Spaces:
Build error
Build error
File size: 101,420 Bytes
5a29263 |
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 2182 2183 2184 2185 2186 2187 2188 2189 2190 2191 2192 2193 2194 2195 2196 2197 2198 2199 2200 2201 2202 2203 2204 2205 2206 2207 2208 2209 2210 2211 2212 2213 2214 2215 2216 2217 2218 2219 2220 2221 2222 2223 2224 2225 2226 2227 2228 2229 2230 2231 2232 2233 2234 2235 2236 2237 2238 2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258 2259 2260 2261 2262 2263 2264 2265 2266 2267 2268 2269 2270 2271 2272 2273 2274 2275 2276 2277 2278 2279 2280 2281 2282 2283 2284 2285 2286 2287 2288 2289 2290 2291 2292 2293 2294 2295 2296 2297 2298 2299 2300 2301 2302 2303 2304 2305 2306 2307 2308 2309 2310 2311 2312 2313 2314 2315 2316 2317 2318 2319 2320 2321 2322 2323 2324 2325 2326 2327 2328 2329 2330 2331 2332 2333 2334 2335 2336 2337 2338 2339 2340 2341 2342 2343 2344 2345 2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361 2362 2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381 2382 2383 2384 2385 2386 2387 2388 2389 2390 2391 2392 2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406 2407 2408 2409 2410 2411 2412 2413 2414 2415 2416 2417 2418 2419 2420 2421 2422 2423 2424 2425 2426 2427 2428 2429 2430 2431 2432 2433 2434 2435 2436 2437 2438 2439 2440 2441 2442 2443 2444 2445 2446 2447 2448 2449 2450 2451 2452 2453 2454 2455 2456 2457 2458 2459 2460 2461 2462 2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474 2475 2476 2477 2478 2479 2480 2481 2482 2483 2484 2485 2486 2487 2488 2489 2490 2491 2492 2493 2494 2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593 2594 2595 2596 2597 2598 |
// Copyright 2024 Mozilla Foundation
//
// Permission is hereby granted, free of charge, to any person obtaining
// a copy of this software and associated documentation files (the
// "Software"), to deal in the Software without restriction, including
// without limitation the rights to use, copy, modify, merge, publish,
// distribute, sublicense, and/or sell copies of the Software, and to
// permit persons to whom the Software is furnished to do so, subject to
// the following conditions:
//
// The above copyright notice and this permission notice shall be
// included in all copies or substantial portions of the Software.
//
// THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
// EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF
// MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
// NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS
// BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
// ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN
// CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
// SOFTWARE.
//
// _ _ ___ _ _ ___
// | |_(_)_ _ _ _| _ ) | /_\ / __|
// | _| | ' \ || | _ \ |__ / _ \\__ \.
// \__|_|_||_\_, |___/____/_/ \_\___/
// |__/
//
// BASIC LINEAR ALGEBRA SUBPROGRAMS
//
//
// This file implements multithreaded CPU matrix multiplication for the
// common contiguous use case C = Aᵀ * B. These kernels are designed to
// have excellent performance[1] for matrices that fit in the CPU cache
// without imposing any overhead such as cache filling or malloc calls.
//
// This implementation does not guarantee any upper bound with rounding
// errors, which grow along with k. Our goal's to maximally exploit the
// hardware for performance, and then use whatever resources remain for
// improving numerical accuracy.
//
// [1] J. Tunney, ‘LLaMA Now Goes Faster on CPUs’, Mar. 2024. [Online].
// Available: https://justine.lol/matmul/. [Accessed: 29-Mar-2024].
#if defined(__GNUC__)
#pragma GCC diagnostic ignored "-Wpedantic"
#pragma GCC diagnostic ignored "-Wignored-attributes"
#endif
#include "sgemm.h"
#include "ggml-impl.h"
#include "ggml-cpu-impl.h"
#include "ggml-quants.h"
#include <atomic>
#include <array>
#ifdef _MSC_VER
#define NOINLINE __declspec(noinline)
#else
#define NOINLINE __attribute__((__noinline__))
#endif
#if defined(__ARM_NEON) || defined(__AVX512F__)
#define VECTOR_REGISTERS 32
#else
#define VECTOR_REGISTERS 16
#endif
#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
namespace {
inline float unhalf(ggml_fp16_t d) {
return GGML_FP16_TO_FP32(d);
}
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED ARITHMETIC OPERATIONS
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
inline __m128 add(__m128 x, __m128 y) { return _mm_add_ps(x, y); }
inline __m128 sub(__m128 x, __m128 y) { return _mm_sub_ps(x, y); }
inline __m128 mul(__m128 x, __m128 y) { return _mm_mul_ps(x, y); }
#endif // __SSE__
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
inline __m256 add(__m256 x, __m256 y) { return _mm256_add_ps(x, y); }
inline __m256 sub(__m256 x, __m256 y) { return _mm256_sub_ps(x, y); }
inline __m256 mul(__m256 x, __m256 y) { return _mm256_mul_ps(x, y); }
#endif // __AVX__
#if defined(__AVX512F__)
inline __m512 add(__m512 x, __m512 y) { return _mm512_add_ps(x, y); }
inline __m512 sub(__m512 x, __m512 y) { return _mm512_sub_ps(x, y); }
inline __m512 mul(__m512 x, __m512 y) { return _mm512_mul_ps(x, y); }
#endif // __AVX512F__
#if defined(__ARM_NEON)
inline float32x4_t add(float32x4_t x, float32x4_t y) { return vaddq_f32(x, y); }
inline float32x4_t sub(float32x4_t x, float32x4_t y) { return vsubq_f32(x, y); }
inline float32x4_t mul(float32x4_t x, float32x4_t y) { return vmulq_f32(x, y); }
#endif // __ARM_NEON
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
inline float16x8_t add(float16x8_t x, float16x8_t y) { return vaddq_f16(x, y); }
inline float16x8_t sub(float16x8_t x, float16x8_t y) { return vsubq_f16(x, y); }
inline float16x8_t mul(float16x8_t x, float16x8_t y) { return vmulq_f16(x, y); }
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(__MMA__)
typedef vector unsigned char vec_t;
typedef __vector_quad acc_t;
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED FUSED MULTIPLY ADD
/**
* Computes a * b + c.
*/
template <typename T, typename U>
inline U madd(T a, T b, U c) {
return add(mul(a, b), c);
}
#if defined(__FMA__)
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
template <>
inline __m256 madd(__m256 a, __m256 b, __m256 c) {
return _mm256_fmadd_ps(a, b, c);
}
#endif
#if defined(__AVX512F__)
template <>
inline __m512 madd(__m512 a, __m512 b, __m512 c) {
return _mm512_fmadd_ps(a, b, c);
}
#endif
#if defined(__AVX512BF16__)
template <>
inline __m512 madd(__m512bh a, __m512bh b, __m512 c) {
return _mm512_dpbf16_ps(c, a, b);
}
template <>
inline __m256 madd(__m256bh a, __m256bh b, __m256 c) {
return _mm256_dpbf16_ps(c, a, b);
}
#endif
#endif
#if defined(__ARM_FEATURE_FMA)
template <>
inline float32x4_t madd(float32x4_t a, float32x4_t b, float32x4_t c) {
return vfmaq_f32(c, b, a);
}
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
template <>
inline float16x8_t madd(float16x8_t a, float16x8_t b, float16x8_t c) {
return vfmaq_f16(c, b, a);
}
#endif
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED HORIZONTAL SUM
#if defined(__ARM_NEON)
inline float hsum(float32x4_t x) {
return vaddvq_f32(x);
}
#endif // __ARM_NEON
#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
inline float hsum(float16x8_t x) {
return vaddvq_f32(vaddq_f32(vcvt_f32_f16(vget_low_f16(x)),
vcvt_f32_f16(vget_high_f16(x))));
}
#endif // __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
inline float hsum(__m128 x) {
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
x = _mm_add_ps(x, _mm_movehl_ps(x, x));
x = _mm_add_ss(x, _mm_movehdup_ps(x));
#else
__m128 t;
t = _mm_shuffle_ps(x, x, _MM_SHUFFLE(2, 3, 0, 1));
x = _mm_add_ps(x, t);
t = _mm_movehl_ps(t, x);
x = _mm_add_ss(x, t);
#endif
return _mm_cvtss_f32(x);
}
#endif
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
inline float hsum(__m256 x) {
return hsum(_mm_add_ps(_mm256_extractf128_ps(x, 1),
_mm256_castps256_ps128(x)));
}
#endif // __AVX__
#if defined(__AVX512F__)
inline float hsum(__m512 x) {
return _mm512_reduce_add_ps(x);
}
#endif // __AVX512F__
////////////////////////////////////////////////////////////////////////////////////////////////////
// VECTORIZED MEMORY LOADING
template <typename T, typename U> T load(const U *);
#if defined(__ARM_NEON)
template <> inline float32x4_t load(const float *p) {
return vld1q_f32(p);
}
#if !defined(_MSC_VER)
// FIXME: this should check for __ARM_FEATURE_FP16_VECTOR_ARITHMETIC
template <> inline float16x8_t load(const ggml_fp16_t *p) {
return vld1q_f16((const float16_t *)p);
}
template <> inline float32x4_t load(const ggml_fp16_t *p) {
return vcvt_f32_f16(vld1_f16((const float16_t *)p));
}
#endif // _MSC_VER
#endif // __ARM_NEON
#if defined(__SSE__) || defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
template <> inline __m128 load(const float *p) {
return _mm_loadu_ps(p);
}
#endif // __SSE__
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
template <> inline __m256 load(const float *p) {
return _mm256_loadu_ps(p);
}
#endif // __AVX__
#if defined(__AVX2__) || defined(__AVX512F__)
template <> inline __m256 load(const ggml_bf16_t *p) {
return _mm256_castsi256_ps(
_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)p)), 16));
}
#endif // __AVX2__
#if defined(__F16C__)
template <> inline __m256 load(const ggml_fp16_t *p) {
return _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)p));
}
#endif // __F16C__
#if defined(__AVX512F__)
template <> inline __m512 load(const float *p) {
return _mm512_loadu_ps(p);
}
template <> inline __m512 load(const ggml_fp16_t *p) {
return _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)p));
}
template <> inline __m512 load(const ggml_bf16_t *p) {
return _mm512_castsi512_ps(
_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)p)), 16));
}
#endif // __AVX512F__
#if defined(__AVX512BF16__)
template <> inline __m512bh load(const ggml_bf16_t *p) {
return (__m512bh)_mm512_loadu_ps((const float *)p);
}
template <> inline __m256bh load(const ggml_bf16_t *p) {
return (__m256bh)_mm256_loadu_ps((const float *)p);
}
template <> inline __m512bh load(const float *p) {
return _mm512_cvtne2ps_pbh(_mm512_loadu_ps(p + 16), _mm512_loadu_ps(p));
}
template <> inline __m256bh load(const float *p) {
return _mm512_cvtneps_pbh(_mm512_loadu_ps(p));
}
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// CONSTANTS
#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
static const int8_t kvalues_iq4nl[16] = {-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113};
static const __m128i iq4nlt = _mm_loadu_si128((const __m128i *) kvalues_iq4nl);
#endif
////////////////////////////////////////////////////////////////////////////////////////////////////
// FLOATING POINT MATRIX MULTIPLICATION
template <int M>
static inline int64_t BLOCK_SIZE(size_t m) {
const int64_t NB_BLOC_M = (m + M - 1) / M;
return (m % NB_BLOC_M == 0) ? m / NB_BLOC_M : (m / NB_BLOC_M) + 1;
}
static constexpr inline int64_t BLOC_POS(int64_t ib, int64_t ibN, int64_t bloc_size) {
return ib < ibN ? ib * bloc_size : ibN * bloc_size + (ib - ibN) * (bloc_size - 1);
}
template <int KN, typename D, typename V, typename TA, typename TB, typename TC>
class tinyBLAS {
public:
tinyBLAS(const ggml_compute_params * params, int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc)
: params(params), A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc) {
}
bool matmul(int64_t m, int64_t n) {
if (k % KN != 0)
return false;
// compute RM for only need tile with size RM&RM-1
#if VECTOR_REGISTERS == 32
if (m % 16 == 0 && (m/16 >= params->nth)) {
const int64_t SIZE_N = BLOCK_SIZE<6>(n);
mnpack<4, 6, 4>(m, n, SIZE_N, 12);
return true;
}
if (m % 8 == 0 ) {
const int64_t SIZE_N = BLOCK_SIZE<6>(n);
mnpack<4, 6, 2>(m, n, SIZE_N, 12);
return true;
}
if (m % 4 == 0) {
const int64_t SIZE_N = BLOCK_SIZE<6>(n);
mnpack<4, 6, 1>(m, n, SIZE_N, 12);
return true;
}
#else // VECTOR_REGISTERS == 16
if (m % 16 == 0 && (m/16 >= params->nth)) {
const int64_t SIZE_N = BLOCK_SIZE<3>(n);
mnpack<4, 3, 4>(m, n, SIZE_N, 24);
return true;
}
if (m % 8 == 0 ) {
const int64_t SIZE_N = BLOCK_SIZE<3>(n);
mnpack<4, 3, 2>(m, n, SIZE_N, 24);
return true;
}
if (m % 4 == 0) {
const int64_t SIZE_N = BLOCK_SIZE<3>(n);
mnpack<4, 3, 1>(m, n, SIZE_N, 24);
return true;
}
#endif
return false;
}
private:
template <int RM, int RN, int BM>
inline void mnpack(int64_t m, int64_t n, int64_t SIZE_N, int64_t BN) {
if (SIZE_N == RN) {
return gemm<RM, RN, BM>(m, n, BN);
}
if constexpr (RN > 1) {
return mnpack<RM, RN-1, BM>(m, n, SIZE_N, BN);
} else {
GGML_LOG_ERROR("mnpack<%d, %d> bloc size not supported\n", RM, (int)SIZE_N);
GGML_ASSERT(false); // we have miss something.
}
}
template <int RM, int RN>
inline void gemm_bloc(int64_t ii, int64_t jj) {
D Cv[RN][RM] = {};
for (int64_t l = 0; l < k; l += KN) {
// help compiler for op order.
if constexpr (RM <= RN) {
V Av[RM];
for (int64_t i = 0; i < RM; ++i) {
Av[i] = load<V>(A + lda * (ii + i) + l);
}
for (int64_t j = 0; j < RN; ++j) {
V Bv = load<V>(B + ldb * (jj + j) + l);
for (int64_t i = 0; i < RM; ++i) {
Cv[j][i] = madd(Av[i], Bv, Cv[j][i]);
}
}
} else {
V Bv[RN];
for (int64_t j = 0; j < RN; ++j) {
Bv[j] = load<V>(B + ldb * (jj + j) + l);
}
for (int64_t i = 0; i < RM; ++i) {
V Av = load<V>(A + lda * (ii + i) + l);
for (int64_t j = 0; j < RN; ++j) {
Cv[j][i] = madd(Av, Bv[j], Cv[j][i]);
}
}
}
}
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
template <int RM, int RN, int BM>
NOINLINE void gemm(int64_t m, int64_t n, int64_t BN) {
static std::atomic<int64_t> current_chunk;
GGML_ASSERT(m % (RM * BM) == 0);
const int64_t ytiles = m / (RM * BM);
const int64_t xtiles = (n + RN -1) / RN;
const int64_t jj_RN = (xtiles - (xtiles * RN - n));
// "round" bloc_size to "nearest" BN
const int64_t NB_BN = xtiles < BN ? 1 : (xtiles + BN / 2) / BN;
const int64_t SIZE_BN = xtiles % NB_BN == 0 ? xtiles / NB_BN : xtiles / NB_BN + 1;
const int64_t jj_BN = (NB_BN - (NB_BN * SIZE_BN - xtiles));
const int64_t nb_job = ytiles * NB_BN;
if (params->ith == 0) {
GGML_ASSERT( jj_BN * SIZE_BN + (NB_BN - jj_BN) * (SIZE_BN - 1) == xtiles);
// Every thread starts at ith, so the first unprocessed chunk is nth. This save a bit of coordination right at the start.
std::atomic_store_explicit(¤t_chunk, (int64_t)params->nth, std::memory_order_relaxed);
}
ggml_barrier(params->threadpool);
int64_t job = params->ith;
while (job < nb_job) {
const int64_t ii = (job % ytiles) * RM * BM;
const int64_t jb = job / ytiles;
const int64_t jr0 = BLOC_POS(jb , jj_BN, SIZE_BN);
const int64_t jrN = BLOC_POS(jb+1, jj_BN, SIZE_BN);
const int64_t jj0 = BLOC_POS(jr0, jj_RN, RN);
const int64_t jj2 = BLOC_POS(jrN, jj_RN, RN);
const int64_t jj1 = jj2 < jj_RN * RN ? jj2 : jj_RN * RN;
for (int64_t bi = 0; bi < BM * RM; bi += RM) {
int64_t jj = jj0;
for (; jj < jj1; jj += RN) {
gemm_bloc<RM, RN>(ii + bi, jj);
}
if constexpr (RN > 1) {
for (; jj < jj2; jj += RN - 1) {
gemm_bloc<RM, RN-1>(ii + bi, jj);
}
}
GGML_ASSERT(jj == jj2);
}
// next step.
job = std::atomic_fetch_add_explicit(¤t_chunk, (int64_t)1, std::memory_order_relaxed);
}
ggml_barrier(params->threadpool);
return;
}
const ggml_compute_params * params;
const TA *const A;
const TB *const B;
TC *const C;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
};
//////////////////////////////////////////////////////////////////////////////////////////
// QUANT ZERO MATRIX MULTIPLICATION
#if defined(__ARM_FEATURE_DOTPROD)
template <typename TA>
class tinyBLAS_Q0_ARM {
public:
tinyBLAS_Q0_ARM(int64_t k,
const TA *A, int64_t lda,
const block_q8_0 *B, int64_t ldb,
float *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
NOINLINE void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
switch ((MIN(m - m0, 3) << 4) | MIN(n - n0, 3ll)) {
case 0x33:
mc = 3;
nc = 3;
gemm<3, 3>(m0, m, n0, n);
break;
case 0x32:
mc = 3;
nc = 2;
gemm<3, 2>(m0, m, n0, n);
break;
case 0x23:
mc = 2;
nc = 3;
gemm<2, 3>(m0, m, n0, n);
break;
case 0x22:
mc = 2;
nc = 2;
gemm<2, 2>(m0, m, n0, n);
break;
case 0x31:
mc = 3;
nc = 1;
gemm<3, 1>(m0, m, n0, n);
break;
case 0x13:
mc = 1;
nc = 3;
gemm<1, 3>(m0, m, n0, n);
break;
case 0x21:
mc = 2;
nc = 1;
gemm<2, 1>(m0, m, n0, n);
break;
case 0x12:
mc = 1;
nc = 2;
gemm<1, 2>(m0, m, n0, n);
break;
case 0x11:
mc = 1;
nc = 1;
gemm<1, 1>(m0, m, n0, n);
break;
default:
return;
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
float32x4_t Cv[RN][RM] = {};
for (int64_t l = 0; l < k; ++l)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
Cv[j][i] = vmlaq_n_f32(Cv[j][i],
vcvtq_f32_s32(vdotq_s32(
vdotq_s32(vdupq_n_s32(0),
load_lo(A + lda * (ii + i) + l),
load_lo(B + ldb * (jj + j) + l)),
load_hi(A + lda * (ii + i) + l),
load_hi(B + ldb * (jj + j) + l))),
unhalf(A[lda * (ii + i) + l].d) *
unhalf(B[ldb * (jj + j) + l].d));
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
inline int8x16_t load_lo(const block_q8_0 *b) {
return vld1q_s8(b->qs);
}
inline int8x16_t load_hi(const block_q8_0 *b) {
return vld1q_s8(b->qs + 16);
}
inline int8x16_t load_lo(const block_q4_0 *b) {
return vsubq_s8(vreinterpretq_s8_u8(vandq_u8(vld1q_u8(b->qs),
vdupq_n_u8(0x0f))),
vdupq_n_s8(0x8));
}
inline int8x16_t load_hi(const block_q4_0 *b) {
return vsubq_s8(vreinterpretq_s8_u8(vshrq_n_u8(vld1q_u8(b->qs), 4)),
vdupq_n_s8(0x8));
}
const TA *const A;
const block_q8_0 *const B;
float *const C;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
#endif // __ARM_FEATURE_DOTPROD
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
template <typename TA, typename TB, typename TC>
class tinyBLAS_Q0_AVX {
public:
tinyBLAS_Q0_AVX(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
switch ((MIN(m - m0, 4) << 4) | MIN(n - n0, 4)) {
#if VECTOR_REGISTERS == 32
case 0x44:
mc = 4;
nc = 4;
#if defined(__AVX2__) && defined(__F16C__)
gemm4xN<4>(m0, m, n0, n);
#else
gemm<4, 4>(m0, m, n0, n);
#endif
break;
case 0x43:
mc = 4;
nc = 3;
#if defined(__AVX2__) && defined(__F16C__)
gemm4xN<3>(m0, m, n0, n);
#else
gemm<4, 3>(m0, m, n0, n);
#endif
break;
case 0x34:
mc = 3;
nc = 4;
#if defined(__AVX2__) && defined(__F16C__)
gemmMx4<3>(m0, m, n0, n);
#else
gemm<3, 4>(m0, m, n0, n);
#endif
break;
case 0x33:
mc = 3;
nc = 3;
gemm<3, 3>(m0, m, n0, n);
break;
case 0x42:
mc = 4;
nc = 2;
#if defined(__AVX2__) && defined(__F16C__)
gemm4xN<2>(m0, m, n0, n);
#else
gemm<4, 2>(m0, m, n0, n);
#endif
break;
case 0x24:
mc = 2;
nc = 4;
#if defined(__AVX2__) && defined(__F16C__)
gemmMx4<2>(m0, m, n0, n);
#else
gemm<2, 4>(m0, m, n0, n);
#endif
break;
#else
case 0x44:
case 0x43:
case 0x42:
mc = 4;
nc = 2;
#if defined(__AVX2__) && defined(__F16C__)
gemm4xN<2>(m0, m, n0, n);
#else
gemm<4, 2>(m0, m, n0, n);
#endif
break;
case 0x34:
case 0x24:
mc = 2;
nc = 4;
#if defined(__AVX2__) && defined(__F16C__)
gemmMx4<2>(m0, m, n0, n);
#else
gemm<2, 4>(m0, m, n0, n);
#endif
break;
case 0x33:
#endif
case 0x32:
mc = 3;
nc = 2;
gemm<3, 2>(m0, m, n0, n);
break;
case 0x23:
mc = 2;
nc = 3;
gemm<2, 3>(m0, m, n0, n);
break;
case 0x41:
mc = 4;
nc = 1;
#if defined(__AVX2__) && defined(__F16C__)
gemm4xN<1>(m0, m, n0, n);
#else
gemm<4, 1>(m0, m, n0, n);
#endif
break;
case 0x22:
mc = 2;
nc = 2;
gemm<2, 2>(m0, m, n0, n);
break;
case 0x14:
mc = 1;
nc = 4;
#if defined(__AVX2__) && defined(__F16C__)
gemmMx4<1>(m0, m, n0, n);
#else
gemm<1, 4>(m0, m, n0, n);
#endif
break;
case 0x31:
mc = 3;
nc = 1;
gemm<3, 1>(m0, m, n0, n);
break;
case 0x13:
mc = 1;
nc = 3;
gemm<1, 3>(m0, m, n0, n);
break;
case 0x21:
mc = 2;
nc = 1;
gemm<2, 1>(m0, m, n0, n);
break;
case 0x12:
mc = 1;
nc = 2;
gemm<1, 2>(m0, m, n0, n);
break;
case 0x11:
mc = 1;
nc = 1;
gemm<1, 1>(m0, m, n0, n);
break;
default:
return;
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
#if defined(__AVX2__) && defined(__F16C__)
// Templated functions for gemm of dimensions 4xN
template <int RN>
NOINLINE void gemm4xN(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / 4;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * 4;
int64_t jj = n0 + job % xtiles * RN;
__m256 Cv[RN][4] = {};
for (int64_t l = 0; l < k; ++l) {
uint64_t a_delta = ((uint64_t)A[lda * (ii + 3) + l].d << 48) | ((uint64_t)A[lda * (ii + 2) + l].d << 32) | ((uint64_t)A[lda * (ii + 1) + l].d << 16) | (A[lda * (ii + 0) + l].d);
// Convert delta values for four blocks to float values
__m128 da = _mm_cvtph_ps(_mm_set_epi64x(0, a_delta));
__m256i avec0 = load(A + lda * (ii + 0) + l);
__m256i avec1 = load(A + lda * (ii + 1) + l);
__m256i avec2 = load(A + lda * (ii + 2) + l);
__m256i avec3 = load(A + lda * (ii + 3) + l);
for (int64_t j = 0; j < RN; ++j) {
__m128 db = _mm_set1_ps(unhalf(B[ldb * (jj + j) + l].d));
// Computation of product of delta values for four blocks and replicate it across 256 bit lane
__m256 dvec = _mm256_castps128_ps256(_mm_mul_ps(da, db));
dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
// Computation of dot product and multiplication with appropriate delta value products
Cv[j][0] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
updot(_mm256_sign_epi8(avec0, avec0),
_mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec0)),
Cv[j][0]);
Cv[j][1] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
updot(_mm256_sign_epi8(avec1, avec1),
_mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec1)),
Cv[j][1]);
Cv[j][2] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
updot(_mm256_sign_epi8(avec2, avec2),
_mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec2)),
Cv[j][2]);
Cv[j][3] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
updot(_mm256_sign_epi8(avec3, avec3),
_mm256_sign_epi8(load(B + ldb * (jj + j) + l), avec3)),
Cv[j][3]);
}
}
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < 4; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
// Templated functions for gemm of dimensions Mx4
template <int RM>
NOINLINE void gemmMx4(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / 4;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * 4;
__m256 Cv[4][RM] = {};
for (int64_t l = 0; l < k; ++l) {
uint64_t b_delta = ((uint64_t)B[ldb * (jj + 3) + l].d << 48) | ((uint64_t)B[ldb * (jj + 2) + l].d << 32) | ((uint64_t)B[ldb * (jj + 1) + l].d << 16) | (B[ldb * (jj + 0) + l].d);
// Convert delta values for four blocks to float values
__m128 db = _mm_cvtph_ps(_mm_set_epi64x(0, b_delta));
__m256i bvec0 = load(B + ldb * (jj + 0) + l);
__m256i bvec1 = load(B + ldb * (jj + 1) + l);
__m256i bvec2 = load(B + ldb * (jj + 2) + l);
__m256i bvec3 = load(B + ldb * (jj + 3) + l);
for (int64_t i = 0; i < RM; ++i) {
__m128 da = _mm_set1_ps(unhalf((A[lda * (ii + i) + l].d)));
// Computation of product of delta values for four blocks and replicate it across 256 bit lane
__m256 dvec = _mm256_castps128_ps256(_mm_mul_ps(da, db));
dvec = _mm256_permute2f128_ps(dvec ,dvec, 0);
// Computation of dot product and multiplication with appropriate delta value products
Cv[0][i] = madd(_mm256_shuffle_ps(dvec, dvec, 0),
updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
load(A + lda * (ii + i) + l)),
_mm256_sign_epi8(bvec0, load(A + lda * (ii + i) + l))),
Cv[0][i]);
Cv[1][i] = madd(_mm256_shuffle_ps(dvec, dvec, 85),
updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
load(A + lda * (ii + i) + l)),
_mm256_sign_epi8(bvec1, load(A + lda * (ii + i) + l))),
Cv[1][i]);
Cv[2][i] = madd(_mm256_shuffle_ps(dvec, dvec, 170),
updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
load(A + lda * (ii + i) + l)),
_mm256_sign_epi8(bvec2, load(A + lda * (ii + i) + l))),
Cv[2][i]);
Cv[3][i] = madd(_mm256_shuffle_ps(dvec, dvec, 255),
updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
load(A + lda * (ii + i) + l)),
_mm256_sign_epi8(bvec3, load(A + lda * (ii + i) + l))),
Cv[3][i]);
}
}
for (int64_t j = 0; j < 4; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
#endif
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
__m256 Cv[RN][RM] = {};
for (int64_t l = 0; l < k; ++l)
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i) {
#if defined(__AVX2__)
__m256 udTmp = updot(_mm256_sign_epi8(load(A + lda * (ii + i) + l),
load(A + lda * (ii + i) + l)),
_mm256_sign_epi8(load(B + ldb * (jj + j) + l),
load(A + lda * (ii + i) + l)));
#else
__m128i ali0 = load0(A + lda * (ii + i) + l);
__m128i ali1 = load1(A + lda * (ii + i) + l);
__m128i blj0 = load0(B + ldb * (jj + j) + l);
__m128i blj1 = load1(B + ldb * (jj + j) + l);
__m128i sepAA0 = _mm_sign_epi8(ali0, ali0);
__m128i sepAA1 = _mm_sign_epi8(ali1, ali1);
__m128i sepBA0 = _mm_sign_epi8(blj0, ali0);
__m128i sepBA1 = _mm_sign_epi8(blj1, ali1);
// updot
const __m128i oneFill = _mm_set1_epi16(1);
__m128i mad0 = _mm_maddubs_epi16(sepAA0, sepBA0);
__m128i mad1 = _mm_maddubs_epi16(sepAA1, sepBA1);
__m256 udTmp = _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_madd_epi16(oneFill, mad1), _mm_madd_epi16(oneFill, mad0)));
#endif
Cv[j][i] = madd(_mm256_set1_ps(unhalf(A[lda * (ii + i) + l].d) *
unhalf(B[ldb * (jj + j) + l].d)),
udTmp,
Cv[j][i]);
}
for (int64_t j = 0; j < RN; ++j)
for (int64_t i = 0; i < RM; ++i)
C[ldc * (jj + j) + (ii + i)] = hsum(Cv[j][i]);
}
}
inline __m256i load(const block_q8_0 *b) {
return _mm256_loadu_si256((const __m256i *)b->qs);
}
inline __m128i load0(const block_q8_0 *b) {
return _mm_loadu_si128((const __m128i *)b->qs);
}
inline __m128i load1(const block_q8_0 *b) {
return _mm_loadu_si128(((const __m128i *)b->qs) + 1);
}
inline __m256i load(const block_q4_0 *b) {
return _mm256_sub_epi8(denibble(b->qs), _mm256_set1_epi8(8));
}
inline __m128i load0(const block_q4_0 *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), x), _mm_set1_epi8(8));
}
inline __m128i load1(const block_q4_0 *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)), _mm_set1_epi8(8));
}
inline __m256i load(const block_q5_0 *b) {
return _mm256_or_si256(denibble(b->qs), bittobyte(b->qh));
}
inline __m128i load0(const block_q5_0* b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
uint32_t x32;
memcpy(&x32, b->qh, sizeof(uint32_t));
__m128i qxl = _mm_and_si128(_mm_set1_epi8(15), x);
__m128i bytesl = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
_mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
_mm_shuffle_epi8(_mm_set1_epi32(x32),
_mm_set_epi64x(0x0101010101010101, 0x0000000000000000))));
bytesl = _mm_andnot_si128(bytesl, _mm_set1_epi8((char)0xF0));
return _mm_or_si128(qxl, bytesl);
}
inline __m128i load1(const block_q5_0* b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
uint32_t x32;
memcpy(&x32, b->qh, sizeof(uint32_t));
__m128i qxh = _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4));
__m128i bytesh = _mm_cmpeq_epi8(_mm_set1_epi64x(-1),
_mm_or_si128(_mm_set1_epi64x(0x7fbfdfeff7fbfdfe),
_mm_shuffle_epi8(_mm_set1_epi32(x32),
_mm_set_epi64x(0x0303030303030303, 0x0202020202020202))));
bytesh = _mm_andnot_si128(bytesh, _mm_set1_epi8((char)0xF0));
return _mm_or_si128(qxh, bytesh);
}
inline __m256i load(const block_iq4_nl *b) {
return MM256_SET_M128I(load1(b), load0(b));
}
inline __m128i load0(const block_iq4_nl *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), x));
}
inline __m128i load1(const block_iq4_nl *b) {
const __m128i x = _mm_loadu_si128((const __m128i *)(b->qs));
return _mm_shuffle_epi8(iq4nlt, _mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(x, 4)));
}
inline __m256 updot(__m256i u, __m256i s) {
__m256i res;
#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
res = _mm256_dpbusd_epi32(_mm256_setzero_si256(), u, s);
#elif defined(__AVXVNNI__)
res = _mm256_dpbusd_avx_epi32(_mm256_setzero_si256(), u, s);
#else
res = _mm256_madd_epi16(_mm256_set1_epi16(1), _mm256_maddubs_epi16(u, s));
#endif
return _mm256_cvtepi32_ps(res);
}
static inline __m256i denibble(const uint8_t *p) {
__m128i x = _mm_loadu_si128((const __m128i *)p);
return _mm256_and_si256(_mm256_set1_epi8(15),
_mm256_insertf128_si256(_mm256_castsi128_si256(x),
_mm_srli_epi16(x, 4), 1));
}
static inline __m256i bittobyte(const uint8_t *p) {
uint32_t x32;
memcpy(&x32, p, sizeof(uint32_t));
__m256i bytes = _mm256_cmpeq_epi8(_mm256_set1_epi64x(-1),
_mm256_or_si256(_mm256_set1_epi64x(0x7fbfdfeff7fbfdfe),
_mm256_shuffle_epi8(_mm256_set1_epi32(x32),
_mm256_set_epi64x(0x0303030303030303, 0x0202020202020202,
0x0101010101010101, 0x0000000000000000))));
return _mm256_andnot_si256(bytes, _mm256_set1_epi8((char)0xF0));
}
const TA *const A;
const TB *const B;
TC *const C;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
#endif // __AVX__
//PPC Implementation
#if defined(__MMA__)
#define SAVE_ACC(ACC, ii, jj) \
__builtin_mma_disassemble_acc(vec_C, ACC); \
for (int I = 0; I < 4; I++) { \
for (int J = 0; J < 4; J++) { \
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&vec_C[I]+J); \
} \
} \
template <typename TA, typename TB, typename TC>
class tinyBLAS_Q0_PPC {
public:
tinyBLAS_Q0_PPC(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
template<int RM, int RN>
inline void save_res(int ii, int jj, int idx, vector float* fin_res) {
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((float*)(C+ii+((jj+J)*ldc)+I)) = *((float*)&fin_res[idx+I]+J);
}
}
}
template<int size>
inline void compute(acc_t* ACC, int c_idx, int s_idx, std::array<int, size>& comparray, vector float* vs, vector float* fin_res) {
vector signed int vec_C[4];
vector float CA[4] = {0};
vector float res[4] = {0};
__builtin_mma_disassemble_acc(vec_C, ACC);
for (int i = 0; i < 4; i++) {
CA[i] = vec_splats((float)(((double)comparray[c_idx+i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[s_idx+i] = vec_madd(res[i], vs[s_idx+i], fin_res[s_idx+i]);
}
}
template<typename VA, typename VB>
void packNormal(const TA* a, int64_t lda, int rows, int cols, VA* vec, bool flip) {
int64_t i, j;
TA *aoffset = NULL;
VA *vecOffset = NULL;
TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
VB c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2]={0};
VB c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2]={0};
VB t1, t2, t3, t4, t5, t6, t7, t8;
vector unsigned char xor_vector;
uint8_t flip_vec = 0x80;
xor_vector = vec_splats(flip_vec);
vector unsigned char swiz1 = {0, 1, 2, 3, 4, 5, 6, 7, 16, 17, 18, 19, 20, 21, 22, 23};
vector unsigned char swiz2 = {8, 9, 10, 11, 12, 13, 14, 15, 24, 25, 26, 27, 28, 29, 30, 31};
vector unsigned char swiz3 = {0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27};
vector unsigned char swiz4 = {4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31};
aoffset = const_cast<TA*>(a);
vecOffset = vec;
j = (rows >> 3);
if (j > 0) {
do {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset += 8 * lda;
i = (cols >> 3);
if (i > 0) {
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4->qs);
C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5->qs);
C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6->qs);
C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7->qs);
C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8->qs);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
__builtin_vsx_disassemble_pair(c5, &C5);
__builtin_vsx_disassemble_pair(c6, &C6);
__builtin_vsx_disassemble_pair(c7, &C7);
__builtin_vsx_disassemble_pair(c8, &C8);
t1 = vec_perm(c1[0], c2[0], swiz1);
t2 = vec_perm(c1[0], c2[0], swiz2);
t3 = vec_perm(c3[0], c4[0], swiz1);
t4 = vec_perm(c3[0], c4[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset);
vec_xst(t6, 0, vecOffset+16);
vec_xst(t7, 0, vecOffset+32);
vec_xst(t8, 0, vecOffset+48);
t1 = vec_perm(c1[1], c2[1], swiz1);
t2 = vec_perm(c1[1], c2[1], swiz2);
t3 = vec_perm(c3[1], c4[1], swiz1);
t4 = vec_perm(c3[1], c4[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+64);
vec_xst(t6, 0, vecOffset+80);
vec_xst(t7, 0, vecOffset+96);
vec_xst(t8, 0, vecOffset+112);
t1 = vec_perm(c5[0], c6[0], swiz1);
t2 = vec_perm(c5[0], c6[0], swiz2);
t3 = vec_perm(c7[0], c8[0], swiz1);
t4 = vec_perm(c7[0], c8[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+128);
vec_xst(t6, 0, vecOffset+144);
vec_xst(t7, 0, vecOffset+160);
vec_xst(t8, 0, vecOffset+176);
t1 = vec_perm(c5[1], c6[1], swiz1);
t2 = vec_perm(c5[1], c6[1], swiz2);
t3 = vec_perm(c7[1], c8[1], swiz1);
t4 = vec_perm(c7[1], c8[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+192);
vec_xst(t6, 0, vecOffset+208);
vec_xst(t7, 0, vecOffset+224);
vec_xst(t8, 0, vecOffset+240);
aoffset1 += lda;
aoffset2 += lda;
aoffset3 += lda;
aoffset4 += lda;
aoffset5 += lda;
aoffset6 += lda;
aoffset7 += lda;
aoffset8 += lda;
vecOffset += 256;
i--;
} while(i > 0);
}
j--;
} while(j > 0);
}
if (rows & 4) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset += 4 * lda;
i = (cols >> 3);
if (i > 0) {
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4->qs);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
t1 = vec_perm(c1[0], c2[0], swiz1);
t2 = vec_perm(c1[0], c2[0], swiz2);
t3 = vec_perm(c3[0], c4[0], swiz1);
t4 = vec_perm(c3[0], c4[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset);
vec_xst(t6, 0, vecOffset+16);
vec_xst(t7, 0, vecOffset+32);
vec_xst(t8, 0, vecOffset+48);
t1 = vec_perm(c1[1], c2[1], swiz1);
t2 = vec_perm(c1[1], c2[1], swiz2);
t3 = vec_perm(c3[1], c4[1], swiz1);
t4 = vec_perm(c3[1], c4[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+64);
vec_xst(t6, 0, vecOffset+80);
vec_xst(t7, 0, vecOffset+96);
vec_xst(t8, 0, vecOffset+112);
aoffset1 += lda;
aoffset2 += lda;
aoffset3 += lda;
aoffset4 += lda;
vecOffset += 128;
i--;
} while(i > 0);
}
}
if (rows & 3) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
i = (cols >> 3);
if (i > 0) {
do {
switch(rows) {
case 3: C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3->qs);
__builtin_vsx_disassemble_pair(c3, &C3);
case 2: C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2->qs);
__builtin_vsx_disassemble_pair(c2, &C2);
case 1: C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1->qs);
__builtin_vsx_disassemble_pair(c1, &C1);
break;
}
t1 = vec_perm(c1[0], c2[0], swiz1);
t2 = vec_perm(c1[0], c2[0], swiz2);
t3 = vec_perm(c3[0], c4[0], swiz1);
t4 = vec_perm(c3[0], c4[0], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset);
vec_xst(t6, 0, vecOffset+16);
vec_xst(t7, 0, vecOffset+32);
vec_xst(t8, 0, vecOffset+48);
t1 = vec_perm(c1[1], c2[1], swiz1);
t2 = vec_perm(c1[1], c2[1], swiz2);
t3 = vec_perm(c3[1], c4[1], swiz1);
t4 = vec_perm(c3[1], c4[1], swiz2);
t5 = vec_perm(t1, t3, swiz3);
t6 = vec_perm(t1, t3, swiz4);
t7 = vec_perm(t2, t4, swiz3);
t8 = vec_perm(t2, t4, swiz4);
if (flip == true) {
t5 = vec_xor(t5, xor_vector);
t6 = vec_xor(t6, xor_vector);
t7 = vec_xor(t7, xor_vector);
t8 = vec_xor(t8, xor_vector);
}
vec_xst(t5, 0, vecOffset+64);
vec_xst(t6, 0, vecOffset+80);
vec_xst(t7, 0, vecOffset+96);
vec_xst(t8, 0, vecOffset+112);
aoffset1 += lda;
aoffset2 += lda;
aoffset3 += lda;
vecOffset += 128;
i--;
} while(i > 0);
}
}
}
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
int m_rem = MIN(m - m0, 8);
int n_rem = MIN(n - n0, 8);
// TO-DO: KERNEL_16x8 and KERNEL_8x16 are having some performance
// issues. After resolving them, below code will be enabled.
/*if (m_rem >= 16 && n_rem >= 8) {
mc = 16;
nc = 8;
gemm<16,8>(m0, m, n0, n);
} else if(m_rem >= 8 && n_rem >= 16) {
mc = 8;
nc = 16;
gemm<8,16>(m0, m, n0, n);
}*/
if (m_rem >= 8 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 8) {
mc = 4;
nc = 8;
gemm<4,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 4) {
mc = 8;
nc = 4;
gemm<8,4>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 4) {
mc = 4;
nc = 4;
gemm_small<4, 4>(m0, m, n0, n);
} else if ((m_rem < 4) && (n_rem > 4)) {
nc = 4;
switch(m_rem) {
case 1:
mc = 1;
gemm_small<1, 4>(m0, m, n0, n);
break;
case 2:
mc = 2;
gemm_small<2, 4>(m0, m, n0, n);
break;
case 3:
mc = 3;
gemm_small<3, 4>(m0, m, n0, n);
break;
default:
return;
}
} else if ((m_rem > 4) && (n_rem < 4)) {
mc = 4;
switch(n_rem) {
case 1:
nc = 1;
gemm_small<4, 1>(m0, m, n0, n);
break;
case 2:
nc = 2;
gemm_small<4, 2>(m0, m, n0, n);
break;
case 3:
nc = 3;
gemm_small<4, 3>(m0, m, n0, n);
break;
default:
return;
}
} else {
switch((m_rem << 4) | n_rem) {
case 0x43:
mc = 4;
nc = 3;
gemm_small<4, 3>(m0, m, n0, n);
break;
case 0x42:
mc = 4;
nc = 2;
gemm_small<4, 2>(m0, m, n0, n);
break;
case 0x41:
mc = 4;
nc = 1;
gemm_small<4, 1>(m0, m, n0, n);
break;
case 0x34:
mc = 3;
nc = 4;
gemm_small<3, 4>(m0, m, n0, n);
break;
case 0x33:
mc = 3;
nc = 3;
gemm_small<3, 3>(m0, m, n0, n);
break;
case 0x32:
mc = 3;
nc = 2;
gemm_small<3, 2>(m0, m, n0, n);
break;
case 0x31:
mc = 3;
nc = 1;
gemm_small<3, 1>(m0, m, n0, n);
break;
case 0x24:
mc = 2;
nc = 4;
gemm_small<2, 4>(m0, m, n0, n);
break;
case 0x23:
mc = 2;
nc = 3;
gemm_small<2, 3>(m0, m, n0, n);
break;
case 0x22:
mc = 2;
nc = 2;
gemm_small<2, 2>(m0, m, n0, n);
break;
case 0x21:
mc = 2;
nc = 1;
gemm_small<2, 1>(m0, m, n0, n);
break;
case 0x14:
mc = 1;
nc = 4;
gemm_small<1, 4>(m0, m, n0, n);
break;
case 0x13:
mc = 1;
nc = 3;
gemm_small<1, 3>(m0, m, n0, n);
break;
case 0x12:
mc = 1;
nc = 2;
gemm_small<1, 2>(m0, m, n0, n);
break;
case 0x11:
mc = 1;
nc = 1;
gemm_small<1, 1>(m0, m, n0, n);
break;
default:
return;
}
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
void KERNEL_4x8(int64_t ii, int64_t jj) {
vec_t vec_A[8], vec_B[16] = {0};
acc_t acc_0, acc_1;
std::array<int, 4> comparray;
vector float fin_res[8] = {0};
vector float vs[8] = {0};
for (int l = 0; l < k; l++) {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 4, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_1, vec_A[x], vec_B[x+8]);
}
for (int I = 0; I<4; I++) {
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
*((float*)&vs[I+4]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
}
}
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < 4; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
compute<4>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<4>(&acc_1, 0, 4, comparray, vs, fin_res);
}
save_res<4, 4>(ii, jj, 0, fin_res);
save_res<4, 4>(ii, jj+4, 4, fin_res);
}
void KERNEL_8x4(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[8] = {0};
acc_t acc_0, acc_1;
std::array<int, 8> comparray;
vector float fin_res[8] = {0};
vector float vs[8] = {0};
for (int l = 0; l < k; l++) {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 4, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
}
for (int I = 0; I<8; I++) {
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
}
}
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < 8; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
}
save_res<4, 4>(ii, jj, 0, fin_res);
save_res<4, 4>(ii+4, jj, 4, fin_res);
}
void KERNEL_8x8(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[16] = {0};
acc_t acc_0, acc_1, acc_2, acc_3;
std::array<int, 8> comparray;
vector float fin_res[16] = {0};
vector float vs[16] = {0};
for (int l = 0; l < k; l++) {
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
__builtin_mma_xxsetaccz(&acc_2);
__builtin_mma_xxsetaccz(&acc_3);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, 8, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, 8, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x++) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_1, vec_A[x+8], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_2, vec_A[x], vec_B[x+8]);
__builtin_mma_xvi8ger4pp(&acc_3, vec_A[x+8], vec_B[x+8]);
}
for (int I = 0; I<8; I++) {
for (int J = 0; J<4; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
*((float*)&vs[I+8]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J+4)*ldb)+l)->d));
}
}
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < 8; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
compute<8>(&acc_0, 0, 0, comparray, vs, fin_res);
compute<8>(&acc_1, 4, 4, comparray, vs, fin_res);
compute<8>(&acc_2, 0, 8, comparray, vs, fin_res);
compute<8>(&acc_3, 4, 12, comparray, vs, fin_res);
}
save_res<4, 4>(ii, jj, 0, fin_res);
save_res<4, 4>(ii+4, jj, 4, fin_res);
save_res<4, 4>(ii, jj+4, 8, fin_res);
save_res<4, 4>(ii+4, jj+4, 12, fin_res);
}
template<int RM, int RN>
void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
vec_t vec_A[8], vec_B[8] = {0};
vector signed int vec_C[4];
acc_t acc_0;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
std::array<int, RM> comparray;
vector float res[4] = {0};
vector float fin_res[4] = {0};
vector float vs[4] = {0};
vector float CA[4] = {0};
__builtin_prefetch((A+(ii*lda)+0)->qs, 0, 1); // prefetch first value
__builtin_prefetch((B+(jj*ldb)+0)->qs, 0, 1); // prefetch first value
for (int l = 0; l < k; l++) {
__builtin_prefetch((A+(ii*lda)+(l+1))->qs, 0, 1); // prefetch one loop ahead
__builtin_prefetch((B+(jj*ldb)+(l+1))->qs, 0, 1); // prefetch one loop ahead
__builtin_mma_xxsetaccz(&acc_0);
packNormal<int8_t, vector signed char>((A+(ii*lda)+l), lda, RM, 8, (int8_t*)vec_A, false);
packNormal<uint8_t, vector unsigned char>((B+(jj*ldb)+l), ldb, RN, 8, (uint8_t*)vec_B, true);
for(int x = 0; x < 8; x+=4) {
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x], vec_B[x]);
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+1], vec_B[x+1]);
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+2], vec_B[x+2]);
__builtin_mma_xvi8ger4pp(&acc_0, vec_A[x+3], vec_B[x+3]);
}
for (int I = 0; I<RM; I++) {
for (int J = 0; J<RN; J++) {
*((float*)&vs[I]+J) = (unhalf((A+((ii+I)*lda)+l)->d) * unhalf((B+((jj+J)*ldb)+l)->d));
}
}
__builtin_mma_disassemble_acc(vec_C, &acc_0);
auto aoffset = A+(ii*lda)+l;
for (int i = 0; i < RM; i++) {
comparray[i] = 0;
int ca = 0;
const int8_t *at = aoffset->qs;
for (int j = 0; j < 32; j++)
ca += (int)*at++;
comparray[i] = ca;
aoffset += lda;
}
for (int i = 0; i < RM; i++) {
CA[i] = vec_splats((float)(((double)comparray[i]) * -128.0));
res[i] = vec_add(vec_ctf(vec_C[i], 0), CA[i]);
fin_res[i] = vec_madd(res[i], vs[i], fin_res[i]);
}
}
save_res<RM, RN>(ii, jj, 0, fin_res);
}
}
template<int RM, int RN>
inline void kernel(int64_t ii, int64_t jj) {
if constexpr(RM == 4 && RN == 8) {
KERNEL_4x8(ii,jj);
} else if constexpr(RM == 8 && RN == 4) {
KERNEL_8x4(ii,jj);
} else if constexpr(RM == 8 && RN == 8) {
KERNEL_8x8(ii,jj);
} else {
static_assert(false, "RN/RM values not supported");
}
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
kernel<RM, RN>(ii, jj);
}
}
const TA *const A;
const TB *const B;
TC *C;
TA *At;
TB *Bt;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
template <typename TA, typename TB, typename TC>
class tinyBLAS_PPC {
public:
tinyBLAS_PPC(int64_t k,
const TA *A, int64_t lda,
const TB *B, int64_t ldb,
TC *C, int64_t ldc,
int ith, int nth)
: A(A), B(B), C(C), k(k), lda(lda), ldb(ldb), ldc(ldc), ith(ith), nth(nth) {
}
void matmul(int64_t m, int64_t n) {
mnpack(0, m, 0, n);
}
private:
void (tinyBLAS_PPC::*kernel)(int64_t, int64_t);
template<typename VA>
void packTranspose(const TA* a, int64_t lda, int rows, int cols, TA* vec) {
int64_t i, j;
TA *aoffset = NULL, *boffset = NULL;
TA *aoffset1 = NULL, *aoffset2 = NULL, *aoffset3 = NULL, *aoffset4 = NULL;
TA *aoffset5 = NULL, *aoffset6 = NULL, *aoffset7 = NULL, *aoffset8 = NULL;
__vector_pair C1, C2, C3, C4, C5, C6, C7, C8;
VA c1[2] = {0}, c2[2] = {0}, c3[2] = {0}, c4[2] = {0};
VA c5[2] = {0}, c6[2] = {0}, c7[2] = {0}, c8[2] = {0};
VA t1, t2, t3, t4, t5, t6, t7, t8;
aoffset = const_cast<TA*>(a);
boffset = vec;
j = (rows >> 3);
if (j > 0) {
do {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset5 = aoffset4 + lda;
aoffset6 = aoffset5 + lda;
aoffset7 = aoffset6 + lda;
aoffset8 = aoffset7 + lda;
aoffset += 8 * lda;
i = (cols >> 3);
if (i > 0) {
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
C5 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset5);
C6 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset6);
C7 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset7);
C8 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset8);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
__builtin_vsx_disassemble_pair(c5, &C5);
__builtin_vsx_disassemble_pair(c6, &C6);
__builtin_vsx_disassemble_pair(c7, &C7);
__builtin_vsx_disassemble_pair(c8, &C8);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergeh(c5[0], c6[0]);
t4 = vec_mergeh(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_mergel(c5[0], c6[0]);
t4 = vec_mergel(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
t1 = vec_mergeh(c1[1], c2[1]);
t2 = vec_mergeh(c3[1], c4[1]);
t3 = vec_mergeh(c5[1], c6[1]);
t4 = vec_mergeh(c7[1], c8[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+32);
vec_xst(t6, 0, boffset+36);
vec_xst(t7, 0, boffset+40);
vec_xst(t8, 0, boffset+44);
t1 = vec_mergel(c1[1], c2[1]);
t2 = vec_mergel(c3[1], c4[1]);
t3 = vec_mergel(c5[1], c6[1]);
t4 = vec_mergel(c7[1], c8[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+48);
vec_xst(t6, 0, boffset+52);
vec_xst(t7, 0, boffset+56);
vec_xst(t8, 0, boffset+60);
aoffset1 += 8*lda;
aoffset2 += 8*lda;
aoffset3 += 8*lda;
aoffset4 += 8*lda;
boffset += 64;
i--;
} while(i > 0);
}
if (cols & 4) {
c1[0] = vec_xl(0, aoffset1);
c2[0] = vec_xl(0, aoffset2);
c3[0] = vec_xl(0, aoffset3);
c4[0] = vec_xl(0, aoffset4);
c5[0] = vec_xl(0, aoffset5);
c6[0] = vec_xl(0, aoffset6);
c7[0] = vec_xl(0, aoffset7);
c8[0] = vec_xl(0, aoffset8);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergeh(c5[0], c6[0]);
t4 = vec_mergeh(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_mergel(c5[0], c6[0]);
t4 = vec_mergel(c7[0], c8[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t3, t4, 0);
t7 = vec_xxpermdi(t1, t2, 3);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
}
j--;
} while(j > 0);
}
if (rows & 4) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
aoffset4 = aoffset3 + lda;
aoffset += 4 * lda;
i = (cols >> 3);
if (i > 0) {
do {
C1 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset1);
C2 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset2);
C3 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset3);
C4 = __builtin_vsx_lxvp(0, (__vector_pair*)aoffset4);
__builtin_vsx_disassemble_pair(c1, &C1);
__builtin_vsx_disassemble_pair(c2, &C2);
__builtin_vsx_disassemble_pair(c3, &C3);
__builtin_vsx_disassemble_pair(c4, &C4);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_mergel(c1[0], c2[0]);
t4 = vec_mergel(c3[0], c4[0]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t1, t2, 3);
t7 = vec_xxpermdi(t3, t4, 0);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset);
vec_xst(t6, 0, boffset+4);
vec_xst(t7, 0, boffset+8);
vec_xst(t8, 0, boffset+12);
t1 = vec_mergeh(c1[1], c2[1]);
t2 = vec_mergeh(c3[1], c4[1]);
t3 = vec_mergel(c1[1], c2[1]);
t4 = vec_mergel(c3[1], c4[1]);
t5 = vec_xxpermdi(t1, t2, 0);
t6 = vec_xxpermdi(t1, t2, 3);
t7 = vec_xxpermdi(t3, t4, 0);
t8 = vec_xxpermdi(t3, t4, 3);
vec_xst(t5, 0, boffset+16);
vec_xst(t6, 0, boffset+20);
vec_xst(t7, 0, boffset+24);
vec_xst(t8, 0, boffset+28);
aoffset1 += 8*lda;
aoffset2 += 8*lda;
aoffset3 += 8*lda;
aoffset4 += 8*lda;
boffset += 32;
i--;
} while(i > 0);
}
if (cols & 4) {
c1[0] = vec_xl(0, aoffset1);
c2[0] = vec_xl(0, aoffset2);
c3[0] = vec_xl(0, aoffset3);
c4[0] = vec_xl(0, aoffset4);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
vec_xst(t4, 0, boffset+12);
}
}
if (rows & 3) {
aoffset1 = aoffset;
aoffset2 = aoffset1 + lda;
aoffset3 = aoffset2 + lda;
if (cols & 4) {
c1[0] = vec_xl(0, aoffset1);
c2[0] = vec_xl(0, aoffset2);
c3[0] = vec_xl(0, aoffset3);
t1 = vec_mergeh(c1[0], c2[0]);
t2 = vec_mergeh(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset);
vec_xst(t4, 0, boffset+4);
t1 = vec_mergel(c1[0], c2[0]);
t2 = vec_mergel(c3[0], c4[0]);
t3 = vec_xxpermdi(t1, t2, 0);
t4 = vec_xxpermdi(t1, t2, 3);
vec_xst(t3, 0, boffset+8);
vec_xst(t4, 0, boffset+12);
}
}
}
void KERNEL_4x4(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[4], vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
for (int l = 0; l < k; l+=4) {
packTranspose<vector float>(A+(ii*lda)+l, lda, 4, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
}
SAVE_ACC(&acc_0, ii, jj);
}
void KERNEL_4x8(int64_t ii, int64_t jj) {
vec_t vec_A[4], vec_B[8], vec_C[4];
acc_t acc_0, acc_1;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
packTranspose<vector float>(A+(ii*lda)+l, lda, 4, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 8, 4, (TA*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], (vec_t)vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[0], (vec_t)vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], (vec_t)vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[1], (vec_t)vec_B[3]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], (vec_t)vec_B[4]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[2], (vec_t)vec_B[5]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], (vec_t)vec_B[6]);
__builtin_mma_xvf32gerpp(&acc_1, vec_A[3], (vec_t)vec_B[7]);
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii, jj+4);
}
void KERNEL_8x4(int64_t ii, int64_t jj) {
vec_t vec_A[8], vec_B[4], vec_C[4];
acc_t acc_0, acc_1;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
for (int64_t l = 0; l < k; l+=4) {
packTranspose<vector float>(A+(ii*lda)+l, lda, 8, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[1], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[2], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[3], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[4], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[5], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[6], vec_B[3]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[7], vec_B[3]);
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii+4, jj);
}
void KERNEL_8x8(int64_t ii, int64_t jj) {
vec_t vec_A[16], vec_B[16], vec_C[4];
acc_t acc_0, acc_1, acc_2, acc_3;
__builtin_mma_xxsetaccz(&acc_0);
__builtin_mma_xxsetaccz(&acc_1);
__builtin_mma_xxsetaccz(&acc_2);
__builtin_mma_xxsetaccz(&acc_3);
for (int l = 0; l < k; l+=8) {
packTranspose<vector float>(A+(ii*lda)+l, lda, 8, 8, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 8, 8, (TA*)vec_B);
for(int x = 0; x < 16; x+=2) {
__builtin_mma_xvf32gerpp(&acc_0, (vec_t)vec_A[x], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_1, (vec_t)vec_A[x], vec_B[x+1]);
__builtin_mma_xvf32gerpp(&acc_2, (vec_t)vec_A[x+1], vec_B[x]);
__builtin_mma_xvf32gerpp(&acc_3, (vec_t)vec_A[x+1], vec_B[x+1]);
}
}
SAVE_ACC(&acc_0, ii, jj);
SAVE_ACC(&acc_1, ii, jj+4);
SAVE_ACC(&acc_2, ii+4, jj);
SAVE_ACC(&acc_3, ii+4, jj+4);
}
void mnpack(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t mc, nc, mp, np;
int m_rem = MIN(m - m0, 16);
int n_rem = MIN(n - n0, 16);
if (m_rem >= 16 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if(m_rem >= 8 && n_rem >= 16) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 8) {
mc = 8;
nc = 8;
gemm<8,8>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 8) {
mc = 4;
nc = 8;
gemm<4,8>(m0, m, n0, n);
} else if (m_rem >= 8 && n_rem >= 4) {
mc = 8;
nc = 4;
gemm<8,4>(m0, m, n0, n);
} else if (m_rem >= 4 && n_rem >= 4) {
mc = 4;
nc = 4;
gemm<4,4>(m0, m, n0, n);
} else if ((m_rem < 4) && (n_rem > 4)) {
nc = 4;
switch(m_rem) {
case 1:
mc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 2:
mc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 3:
mc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
} else if ((m_rem > 4) && (n_rem < 4)) {
mc = 4;
switch(n_rem) {
case 1:
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 2:
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 3:
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
} else {
switch((m_rem << 4) | n_rem) {
case 0x43:
mc = 4;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x42:
mc = 4;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x41:
mc = 4;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x34:
mc = 3;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x33:
mc = 3;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x32:
mc = 3;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x31:
mc = 3;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x24:
mc = 2;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x23:
mc = 2;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x22:
mc = 2;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x21:
mc = 2;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x14:
mc = 1;
nc = 4;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x13:
mc = 1;
nc = 3;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x12:
mc = 1;
nc = 2;
gemm_small(m0, m, n0, n, mc, nc);
break;
case 0x11:
mc = 1;
nc = 1;
gemm_small(m0, m, n0, n, mc, nc);
break;
default:
return;
}
}
mp = m0 + (m - m0) / mc * mc;
np = n0 + (n - n0) / nc * nc;
mnpack(mp, m, n0, np);
mnpack(m0, m, np, n);
}
void gemm_small(int64_t m0, int64_t m, int64_t n0, int64_t n, int RM, int RN) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
vec_t vec_C[4];
acc_t acc_0;
__builtin_mma_xxsetaccz(&acc_0);
vec_t vec_A[4], vec_B[4];
for (int l=0; l<k; l+=4) {
if (RN >= 4 && RM == 1) {
TA* a = const_cast<TA*>(A+(ii)*lda+l);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, 4, 4, (TA*)vec_B);
vec_A[0] = (vec_t)vec_xl(0,a);
vec_A[1] = (vec_t)vec_splats(*((TA*)&vec_A+1));
vec_A[2] = (vec_t)vec_splats(*((TA*)&vec_A+2));
vec_A[3] = (vec_t)vec_splats(*((TA*)&vec_A+3));
} else {
packTranspose<vector float>(A+(ii*lda)+l, lda, RM, 4, (TA*)vec_A);
packTranspose<vector float>(B+(jj*ldb)+l, ldb, RN, 4, (TA*)vec_B);
}
__builtin_mma_xvf32gerpp(&acc_0, vec_A[0], vec_B[0]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[1], vec_B[1]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[2], vec_B[2]);
__builtin_mma_xvf32gerpp(&acc_0, vec_A[3], vec_B[3]);
}
__builtin_mma_disassemble_acc(vec_C, &acc_0);
for (int I = 0; I < RM; I++) {
for (int J = 0; J < RN; J++) {
*((TC*)(C+ii+((jj+J)*ldc)+I)) = *((TC*)&vec_C[I]+J);
}
}
}
}
template <int RM, int RN>
NOINLINE void gemm(int64_t m0, int64_t m, int64_t n0, int64_t n) {
int64_t ytiles = (m - m0) / RM;
int64_t xtiles = (n - n0) / RN;
int64_t tiles = xtiles * ytiles;
int64_t duty = (tiles + nth - 1) / nth;
int64_t start = duty * ith;
int64_t end = start + duty;
if (RM == 4 && RN == 4) {
kernel = &tinyBLAS_PPC::KERNEL_4x4;
} else if (RM == 4 && RN == 8) {
kernel = &tinyBLAS_PPC::KERNEL_4x8;
} else if (RM == 8 && RN == 4) {
kernel = &tinyBLAS_PPC::KERNEL_8x4;
} else if (RM == 8 && RN == 8) {
kernel = &tinyBLAS_PPC::KERNEL_8x8;
}
if (end > tiles)
end = tiles;
for (int64_t job = start; job < end; ++job) {
int64_t ii = m0 + job / xtiles * RM;
int64_t jj = n0 + job % xtiles * RN;
(this->*kernel)(ii, jj);
}
}
const TA *const A;
const TB *const B;
TC *C;
TA *At;
TB *Bt;
const int64_t k;
const int64_t lda;
const int64_t ldb;
const int64_t ldc;
const int ith;
const int nth;
};
#endif
} // namespace
/**
* Performs optimized matrix multiplication on CPU.
*
* This subroutine may compute C = Aᵀ * B with column major ordering.
* Despite its name, this isn't a generalized implementation. Work is
* only performed when a handwritten kernel is written and available.
* Otherwise the caller should fall back to a general matmul routine.
*
* For example, for single-threaded single-precision GEMM you can say
*
* llamafile_sgemm(m, n, k, A, lda, B, ldb, C, ldc,
* 0, 1,
* GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32);
*
* @param m is rows in `A` and `C`
* @param n is cols in `B` and `C`
* @param k is cols in `A` and rows in `B`
* @param A is first input matrix (always transposed)
* @param lda is row stride of `A`
* @param B is second input matrix (never transposed)
* @param ldb is row stride of `B`
* @param C is input/output array of output matrices
* @param ldc is row stride of `C`
* @param ith is thread id (must be less than `nth`)
* @param nth is number of threads (must be greater than zero)
* @param Atype is GGML data type of `A`
* @param Btype is GGML data type of `B`
* @param Ctype is GGML data type of `C`
* @return true if this function was able to service the matmul request
*/
bool llamafile_sgemm(const struct ggml_compute_params * params, int64_t m, int64_t n, int64_t k,
const void *A, int64_t lda, const void *B, int64_t ldb, void *C,
int64_t ldc, int Atype, int Btype, int Ctype) {
assert(m >= 0);
assert(n >= 0);
assert(k >= 0);
assert(lda >= k);
assert(ldb >= k);
assert(ldc >= m);
assert(params->nth > 0);
assert(params->ith < params->nth);
// only enable sgemm for prompt processing
if (n < 2)
return false;
if (Ctype != GGML_TYPE_F32)
return false;
switch (Atype) {
case GGML_TYPE_F32: {
if (Btype != GGML_TYPE_F32)
return false;
#if defined(__AVX512F__)
tinyBLAS<16, __m512, __m512, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__AVX__) || defined(__AVX2__)
tinyBLAS<8, __m256, __m256, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__ARM_NEON)
if (n < 4)
return false;
tinyBLAS<4, float32x4_t, float32x4_t, float, float, float> tb{ params,
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
#elif defined(__MMA__)
if (k % 8)
return false;
tinyBLAS_PPC<float, float, float> tb{
k, (const float *)A, lda,
(const float *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
case GGML_TYPE_BF16: {
#if defined(__AVX512BF16__)
if (Btype == GGML_TYPE_BF16) {
tinyBLAS<32, __m512, __m512bh, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
(const ggml_bf16_t *)A, lda,
(const ggml_bf16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__AVX512F__)
if (Btype == GGML_TYPE_BF16) {
tinyBLAS<16, __m512, __m512, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
(const ggml_bf16_t *)A, lda,
(const ggml_bf16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__AVX2__)
if (Btype == GGML_TYPE_BF16) {
tinyBLAS<8, __m256, __m256, ggml_bf16_t, ggml_bf16_t, float> tb{ params, k,
(const ggml_bf16_t *)A, lda,
(const ggml_bf16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#endif
return false;
}
case GGML_TYPE_F16: {
#if defined(__AVX512F__)
if (Btype == GGML_TYPE_F16) {
tinyBLAS<16, __m512, __m512, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
(const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif (defined(__AVX__) || defined(__AVX2__)) && defined(__F16C__)
if (Btype == GGML_TYPE_F16) {
tinyBLAS<8, __m256, __m256, ggml_fp16_t, ggml_fp16_t, float> tb{ params, k,
(const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC) && !defined(_MSC_VER)
if (n < 8)
return false;
if (Btype == GGML_TYPE_F16) {
tinyBLAS<8, float16x8_t, float16x8_t, ggml_fp16_t, ggml_fp16_t, float> tb{ params,
k, (const ggml_fp16_t *)A, lda,
(const ggml_fp16_t *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#elif defined(__ARM_NEON) && !defined(_MSC_VER)
if (Btype == GGML_TYPE_F32) {
tinyBLAS<4, float32x4_t, float32x4_t, ggml_fp16_t, float, float> tb{ params,
k, (const ggml_fp16_t *)A, lda,
(const float *)B, ldb,
(float *)C, ldc};
return tb.matmul(m, n);
}
#endif
return false;
}
case GGML_TYPE_Q8_0: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_q8_0, block_q8_0, float> tb{
k, (const block_q8_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_DOTPROD)
tinyBLAS_Q0_ARM<block_q8_0> tb{
k, (const block_q8_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#elif defined(__MMA__)
if (n < 8 && n != 4)
return false;
if (m < 8 && m != 4)
return false;
tinyBLAS_Q0_PPC<block_q8_0, block_q8_0, float> tb{
k, (const block_q8_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
case GGML_TYPE_Q4_0: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_q4_0, block_q8_0, float> tb{
k, (const block_q4_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#elif defined(__ARM_FEATURE_DOTPROD)
tinyBLAS_Q0_ARM<block_q4_0> tb{
k, (const block_q4_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
case GGML_TYPE_Q5_0: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_q5_0, block_q8_0, float> tb{
k, (const block_q5_0 *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
case GGML_TYPE_IQ4_NL: {
if (Btype != GGML_TYPE_Q8_0)
return false;
#if defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX__)
tinyBLAS_Q0_AVX<block_iq4_nl, block_q8_0, float> tb{
k, (const block_iq4_nl *)A, lda,
(const block_q8_0 *)B, ldb,
(float *)C, ldc,
params->ith, params->nth};
tb.matmul(m, n);
return true;
#else
return false;
#endif
}
default:
return false;
}
(void)params;
(void)m;
(void)n;
(void)k;
(void)A;
(void)lda;
(void)B;
(void)ldb;
(void)C;
(void)ldc;
(void)Atype;
(void)Btype;
(void)Ctype;
}
|