.s\n"
"ld1rqb { z18.b }, p1/Z, [x25, #16]\n"
"ld1h { z5.s }, p0/Z, [x20]\n"
"fcvt z5.s, p1/m, z5.h\n"
"fmla z25.s, p1/M, z22.s, z7.s\n"
"mov z22.s, #0x0\n"
"mov z7.s, #0x0\n"
".inst 0x451f9936 // smmla z22.s, z9.b, z31.b\n"
".inst 0x45069927 // smmla z7.s, z9.b, z6.b\n"
"ld1rqb { z9.b }, p1/Z, [x25, #32]\n"
"mov z5.q, z5.q[0]\n"
".inst 0x450e9936 // smmla z22.s, z9.b, z14.b\n"
".inst 0x45029927 // smmla z7.s, z9.b, z2.b\n"
"ld1rqb { z9.b }, p1/Z, [x25, #64]\n"
".inst 0x451e9936 // smmla z22.s, z9.b, z30.b\n"
".inst 0x45159927 // smmla z7.s, z9.b, z21.b\n"
"ld1rqb { z9.b }, p1/Z, [x25, #96]\n"
".inst 0x45049936 // smmla z22.s, z9.b, z4.b\n"
".inst 0x45119927 // smmla z7.s, z9.b, z17.b\n"
"uzp1 z9.d, z22.d, z7.d\n"
"scvtf z9.s, p1/m, z9.s\n"
"uzp2 z22.d, z22.d, z7.d\n"
"fmul z7.s, z23.s, z3.s[0]\n"
"scvtf z22.s, p1/m, z22.s\n"
"fmla z11.s, p1/M, z9.s, z7.s\n"
"ld1rqb { z9.b }, p1/Z, [x24]\n"
"fmul z7.s, z23.s, z3.s[1]\n"
"fmla z16.s, p1/M, z22.s, z7.s\n"
"mov z22.s, #0x0\n"
"mov z7.s, #0x0\n"
".inst 0x451f9a56 // smmla z22.s, z18.b, z31.b\n"
".inst 0x45069a47 // smmla z7.s, z18.b, z6.b\n"
"ld1rqb { z18.b }, p1/Z, [x25, #48]\n"
".inst 0x450e9a56 // smmla z22.s, z18.b, z14.b\n"
".inst 0x45029a47 // smmla z7.s, z18.b, z2.b\n"
"ld1rqb { z18.b }, p1/Z, [x25, #80]\n"
".inst 0x451e9a56 // smmla z22.s, z18.b, z30.b\n"
".inst 0x45159a47 // smmla z7.s, z18.b, z21.b\n"
"ld1rqb { z18.b }, p1/Z, [x25, #112]\n"
"add x25, x25, #0x88\n"
".inst 0x45049a56 // smmla z22.s, z18.b, z4.b\n"
".inst 0x45119a47 // smmla z7.s, z18.b, z17.b\n"
"uzp1 z18.d, z22.d, z7.d\n"
"scvtf z18.s, p1/m, z18.s\n"
"uzp2 z7.d, z22.d, z7.d\n"
"fmul z22.s, z23.s, z3.s[2]\n"
"fmul z3.s, z23.s, z3.s[3]\n"
"scvtf z7.s, p1/m, z7.s\n"
"fmla z19.s, p1/M, z18.s, z22.s\n"
"ld1rqb { z18.b }, p1/Z, [x24, #16]\n"
"fmul z22.s, z23.s, z5.s[0]\n"
"fmla z26.s, p1/M, z7.s, z3.s\n"
"mov z3.s, #0x0\n"
"mov z7.s, #0x0\n"
".inst 0x451f9923 // smmla z3.s, z9.b, z31.b\n"
".inst 0x45069927 // smmla z7.s, z9.b, z6.b\n"
"ld1rqb { z9.b }, p1/Z, [x24, #32]\n"
".inst 0x450e9923 // smmla z3.s, z9.b, z14.b\n"
".inst 0x45029927 // smmla z7.s, z9.b, z2.b\n"
"mov z9.s, #0x0\n"
".inst 0x451f9a49 // smmla z9.s, z18.b, z31.b\n"
"mov z31.s, #0x0\n"
".inst 0x45069a5f // smmla z31.s, z18.b, z6.b\n"
"ld1rqb { z6.b }, p1/Z, [x24, #48]\n"
"ld1rqb { z18.b }, p1/Z, [x24, #64]\n"
".inst 0x450e98c9 // smmla z9.s, z6.b, z14.b\n"
"fmul z14.s, z23.s, z5.s[1]\n"
".inst 0x450298df // smmla z31.s, z6.b, z2.b\n"
"ld1rqb { z6.b }, p1/Z, [x24, #80]\n"
"fmul z2.s, z23.s, z5.s[2]\n"
"fmul z23.s, z23.s, z5.s[3]\n"
".inst 0x451e9a43 // smmla z3.s, z18.b, z30.b\n"
".inst 0x45159a47 // smmla z7.s, z18.b, z21.b\n"
"ld1rqb { z5.b }, p1/Z, [x24, #96]\n"
".inst 0x451e98c9 // smmla z9.s, z6.b, z30.b\n"
".inst 0x451598df // smmla z31.s, z6.b, z21.b\n"
"ld1rqb { z18.b }, p1/Z, [x24, #112]\n"
"add x24, x24, #0x88\n"
".inst 0x450498a3 // smmla z3.s, z5.b, z4.b\n"
".inst 0x451198a7 // smmla z7.s, z5.b, z17.b\n"
".inst 0x45049a49 // smmla z9.s, z18.b, z4.b\n"
".inst 0x45119a5f // smmla z31.s, z18.b, z17.b\n"
"uzp1 z18.d, z3.d, z7.d\n"
"uzp2 z5.d, z3.d, z7.d\n"
"scvtf z18.s, p1/m, z18.s\n"
"uzp1 z6.d, z9.d, z31.d\n"
"uzp2 z9.d, z9.d, z31.d\n"
"scvtf z5.s, p1/m, z5.s\n"
"fmla z8.s, p1/M, z18.s, z22.s\n"
"scvtf z6.s, p1/m, z6.s\n"
"scvtf z9.s, p1/m, z9.s\n"
"fmla z29.s, p1/M, z5.s, z14.s\n"
"fmla z27.s, p1/M, z6.s, z2.s\n"
"fmla z10.s, p1/M, z9.s, z23.s\n"
"bgt 3b\n"
"mov x20, %x[res_ptr]\n"
"subs x10, x10, #0x8\n"
"add %x[res_ptr], %x[res_ptr], #0x20\n"
"st1w { z24.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z15.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z12.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z0.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z13.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z1.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z20.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z25.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z11.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z16.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z19.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z26.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z8.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z29.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z27.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"st1w { z10.s }, p1, [x20]\n"
"bne 2b\n"
"mov x20, #0x4\n"
"sub x13, x13, #0x10\n"
"cmp x13, #0x10\n"
"mov %x[res_ptr], x9\n"
"madd %x[a_ptr], x20, x12, %x[a_ptr]\n"
"bge 1b\n"
"4:" // Row loop skip
"cbz x13, 9f\n"
"5:" // Row tail: Row loop
"add x25, %x[b_ptr], #0x10\n"
"mov x24, %x[nc]\n"
"add x23, %x[res_ptr], %x[res_stride], LSL #2\n"
"6:" // Row tail: Column loop
"mov z24.b, #0x0\n"
"mov z15.b, #0x0\n"
"add x28, %x[a_ptr], #0x8\n"
"mov x22, %x[nb]\n"
"mov z12.b, #0x0\n"
"mov z0.b, #0x0\n"
"7:" // Row tail: Block loop
"ld1b { z3.b }, p1/Z, [x25]\n"
"ld1b { z6.b }, p1/Z, [x25, #1, MUL VL]\n"
"mov z2.s, #0x0\n"
"mov z25.s, #0x0\n"
"ld1rqb { z26.b }, p1/Z, [x28]\n"
"ld1rqb { z21.b }, p1/Z, [x28, #16]\n"
"mov z27.s, #0x0\n"
"mov z19.s, #0x0\n"
"ld1b { z29.b }, p1/Z, [x25, #2, MUL VL]\n"
"ld1b { z16.b }, p1/Z, [x25, #3, MUL VL]\n"
"sub x21, x25, #0x10\n"
"sub x20, x28, #0x8\n"
"lsl z20.b, z3.b, #0x4\n"
"lsl z4.b, z6.b, #0x4\n"
"ld1rqb { z10.b }, p1/Z, [x28, #32]\n"
"ld1rqb { z23.b }, p1/Z, [x28, #48]\n"
"and z3.b, z3.b, #0xf0\n"
"and z6.b, z6.b, #0xf0\n"
"ld1rqb { z11.b }, p1/Z, [x28, #64]\n"
"ld1rqb { z7.b }, p1/Z, [x28, #80]\n"
"lsl z8.b, z29.b, #0x4\n"
"lsl z14.b, z16.b, #0x4\n"
"ld1rqb { z18.b }, p1/Z, [x28, #96]\n"
"ld1rqb { z30.b }, p1/Z, [x28, #112]\n"
".inst 0x45149b42 // smmla z2.s, z26.b, z20.b\n"
".inst 0x45049b59 // smmla z25.s, z26.b, z4.b\n"
"and z29.b, z29.b, #0xf0\n"
"ld1h { z17.s }, p1/Z, [x21]\n"
".inst 0x45149abb // smmla z27.s, z21.b, z20.b\n"
".inst 0x45049ab3 // smmla z19.s, z21.b, z4.b\n"
"and z16.b, z16.b, #0xf0\n"
"ld1h { z4.s }, p0/Z, [x20]\n"
"subs x22, x22, #0x1\n"
"add x28, x28, #0x88\n"
"fcvt z17.s, p1/m, z17.h\n"
"add x25, x25, #0x90\n"
".inst 0x45089942 // smmla z2.s, z10.b, z8.b\n"
".inst 0x450e9959 // smmla z25.s, z10.b, z14.b\n"
"fcvt z4.s, p1/m, z4.h\n"
".inst 0x45089afb // smmla z27.s, z23.b, z8.b\n"
".inst 0x450e9af3 // smmla z19.s, z23.b, z14.b\n"
"fscale z17.s, p1/m, z17.s, z28.s\n"
"mov z4.q, z4.q[0]\n"
".inst 0x45039962 // smmla z2.s, z11.b, z3.b\n"
".inst 0x45069979 // smmla z25.s, z11.b, z6.b\n"
"fmul z23.s, z17.s, z4.s[0]\n"
"fmul z9.s, z17.s, z4.s[1]\n"
"fmul z21.s, z17.s, z4.s[2]\n"
"fmul z4.s, z17.s, z4.s[3]\n"
".inst 0x450398fb // smmla z27.s, z7.b, z3.b\n"
".inst 0x450698f3 // smmla z19.s, z7.b, z6.b\n"
".inst 0x451d9a42 // smmla z2.s, z18.b, z29.b\n"
".inst 0x45109a59 // smmla z25.s, z18.b, z16.b\n"
".inst 0x451d9bdb // smmla z27.s, z30.b, z29.b\n"
".inst 0x45109bd3 // smmla z19.s, z30.b, z16.b\n"
"uzp1 z31.d, z2.d, z25.d\n"
"uzp2 z13.d, z2.d, z25.d\n"
"scvtf z31.s, p1/m, z31.s\n"
"uzp1 z17.d, z27.d, z19.d\n"
"uzp2 z18.d, z27.d, z19.d\n"
"scvtf z13.s, p1/m, z13.s\n"
"fmla z24.s, p1/M, z31.s, z23.s\n"
"scvtf z17.s, p1/m, z17.s\n"
"scvtf z18.s, p1/m, z18.s\n"
"fmla z15.s, p1/M, z13.s, z9.s\n"
"fmla z12.s, p1/M, z17.s, z21.s\n"
"fmla z0.s, p1/M, z18.s, z4.s\n"
"bgt 7b\n"
"mov x20, %x[res_ptr]\n"
"cmp x13, #0x1\n"
"st1w { z24.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"cmp x13, #0x2\n"
"st1w { z15.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"cmp x13, #0x3\n"
"st1w { z12.s }, p1, [x20]\n"
"add x20, x20, %x[res_stride]\n"
"ble 8f\n"
"st1w { z0.s }, p1, [x20]\n"
"8:" // Row tail: Accumulator store skip
"subs x24, x24, #0x8\n"
"add %x[res_ptr], %x[res_ptr], #0x20\n"
"bne 6b\n"
"subs x13, x13, #0x4\n"
"add %x[a_ptr], %x[a_ptr], x12\n"
"mov %x[res_ptr], x23\n"
"bgt 5b\n"
"9:" // Row tail: Row loop skip
: [a_ptr] "+&r" (a_ptr), [res_ptr] "+&r" (res_ptr)
: [b_ptr] "r" (b_ptr), [nr] "r" (nr), [nb] "r" (nb), [res_stride] "r" (res_stride), [nc] "r" (nc)
: "cc", "memory", "p0", "p1", "x9", "x10", "x11", "x12", "x13", "x20", "x21", "x22", "x23", "x24", "x25", "x26", "x27", "x28", "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z8", "z9", "z10", "z11", "z12", "z13", "z14", "z15", "z16", "z17", "z18", "z19", "z20", "z21", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30", "z31"
);
return;
}
else if (ggml_cpu_has_neon() && ggml_cpu_has_matmul_int8()) {
GGML_ASSERT((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) &&
"__ARM_FEATURE_SVE for vector size of 256-bits not defined, use the Q4_0_4_8 quantization format for optimal "
"performance");
}
else if (ggml_cpu_has_neon()) {
GGML_ASSERT(((ggml_cpu_has_sve() && (ggml_sve_cnt_b == QK8_0)) || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE for vector size of 256-bits and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 "
"quantization format for optimal performance");
}
#endif
#if defined(__ARM_NEON) && defined(__ARM_FEATURE_MATMUL_INT8)
GGML_ASSERT(ggml_cpu_has_sve() &&
"__ARM_FEATURE_SVE not defined, use the Q4_0_4_8 quantization format for optimal performance");
#elif defined(__ARM_NEON) && defined(__aarch64__)
GGML_ASSERT((ggml_cpu_has_sve() || ggml_cpu_has_matmul_int8()) &&
"__ARM_FEATURE_SVE and __ARM_FEATURE_MATMUL_INT8 not defined, use the Q4_0_4_4 quantization format for optimal "
"performance");
#else
float sumf[4][8];
int sumi;
for (int y = 0; y < nr / 4; y++) {
const block_q8_0x4 * a_ptr = (const block_q8_0x4 *) vy + (y * nb);
for (int x = 0; x < nc / ncols_interleaved; x++) {
const block_q4_0x8 * b_ptr = (const block_q4_0x8 *) vx + (x * nb);
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) sumf[m][j] = 0.0;
}
for (int l = 0; l < nb; l++) {
for (int k = 0; k < (qk / (2 * blocklen)); k++) {
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++) {
sumi = 0;
for (int i = 0; i < blocklen; ++i) {
const int v0 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] << 4);
const int v1 = (int8_t) (b_ptr[l].qs[k * ncols_interleaved * blocklen + j * blocklen + i] & 0xF0);
sumi += ((v0 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i]) +
(v1 * a_ptr[l].qs[k * 4 * blocklen + m * blocklen + i + qk / 2 * 4])) >> 4;
}
sumf[m][j] += sumi * GGML_FP16_TO_FP32(b_ptr[l].d[j]) * GGML_FP16_TO_FP32(a_ptr[l].d[m]);
}
}
}
}
for (int m = 0; m < 4; m++) {
for (int j = 0; j < ncols_interleaved; j++)
s[(y * 4 + m) * bs + x * ncols_interleaved + j] = sumf[m][j];
}
}
}
#endif
}
|