Spaces:
Running
Running
File size: 101,167 Bytes
c61ccee |
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 2599 2600 2601 2602 2603 2604 2605 2606 2607 2608 2609 2610 2611 2612 2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 |
from __future__ import annotations
import base64
import copyreg
import dataclasses
import functools
import hashlib
import importlib
import io
import json
import logging
import multiprocessing
import os
import pathlib
import pickle
import pkgutil
import platform
import re
import shlex
import shutil
import signal
import subprocess
import sys
import sysconfig
import tempfile
import textwrap
import threading
import warnings
import weakref
from bisect import bisect_right
from concurrent.futures import Future, ProcessPoolExecutor, ThreadPoolExecutor
from copy import copy
from ctypes import c_void_p, cdll, CDLL
from functools import partial
from pathlib import Path
from threading import Thread
from time import sleep, time
from types import ModuleType
from typing import Any, Callable, Dict, List, Optional, Set, Tuple, TYPE_CHECKING, Union
import torch
from torch._dynamo.device_interface import (
get_interface_for_device,
get_registered_device_interfaces,
)
from torch._dynamo.utils import counters, dynamo_timed
from torch._inductor import config, exc, metrics
from torch._inductor.codegen.cuda import cuda_env
from torch._inductor.utils import cache_dir, developer_warning, is_linux
from torch._subclasses.fake_tensor import (
extract_tensor_metadata,
FakeTensor,
TensorMetadata,
)
from torch.fx.experimental.symbolic_shapes import has_hint, hint_int, ShapeEnv
if TYPE_CHECKING:
from torch._inductor.graph import GraphLowering
from torch._inductor.select_algorithm import ChoiceCaller
from torch.hub import _Faketqdm, tqdm
_HERE = os.path.abspath(__file__)
_TORCH_PATH = os.path.dirname(os.path.dirname(_HERE))
_LINKER_SCRIPT = os.path.join(_TORCH_PATH, "_inductor/script.ld")
if config.is_fbcode():
from triton.fb import build_paths
from triton.fb.build import _run_build_command
from torch._inductor.fb.utils import (
log_global_cache_errors,
log_global_cache_stats,
log_global_cache_vals,
use_global_cache,
)
else:
def log_global_cache_errors(*args, **kwargs):
pass
def log_global_cache_stats(*args, **kwargs):
pass
def log_global_cache_vals(*args, **kwargs):
pass
def use_global_cache() -> bool:
return False
LOCK_TIMEOUT = 600
# timing metrics for time spent in the compilation
_cumulative_compile_time = 0.0
_t0: Optional[float] = None
def _compile_start() -> None:
global _t0
if _t0 is None:
_t0 = time()
def _compile_end() -> None:
global _cumulative_compile_time, _t0
if _t0 is not None:
t1 = time()
_cumulative_compile_time += t1 - _t0
_t0 = None
# print("CUMULATIVE COMPILE TIME", _cumulative_compile_time)
log = logging.getLogger(__name__)
def cpp_wrapper_cache_dir(name: str) -> str:
cu_str = (
"cpu"
if torch.version.cuda is None
else f'cu{torch.version.cuda.replace(".", "")}'
)
python_version = f"py{sys.version_info.major}{sys.version_info.minor}"
build_folder = f"{python_version}_{cu_str}"
cpp_wrapper_dir = os.path.join(cache_dir(), build_folder)
cpp_wrapper_build_directory = os.path.join(cpp_wrapper_dir, name)
os.makedirs(cpp_wrapper_build_directory, exist_ok=True)
return cpp_wrapper_build_directory
def get_cpp_wrapper_cubin_path_name():
return "cubin_path" if torch.version.hip is None else "hsaco_path"
class CacheBase:
@staticmethod
@functools.lru_cache(None)
def get_system() -> Dict[str, Any]:
try:
import triton
triton_version = triton.__version__
except ModuleNotFoundError:
triton_version = None
try:
system: Dict[str, Any] = {
"device": {
"name": torch.cuda.get_device_properties(
torch.cuda.current_device()
).name,
},
"version": {
"cuda": torch.version.cuda,
"triton": triton_version,
},
}
except (AssertionError, RuntimeError):
# If cuda is not installed, none of the above config is relevant.
system = {}
system["hash"] = hashlib.sha256(
json.dumps(system, sort_keys=True).encode("utf-8")
).hexdigest()
return system
@staticmethod
@functools.lru_cache(None)
def get_local_cache_path() -> Path:
return Path(os.path.join(cache_dir(), "cache", CacheBase.get_system()["hash"]))
@staticmethod
@functools.lru_cache(None)
def get_global_cache_path() -> Optional[Path]:
return (
Path(os.path.join(config.global_cache_dir, CacheBase.get_system()["hash"]))
if config.global_cache_dir is not None
else None
)
def __init__(self) -> None:
if not torch.cuda.is_available():
return
self.system = CacheBase.get_system()
self.local_cache_path = CacheBase.get_local_cache_path()
self.global_cache_path = CacheBase.get_global_cache_path()
def get_local_cache(self) -> Dict[str, Any]:
if not self.local_cache_path.is_file():
return {}
with open(self.local_cache_path) as local_cache_fp:
local_cache = json.load(local_cache_fp)
return local_cache["cache"]
def update_local_cache(self, local_cache: Dict[str, Any]) -> None:
if not os.path.exists(self.local_cache_path.parent):
os.makedirs(self.local_cache_path.parent, exist_ok=True)
write_atomic(
str(self.local_cache_path),
json.dumps({"system": self.system, "cache": local_cache}, indent=4),
)
class LocalCache(CacheBase):
def lookup(self, *keys: str) -> Optional[Dict[str, Any]]:
cache = self.get_local_cache()
sub_cache = cache
for key in keys:
if key in cache:
sub_cache = cache[key]
else:
return None
return sub_cache
def set_value(self, *keys: str, value: Any) -> None:
cache = self.get_local_cache()
sub_cache = cache
for key in keys[0:-1]:
sub_cache.setdefault(key, {})
sub_cache = sub_cache[key]
sub_cache[keys[-1]] = value
self.update_local_cache(cache)
class PersistentCache(CacheBase):
@functools.lru_cache(None)
def get_global_cache(self):
if self.global_cache_path is None or not self.global_cache_path.is_file():
return {}
with open(self.global_cache_path) as global_cache_fp:
global_cache = json.load(global_cache_fp)
return global_cache["cache"]
def lookup(
self,
choices: List[ChoiceCaller],
op: str,
inputs: str,
benchmark: Callable[[Any], Dict[ChoiceCaller, float]],
) -> Dict[ChoiceCaller, float]:
"""
Check to see if we have benchmarked the given choice callers. For each
choice caller:
1. Check global_cache[op][inputs][choice][precision], return benchmark if cached.
2. Check local_cache[op][inputs][choice][precision], return benchmark if cached.
3.
a. `max_autotune_gemm=True`: benchmark the choice, update
local_cache[op][inputs][choice], and return the benchmark.
b. `max_autotune_gemm=False`: don't benchmark the choice, return nothing.
"""
precision = torch.get_float32_matmul_precision()
log_stats = partial(log_global_cache_stats, self.system, op, inputs, precision)
log_vals = partial(log_global_cache_vals, self.system, op, inputs, precision)
log_errors = partial(
log_global_cache_errors, self.system, op, inputs, precision
)
timings = {}
def check_cache(cache, callback=None) -> bool:
"""Check if `cache` contains data for all the choices"""
hit = True
for choice in choices:
choice_hash = choice.hash_key()
if choice_hash in cache.get(op, {}).get(inputs, {}).get(precision, {}):
# cache hit
timings[choice] = cache[op][inputs][precision][choice_hash]
else:
# cache miss
hit = False
break
if callback:
callback(cached=hit)
return hit
if config.max_autotune or config.max_autotune_gemm:
local_cache = self.get_local_cache()
# check local cache first since it is data specific to the current machine
if not check_cache(local_cache) and not (
use_global_cache()
and check_cache(self.get_global_cache(), callback=log_stats)
):
try:
# re-benchmark everything to try to get consistent numbers from the same machine
timings = benchmark(choices)
assert all(choice in timings for choice in choices)
local_cache.setdefault(op, {})
local_cache[op].setdefault(inputs, {}).setdefault(precision, {})
for choice, timing in timings.items():
local_cache[op][inputs][precision][choice.hash_key()] = timing
except RuntimeError as e:
# catch and log autotuning failures
log_errors(e)
raise e
self.update_local_cache(local_cache)
timings_to_log = {
choice.hash_key(): timings[choice] for choice in choices
}
log_vals(timings_to_log)
elif use_global_cache():
# only check global cache, not local one
check_cache(self.get_global_cache(), callback=log_stats)
# may have a partial cache hit, where not everything is benchmarked
return timings
def get_lock_dir() -> str:
lock_dir = os.path.join(cache_dir(), "locks")
if not os.path.exists(lock_dir):
os.makedirs(lock_dir, exist_ok=True)
return lock_dir
def sha256_hash(data: bytes) -> str:
# [:51] to strip off the "Q====" suffix common to every hash value.
return base64.b32encode(hashlib.sha256(data).digest())[:51].decode("utf-8").lower()
def code_hash(code: Union[str, bytes], extra: str = ""):
hashing_str = code if isinstance(code, bytes) else code.encode("utf-8")
if extra != "":
hashing_str = hashing_str + b"||" + extra.encode("utf-8")
return "c" + sha256_hash(hashing_str)
def get_path(
basename: str, extension: str, specified_dir: str = ""
) -> Tuple[str, str, str]:
if specified_dir:
if os.path.isabs(specified_dir):
subdir = specified_dir
else:
subdir = os.path.join(cache_dir(), specified_dir)
else:
subdir = os.path.join(cache_dir(), basename[1:3])
path = os.path.join(subdir, f"{basename}.{extension}")
return basename, subdir, path
def get_hash(content: Union[str, bytes], extra: str = "", hash_type: str = "code"):
if hash_type == "code":
return code_hash(content, extra)
if hash_type in ["cubin", "hsaco"]:
return code_hash(repr(content))
raise AssertionError(f"Unknown hash type {hash_type}")
def write(
content: Union[str, bytes],
extension: str,
extra: str = "",
hash_type: str = "code",
specified_dir: str = "",
) -> Tuple[str, str]:
# use striped content to compute hash so we don't end up with different
# hashes just because the content begins/ends with differnet number of
# spaces.
key: str = get_hash(content.strip(), extra, hash_type)
basename, subdir, path = get_path(key, extension, specified_dir)
if not os.path.exists(subdir):
os.makedirs(subdir, exist_ok=True)
if not os.path.exists(path):
write_atomic(path, content)
return basename, path
def write_atomic(path: str, content: Union[str, bytes]) -> None:
# Write into temporary file first to avoid conflicts between threads
# Avoid using a named temporary file, as those have restricted permissions
assert isinstance(
content, (str, bytes)
), "Only strings and byte arrays can be saved in the cache"
path = pathlib.Path(path)
tmp_path = path.parent / f".{os.getpid()}.{threading.get_ident()}.tmp"
write_mode = "w" if isinstance(content, str) else "wb"
with tmp_path.open(write_mode) as f:
f.write(content)
tmp_path.rename(path)
@dataclasses.dataclass
class TensorMetadataAndValues:
"""
TensorMetadata plus the elements as a list of raw values.
Used for hashing inlined constants.
"""
tensor_metadata: TensorMetadata
values: List[Any]
def _ident(x: Any) -> Any:
return x
def _reduce_fake_tensor(t):
"""
See FxGraphCachePickler. Custom reducer to pickle FakeTensors.
"""
metadata = extract_tensor_metadata(t)
return (_ident, (metadata,))
def _reduce_tensor(t):
"""
See FxGraphCachePickler. Custom reducer to pickle Tensors.
"""
if t.is_mkldnn:
# TODO: These tensors don't currently pickle, so we can't cache a
# compiled graph containing them. Just fail now. If mkldnn tensors
# get pickling support, we can remove this.
raise BypassFxGraphCache()
# If we see tensors, we know they're constants stored as attributes on
# the GraphModule. See tensor lowering; small constants are inlined. If
# we see a small tensor, therefore, no reference will ultimately remain
# in the generated code. So we need to include its value in the cache key.
# Large constants are effectively treated as inputs and we consider only
# their metadata.
metadata = extract_tensor_metadata(t)
if len(t.shape) == 0 or torch._inductor.graph.GraphLowering.can_inline_constant(t):
return (_ident, (TensorMetadataAndValues(metadata, t.tolist()),))
else:
return (_ident, (metadata,))
def _reduce_symint(s):
"""
See FxGraphCachePickler. Custom reducer to pickle SymInts.
"""
# For hashing purposes, we only care about the name of the symbol and
# not the backed value. We evaluate guards stored with a cached graph
# to ensure a cached entity with SymInt args is safe to reuse.
return (_ident, (str(s),))
class FxGraphCachePickler(pickle.Pickler):
"""
Custom pickler to customize the pickling of some objects (Tensors), only for the
purpose of computing a hash for keying into the FxGraphCache. Tensors contain
objects that don't pickle and/or vary between runs, and we want to capture the
data that allow us to compute a stable, but safe hash.
"""
dispatch_table = copyreg.dispatch_table.copy()
dispatch_table[FakeTensor] = _reduce_fake_tensor
dispatch_table[torch.Tensor] = _reduce_tensor
dispatch_table[torch.SymInt] = _reduce_symint
@staticmethod
def dumps(obj) -> bytes:
"""
Pickle an object using the FxGraphCachePickler.
"""
with io.BytesIO() as stream:
pickler = FxGraphCachePickler(stream)
pickler.dump(obj)
return stream.getvalue()
@staticmethod
def get_hash(obj: Any) -> str:
"""
Serialize an object using the FxGraphCachePickler and return a hash
of the pickled object.
"""
serialized_data = FxGraphCachePickler.dumps(obj)
return sha256_hash(serialized_data)
@functools.lru_cache(None)
def get_inductor_code_hash() -> bytes:
"""
Compute a hash of all inductor code modules. Used by the FxGraph cache
so any inductor code changes would result in new cache keys.
"""
inductor_root = os.path.dirname(__file__)
contents: Dict[str, bytes] = {}
for lib in pkgutil.iter_modules([inductor_root]):
spec = lib.module_finder.find_spec(lib.name, None)
assert spec is not None
module = spec.origin
assert module is not None
with open(module, "rb") as f:
contents[module] = f.read()
return hashlib.sha256(pickle.dumps(contents)).digest()
@dataclasses.dataclass
class OrderedSetHolder:
"""
See FxGraphHashDetails. Holds a sorted list to support stable hashing
of set kwargs.
"""
items: List[Any]
class BypassFxGraphCache(Exception):
"""
Exception to indicate that the FxGraphCache should be bypassed.
"""
pass
class FxGraphHashDetails:
"""
Object to capture all the details for a compiled FX graph relevant to computing
a safe and stable cache key.
"""
# Excluded kwargs param that are not stable between runs
EXCLUDED_KWARGS = ["graph_id"]
def __init__(
self,
gm: torch.fx.GraphModule,
example_inputs: List[torch.Tensor],
fx_kwargs: Dict[str, Any],
):
self.gm = gm
self.example_inputs = example_inputs
# Order kwargs so hashing is stable to changes in kwarg order.
self.fx_kwargs = {}
for k in sorted(fx_kwargs):
if k not in self.EXCLUDED_KWARGS:
if type(fx_kwargs[k]) is set:
# Special case to handle set params. Python sets can't be
# ordered, so sort the elements and store them in a proxy.
self.fx_kwargs[k] = OrderedSetHolder(sorted(fx_kwargs[k]))
else:
self.fx_kwargs[k] = fx_kwargs[k]
# 'Deterministic algorithms' can affect codegen via lowering to cuda kernels.
self.deterministic_algorithms_settings = (
torch.are_deterministic_algorithms_enabled(),
torch.is_deterministic_algorithms_warn_only_enabled(),
torch.utils.deterministic.fill_uninitialized_memory, # type: ignore[attr-defined]
)
# Global settings affecting matmul codegen.
self.cuda_matmul_settings = (
torch.backends.cuda.matmul.allow_tf32,
torch.backends.cuda.matmul.allow_fp16_reduced_precision_reduction,
torch.backends.cuda.matmul.allow_bf16_reduced_precision_reduction,
)
# Also hash on various system info (including the triton compiler version).
self.torch_version = torch.__version__
self.system_info = CacheBase.get_system()
# And the inductor configuration and code.
self.inductor_code_hash = get_inductor_code_hash()
try:
self.inductor_config = config.save_config()
except TypeError as e:
# Some configs options are callables, e.g., post_grad_custom_pre_pass,
# and may not pickle.
log.debug("Can't pickle inductor config: %s", e)
raise BypassFxGraphCache() from e
def debug_str(self) -> str:
"""
Get a printable string describing in more detail all the attributes
comprising this object. Useful for debugging when one graph hashes
to a different value than another.
"""
def get_str(obj) -> str:
if isinstance(obj, torch.Tensor):
return str(extract_tensor_metadata(obj))
elif isinstance(obj, bytes):
return "<bytes>"
else:
return str(obj)
lines = []
for attr, obj in vars(self).items():
if isinstance(obj, list):
for ii in range(len(obj)):
h = FxGraphCachePickler.get_hash(obj[ii])
lines.append(f"[{h}] {attr}[{ii}]: {get_str(obj[ii])}")
elif isinstance(obj, dict):
for k, v in obj.items():
h = FxGraphCachePickler.get_hash(v)
lines.append(f"[{h}] {attr}[{k}]: {get_str(v)}")
else:
h = FxGraphCachePickler.get_hash(obj)
lines.append(f"[{h}] {attr}: {get_str(obj)}")
return "\n".join(lines)
def compiled_fx_graph_hash(
gm: torch.fx.GraphModule,
example_inputs: List[torch.Tensor],
fx_kwargs: Dict[str, Any],
) -> str:
"""
Generate a unique hash of the FX graph for caching.
"""
details = FxGraphHashDetails(gm, example_inputs, fx_kwargs)
# The prefix distinguishes among the other kinds of objects we
# cache in this module.
key = "f" + FxGraphCachePickler.get_hash(details)
log.debug("FX graph cache hash details for key %s:\n%s", key, details.debug_str())
return key
class FxGraphCache:
"""
Supports caching and reusing compiled Fx graphs.
The overall strategy is as follows:
- This cache stores entries on disk. When saving an entry, we can't
serialize callables (that could be C++, Triton, etc.), so we serialize
their own disk cache location. We then recreate the compiled artifact
after fetching from disk.
- For indexing the cache, we gather the fields relevant to identifying an
FxGraph (the graph module, graph inputs, system settings etc.) into an
FxGraphCacheDetails object, pickle it, and compute a hash for the key.
See FxGraphCachePickler.
- Among the metadata we store, we also include a guards expression that's
appropriate for validating any symbols for Tensor arguments that have
symbolic bounds. On cache lookup then, we evaluate those guards in the
current context to validate that a cached entry can be served.
- A given graph could have multiple compiled versions, corresponding to
different sets of guards. Therefore, we store cache entries in the form:
<temp dir>/<fx graph hash>/<serialized metatdata>
- On lookup, we compute the key from the graph details, iterate over all
leaf files in the corresponding subdirectory, deserialize the entry, and
evaluate its guards expression. If the evaluation succeeds, we have a
cache hit. If it fails, we compile the graph and store a new entry.
- Finally, on a cache hit, we need to make sure any guards that would
have been created during compilation are added to the current context.
"""
# TODO(masnesral): Investigate whether it's beneficial to store compiled graphs
# in an in-memory cache after loading from disk.
@staticmethod
def _get_tmp_dir() -> str:
"""
Get the toplevel temporary directory for storing compiled graphs.
"""
return os.path.join(cache_dir(), "fxgraph")
@staticmethod
def _get_tmp_dir_for_key(key: str) -> str:
"""
Return the disk location for a given cache key.
"""
return os.path.join(FxGraphCache._get_tmp_dir(), key[1:3], key)
@staticmethod
def _filter_symints(inputs: List[Any]) -> List[torch.SymInt]:
"""
Get the SymInt objects from the input list.
"""
return [s for s in inputs if isinstance(s, torch.SymInt)]
@staticmethod
def _get_shape_env() -> Optional[ShapeEnv]:
"""
Helper to get the shape env from the tracing context.
"""
ctx = torch._guards.TracingContext.try_get()
if not ctx:
return None
return ctx.fake_mode.shape_env
@staticmethod
def _lookup_graph(
key: str,
example_inputs: List[torch.Tensor],
) -> Optional[CompiledFxGraph]:
"""
Lookup a compiled graph in the cache by key. On a hit, return the
deserialized CompiledFxGraph object. On a miss, return None.
"""
subdir = FxGraphCache._get_tmp_dir_for_key(key)
if not os.path.exists(subdir):
return None
shape_env = FxGraphCache._get_shape_env()
assert shape_env is not None
# Iterate over any entries in the subdir for this key and evaluate
# their guards to determine whether there's a hit.
graph = None
for path in sorted(os.listdir(subdir)):
with open(os.path.join(subdir, path), "rb") as f:
candidate: CompiledFxGraph = pickle.load(f)
guards_expr = candidate.guards_expr
if not guards_expr:
# No guards to evaluate, so this is a hit.
graph = candidate
break
# Evaluate the guard expression in the current context.
symints = FxGraphCache._filter_symints(example_inputs)
# If there's not a cache hit, we don't want the evaluation to
# affect the current env, e.g., cause the creation of new guards,
# so we evaluate with the hints instead of the symbols.
assert all(has_hint(s) for s in symints)
hints = [hint_int(s) for s in symints]
hit = bool(shape_env.evaluate_guards_expression(guards_expr, hints))
log.debug(
"fx graph cache key %s evaluating guards for %s with values %s => %s",
key,
guards_expr,
hints,
hit,
)
if hit:
# Now re-evaluate with the symints to add any guards to the current env.
check = bool(shape_env.evaluate_guards_expression(guards_expr, symints))
assert check is True
log.debug(
"fx graph cache key %s post-load guards: %s", key, shape_env.guards
)
graph = candidate
break
# Increment the cached metrics by the amounts recorded when the FX
# graph was compiled for this cache entry. Pretending these counters
# were incremented normally is useful for testing with the cache enabled.
if graph is not None:
metrics.CachedMetricsHelper.apply_deltas(graph.metrics_deltas)
return graph
@staticmethod
def _save_graph(
key: str, compiled_graph: CompiledFxGraph, example_inputs: List[torch.Tensor]
):
"""
Store a serialized CompiledFxGraph on disk.
"""
disk_compiled_graph = copy(compiled_graph)
# Important as compiled models are not pickleable:
disk_compiled_graph.compiled_artifact = None
# Before serializing, compute the guard expression that will be used to
# ensure that a CompiledFxGraph is valid when loaded from the cache. It's
# sufficient to consider only the SymInt args to the fx graph since the
# Tensor shapes are already captured in the hash for the cache key. Any
# Tensor arg with a symbolic shape will have a SymInt arg for the graph.
shape_env = FxGraphCache._get_shape_env()
assert shape_env is not None
symints = FxGraphCache._filter_symints(example_inputs)
disk_compiled_graph.guards_expr = shape_env.produce_guards_expression(symints)
try:
content = pickle.dumps(disk_compiled_graph)
except Exception as e:
log.debug("fx graph cache unable to serialize compiled graph: %s", e)
counters["inductor"]["fxgraph_cache_pickle_error"] += 1
return
subdir = FxGraphCache._get_tmp_dir_for_key(key)
if not os.path.exists(subdir):
os.makedirs(subdir, exist_ok=True)
# Use a hash of the serialized CompiledFxGraph to get a unique file
# name. The specific name doesn't matter since a lookup involves
# iterating over all entries in the parent subdir.
path = os.path.join(subdir, sha256_hash(content))
write_atomic(path, content)
@staticmethod
def _check_can_cache():
"""
Check some conditions that would preclude caching and raise BypassFxGraphCache
to bypass in case caching is not possible.
"""
if config.freezing or config.aot_inductor.use_runtime_constant_folding:
# Freezing can embed constants that wouldn't be static across runs.
raise BypassFxGraphCache()
if FxGraphCache._get_shape_env() is None:
# The treatment of guards in the caching implementation requires that
# we have a shape env.
log.debug("fx graph cache no shape env")
raise BypassFxGraphCache()
@staticmethod
def load(
compile_fx_fn: Callable[..., Any],
gm: torch.fx.GraphModule,
example_inputs: List[torch.Tensor],
fx_kwargs: Dict[str, Any],
):
"""
Load a compiled graph from the cache. If a cached entry does not exist,
compile the graph and save it to the cache.
"""
from filelock import FileLock
compiled_graph = None
try:
FxGraphCache._check_can_cache()
key = compiled_fx_graph_hash(gm, example_inputs, fx_kwargs)
lock_path = os.path.join(get_lock_dir(), key + ".lock")
with FileLock(lock_path, timeout=LOCK_TIMEOUT):
compiled_graph = FxGraphCache._lookup_graph(key, example_inputs)
if compiled_graph is None:
log.debug("fx graph cache miss for key %s", key)
counters["inductor"]["fxgraph_cache_miss"] += 1
compiled_graph = compile_fx_fn(gm, example_inputs, **fx_kwargs)
FxGraphCache._save_graph(key, compiled_graph, example_inputs)
else:
log.debug("fx graph cache hit for key %s", key)
counters["inductor"]["fxgraph_cache_hit"] += 1
except BypassFxGraphCache:
counters["inductor"]["fxgraph_cache_bypass"] += 1
if not compiled_graph:
compiled_graph = compile_fx_fn(gm, example_inputs, **fx_kwargs)
return compiled_graph
@staticmethod
def clear():
"""
Clear out the on-disk cache.
"""
try:
shutil.rmtree(FxGraphCache._get_tmp_dir())
except FileNotFoundError:
pass
@dataclasses.dataclass
class CompiledFxGraph:
"""
Class holding a compiled FX graph. This is the object serialized on disk
to support FxGraph caching.
"""
compiled_artifact: Optional[Callable[..., Any]]
current_callable: Optional[Callable[..., Any]]
cache_key: Optional[str]
artifact_path: Optional[str]
cache_linemap: Optional[List[Tuple[int, str]]]
device_types: Set[str]
device_idxs: Set[int]
mutated_inputs: Set[str]
mutated_input_idxs: Set[int]
constants: Dict[str, torch.Tensor]
output_strides: Optional[List[Optional[Tuple[int, ...]]]]
disabled_cudagraphs_reason: Optional[str]
metrics_deltas: metrics.CachedMetricsDeltas
# This is a string representation of an expression we serialize
# with the object so the guards can be evaluated in a different
# context in order to verify the validity of serving a cached
# fx graph. The expression must be generated by:
# ShapeEnv.produce_guards_expression()
guards_expr: Optional[str]
_boxed_call: Optional[bool] = None
def __init__(
self,
compiled_artifact: Optional[Callable[..., Any]],
graph: GraphLowering,
output_strides: List[Optional[Tuple[int, ...]]],
disabled_cudagraphs_reason: Optional[str],
metrics_deltas: metrics.CachedMetricsDeltas,
):
self.compiled_artifact = compiled_artifact
self.current_callable = None
self.cache_key = graph.cache_key
self.artifact_path = graph.cache_path
self.cache_linemap = graph.cache_linemap
self.device_types = graph.device_types
self.device_idxs = graph.device_idxs
self.mutated_inputs = graph.mutated_inputs
self.mutated_input_idxs = set(graph.mutated_input_idxs)
self.constants = graph.constants
self.output_strides = output_strides
self.disabled_cudagraphs_reason = disabled_cudagraphs_reason
self.metrics_deltas = metrics_deltas
self.guards_expr = None
def __call__(self, inputs: List[Any]) -> Any:
return self.get_current_callable()(inputs)
def get_current_callable(self) -> Callable[..., Any]:
if self.current_callable is None:
# This prevents a circular reference that makes CompiledFxGraph
# get stuck without getting garbage collected
return functools.partial(_run_from_cache, weakref.proxy(self))
else:
return self.current_callable
def _run_from_cache(compiled_graph: CompiledFxGraph, inputs: List[Any]) -> Any:
# We can't really serialize callables that may be C++/Triton/etc.,
# so we serialize their disk cache location instead
# TODO: When making an API that can save compiled models e2e to disk
# this will need to be better
if compiled_graph.compiled_artifact is None:
from .codecache import PyCodeCache
assert compiled_graph.cache_key
assert compiled_graph.artifact_path
compiled_graph.compiled_artifact = PyCodeCache.load_by_key_path(
compiled_graph.cache_key,
compiled_graph.artifact_path,
compiled_graph.cache_linemap,
compiled_graph.constants,
).call
return compiled_graph.compiled_artifact(inputs)
def cpp_compiler() -> str:
if config.is_fbcode():
return build_paths.cc()
if isinstance(config.cpp.cxx, (list, tuple)):
search = tuple(config.cpp.cxx)
else:
search = (config.cpp.cxx,)
return cpp_compiler_search(search)
@functools.lru_cache(1)
def cpp_compiler_search(search: str) -> str:
for cxx in search:
try:
if cxx is None:
# gxx package is only available for Linux
# according to https://anaconda.org/conda-forge/gxx/
if sys.platform != "linux":
continue
# Do not install GXX by default
if not os.getenv("TORCH_INDUCTOR_INSTALL_GXX"):
continue
from filelock import FileLock
lock_dir = get_lock_dir()
lock = FileLock(
os.path.join(lock_dir, "g++.lock"), timeout=LOCK_TIMEOUT
)
with lock:
cxx = install_gcc_via_conda()
subprocess.check_output([cxx, "--version"])
return cxx
except (subprocess.SubprocessError, FileNotFoundError, ImportError):
continue
raise exc.InvalidCxxCompiler()
def install_gcc_via_conda() -> str:
"""On older systems, this is a quick way to get a modern compiler"""
prefix = os.path.join(cache_dir(), "gcc")
cxx_path = os.path.join(prefix, "bin", "g++")
if not os.path.exists(cxx_path):
log.info("Downloading GCC via conda")
conda = os.environ.get("CONDA_EXE", "conda")
if conda is None:
conda = shutil.which("conda")
if conda is not None:
subprocess.check_call(
[
conda,
"create",
f"--prefix={prefix}",
"--channel=conda-forge",
"--quiet",
"-y",
"python=3.8",
"gxx",
],
stdout=subprocess.PIPE,
)
return cxx_path
def is_gcc() -> bool:
return bool(re.search(r"(gcc|g\+\+)", cpp_compiler()))
def is_clang() -> bool:
return bool(re.search(r"(clang|clang\+\+)", cpp_compiler()))
@functools.lru_cache(None)
def is_apple_clang() -> bool:
cxx = cpp_compiler()
version_string = subprocess.check_output([cxx, "--version"]).decode("utf8")
return "Apple" in version_string.splitlines()[0]
class VecISA:
_bit_width: int
_macro: str
_arch_flags: str
_dtype_nelements: Dict[torch.dtype, int]
# Note [Checking for Vectorized Support in Inductor]
# TorchInductor CPU vectorization reuses PyTorch vectorization utility functions
# Hence, TorchInductor would depend on Sleef* to accelerate mathematical functions
# like exp, pow, sin, cos and etc.
# But PyTorch and TorchInductor might use different compilers to build code. If
# PyTorch uses gcc-7/g++-7 to build the release package, the libtorch_cpu.so
# will not expose the Sleef* AVX512 symbols since gcc-7/g++-7 cannot pass
# avx512 check in CMake - FindAVX.cmake. But TorchInductor install the latest
# gcc/g++ compiler by default while it could support the AVX512 compilation.
# Therefore, there would be a conflict sleef version between PyTorch and
# TorchInductor. Hence, we dry-compile the following code to check whether current
# HW platform and PyTorch both could support AVX512 or AVX2. And suppose ARM
# also needs the logic
# In fbcode however, we are using the same compiler for pytorch and for inductor codegen,
# making the runtime check unnecessary.
_avx_code = """
#if defined(CPU_CAPABILITY_AVX512) || defined(CPU_CAPABILITY_AVX2) || defined(CPU_CAPABILITY_ZVECTOR)
#include <ATen/cpu/vec/functional.h>
#include <ATen/cpu/vec/vec.h>
#endif
__attribute__((aligned(64))) float in_out_ptr0[16] = {0.0};
extern "C" void __avx_chk_kernel() {
auto tmp0 = at::vec::Vectorized<float>(1);
auto tmp1 = tmp0.exp();
tmp1.store(in_out_ptr0);
}
""" # noqa: B950
_avx_py_load = """
import torch
from ctypes import cdll
cdll.LoadLibrary("__lib_path__")
"""
def bit_width(self) -> int:
return self._bit_width
def nelements(self, dtype: torch.dtype = torch.float) -> int:
return self._dtype_nelements[dtype]
def build_macro(self) -> str:
return self._macro
def build_arch_flags(self) -> str:
return self._arch_flags
def __hash__(self) -> int:
return hash(str(self))
@functools.lru_cache(None)
def __bool__(self) -> bool:
if config.cpp.vec_isa_ok is not None:
return config.cpp.vec_isa_ok
if config.is_fbcode():
return True
key, input_path = write(VecISA._avx_code, "cpp")
from filelock import FileLock
lock_dir = get_lock_dir()
lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT)
with lock:
output_path = input_path[:-3] + "so"
build_cmd = shlex.split(
cpp_compile_command(
input_path, output_path, warning_all=False, vec_isa=self
)
)
try:
# Check build result
compile_file(input_path, output_path, build_cmd)
subprocess.check_call(
[
sys.executable,
"-c",
VecISA._avx_py_load.replace("__lib_path__", output_path),
],
stderr=subprocess.DEVNULL,
env={**os.environ, "PYTHONPATH": ":".join(sys.path)},
)
except Exception as e:
return False
return True
@dataclasses.dataclass
class VecAVX512(VecISA):
_bit_width = 512
_macro = "-DCPU_CAPABILITY_AVX512"
_arch_flags = "-mavx512f -mavx512dq -mavx512vl -mavx512bw -mfma"
_dtype_nelements = {torch.float: 16, torch.bfloat16: 32, torch.float16: 32}
def __str__(self) -> str:
return "avx512"
__hash__: Callable[[VecISA], Any] = VecISA.__hash__
@dataclasses.dataclass
class VecAVX2(VecISA):
_bit_width = 256
_macro = "-DCPU_CAPABILITY_AVX2"
_arch_flags = "-mavx2 -mfma"
_dtype_nelements = {torch.float: 8, torch.bfloat16: 16, torch.float16: 16}
def __str__(self) -> str:
return "avx2"
__hash__: Callable[[VecISA], Any] = VecISA.__hash__
@dataclasses.dataclass
class VecZVECTOR(VecISA):
_bit_width = 256
_macro = "-DCPU_CAPABILITY_ZVECTOR -DCPU_CAPABILITY=ZVECTOR -DHAVE_ZVECTOR_CPU_DEFINITION"
_arch_flags = "-mvx -mzvector"
_dtype_nelements = {torch.float: 8, torch.bfloat16: 16, torch.float16: 16}
def __str__(self) -> str:
return "zvector"
__hash__: Callable[[VecISA], Any] = VecISA.__hash__
class InvalidVecISA(VecISA):
_bit_width = 0
_macro = ""
_arch_flags = ""
_dtype_nelements = {}
def __str__(self) -> str:
return "INVALID_VEC_ISA"
def __bool__(self) -> bool: # type: ignore[override]
return False
__hash__: Callable[[VecISA], Any] = VecISA.__hash__
invalid_vec_isa = InvalidVecISA()
supported_vec_isa_list = [VecAVX512(), VecAVX2()]
# Cache the cpuinfo to avoid I/O overhead. Meanwhile, the cpuinfo content
# might have too much redundant content that is useless for ISA check. Hence,
# we only cache some key isa information.
@functools.lru_cache(None)
def valid_vec_isa_list() -> List[VecISA]:
if sys.platform != "linux":
return []
if platform.machine() == "s390x":
return [VecZVECTOR()]
isa_list = []
with open("/proc/cpuinfo") as _cpu_info:
_cpu_info_content = _cpu_info.read()
for isa in supported_vec_isa_list:
if str(isa) in _cpu_info_content and isa:
isa_list.append(isa)
return isa_list
def pick_vec_isa() -> VecISA:
if config.is_fbcode():
return VecAVX2()
_valid_vec_isa_list: List[VecISA] = valid_vec_isa_list()
if not _valid_vec_isa_list:
return invalid_vec_isa
# If the simdlen is None, it indicates determin the vectorization length automatically
if config.cpp.simdlen is None:
assert _valid_vec_isa_list
return _valid_vec_isa_list[0]
for isa in _valid_vec_isa_list:
if config.cpp.simdlen == isa.bit_width():
return isa
return invalid_vec_isa
def get_compile_only(compile_only: bool = True) -> str:
return "-c" if compile_only else ""
def get_shared(shared: bool = True, compile_only: bool = False) -> str:
if not shared:
return ""
if compile_only:
return "-fPIC"
if platform.system() == "Darwin" and "clang" in cpp_compiler():
# This causes undefined symbols to behave the same as linux
return "-shared -fPIC -undefined dynamic_lookup"
else:
return "-shared -fPIC"
def get_warning_all_flag(warning_all: bool = True) -> str:
return "-Wall" if warning_all else ""
def get_glibcxx_abi_build_flags() -> str:
return "-D_GLIBCXX_USE_CXX11_ABI=" + str(int(torch._C._GLIBCXX_USE_CXX11_ABI))
def cpp_flags() -> str:
flags = ["-std=c++17", "-Wno-unused-variable", "-Wno-unknown-pragmas"]
if is_clang():
flags.append("-Werror=ignored-optimization-argument")
return " ".join(flags)
def cpp_wrapper_flags() -> str:
return "-DTORCH_INDUCTOR_CPP_WRAPPER"
def optimization_flags() -> str:
base_flags = "-O0 -g" if config.aot_inductor.debug_compile else "-O3 -DNDEBUG"
base_flags += " -ffast-math -fno-finite-math-only"
if not config.cpp.enable_unsafe_math_opt_flag:
base_flags += " -fno-unsafe-math-optimizations"
if not config.cpp.enable_floating_point_contract_flag:
base_flags += " -ffp-contract=off"
if config.is_fbcode():
# FIXME: passing `-fopenmp` adds libgomp.so to the generated shared library's dependencies.
# This causes `ldopen` to fail in fbcode, because libgomp does not exist in the default paths.
# We will fix it later by exposing the lib path.
return base_flags
if sys.platform == "darwin":
# Per https://mac.r-project.org/openmp/ right way to pass `openmp` flags to MacOS is via `-Xclang`
# Also, `-march=native` is unrecognized option on M1
base_flags += " -Xclang"
else:
if platform.machine() == "ppc64le":
base_flags += " -mcpu=native"
else:
base_flags += " -march=native"
# Internal cannot find libgomp.so
if not config.is_fbcode():
base_flags += " -fopenmp"
return base_flags
def use_custom_generated_macros() -> str:
return "-D C10_USING_CUSTOM_GENERATED_MACROS"
def use_fb_internal_macros() -> str:
if config.is_fbcode():
openmp_lib = build_paths.openmp_lib()
preprocessor_flags = " ".join(
(
"-D C10_USE_GLOG",
"-D C10_USE_MINIMAL_GLOG",
"-D C10_DISABLE_TENSORIMPL_EXTENSIBILITY",
)
)
return f"-Wp,-fopenmp {openmp_lib} {preprocessor_flags}"
else:
return ""
def use_standard_sys_dir_headers() -> str:
if config.is_fbcode():
return "-nostdinc"
else:
return ""
@functools.lru_cache(None)
def is_conda_llvm_openmp_installed() -> bool:
try:
command = "conda list llvm-openmp --json"
output = subprocess.check_output(command.split()).decode("utf8")
return len(json.loads(output)) > 0
except subprocess.SubprocessError:
return False
@functools.lru_cache(None)
def homebrew_libomp() -> Tuple[bool, str]:
try:
# check if `brew` is installed
subprocess.check_output(["which", "brew"])
# get the location of `libomp` if it is installed
# this is the location that `libomp` **would** be installed
# see https://github.com/Homebrew/brew/issues/10261#issuecomment-756563567 for details
libomp_path = (
subprocess.check_output(["brew", "--prefix", "libomp"])
.decode("utf8")
.strip()
)
# check if `libomp` is installed
omp_available = os.path.exists(libomp_path)
return omp_available, libomp_path
except subprocess.SubprocessError:
return False, ""
def get_include_and_linking_paths(
include_pytorch: bool = False,
vec_isa: VecISA = invalid_vec_isa,
cuda: bool = False,
aot_mode: bool = False,
) -> Tuple[List[str], str, str, str, str]:
if (
config.is_fbcode()
and "CUDA_HOME" not in os.environ
and "CUDA_PATH" not in os.environ
):
os.environ["CUDA_HOME"] = os.path.dirname(build_paths.cuda())
from torch.utils import cpp_extension
macros = ""
build_arch_flags = ""
if sys.platform == "linux" and (
include_pytorch
or vec_isa != invalid_vec_isa
or cuda
or config.cpp.enable_kernel_profile
):
# Note - We include pytorch only on linux right now. There is more work
# to do to enable OMP build on darwin where PyTorch is built with IOMP
# and we need a way to link to what PyTorch links.
ipaths = cpp_extension.include_paths(cuda) + [sysconfig.get_path("include")]
lpaths = cpp_extension.library_paths(cuda) + [
sysconfig.get_config_var("LIBDIR")
]
libs = []
# No need to manually specify libraries in fbcode.
if not config.is_fbcode():
libs += ["torch", "torch_cpu"]
libs += ["gomp"]
if not aot_mode:
libs += ["torch_python"]
else:
# internal remote execution is able to find omp, but not gomp
libs += ["omp"]
if aot_mode:
ipaths += [os.path.dirname(cpp_prefix_path())]
if cuda:
# This is a special treatment for Meta internal cuda-12 where all libs
# are in lib/cuda-12 and lib/cuda-12/stubs
for i, path in enumerate(lpaths):
if path.startswith(
os.environ["CUDA_HOME"]
) and not os.path.exists(f"{path}/libcudart_static.a"):
for root, dirs, files in os.walk(path):
if "libcudart_static.a" in files:
lpaths[i] = os.path.join(path, root)
lpaths.append(os.path.join(lpaths[i], "stubs"))
break
macros = vec_isa.build_macro()
if macros:
if config.is_fbcode() and vec_isa != invalid_vec_isa:
cap = str(vec_isa).upper()
macros = " ".join(
[
vec_isa.build_arch_flags(),
f"-D CPU_CAPABILITY={cap}",
f"-D CPU_CAPABILITY_{cap}",
f"-D HAVE_{cap}_CPU_DEFINITION",
]
)
if cuda:
if macros is None:
macros = ""
macros += " -D USE_ROCM" if torch.version.hip else " -D USE_CUDA"
if cuda:
if torch.version.hip is not None:
libs += ["c10_hip", "torch_hip"]
macros += " -D __HIP_PLATFORM_AMD__"
else:
if config.is_fbcode():
libs += ["cuda"]
else:
libs += ["c10_cuda", "cuda", "torch_cuda"]
build_arch_flags = vec_isa.build_arch_flags()
else:
# Note - this is effectively a header only inclusion. Usage of some header files may result in
# symbol not found, if those header files require a library.
# For those cases, include the lpath and libs command as we do for pytorch above.
# This approach allows us to only pay for what we use.
ipaths = cpp_extension.include_paths(cuda) + [sysconfig.get_path("include")]
if aot_mode:
ipaths += [os.path.dirname(cpp_prefix_path())]
lpaths = []
if sys.platform == "darwin":
# only Apple builtin compilers (Apple Clang++) require openmp
omp_available = not is_apple_clang()
# check the `OMP_PREFIX` environment first
if os.getenv("OMP_PREFIX") is not None:
header_path = os.path.join(os.getenv("OMP_PREFIX"), "include", "omp.h") # type: ignore[arg-type]
valid_env = os.path.exists(header_path)
if valid_env:
ipaths.append(os.path.join(os.getenv("OMP_PREFIX"), "include")) # type: ignore[arg-type]
lpaths.append(os.path.join(os.getenv("OMP_PREFIX"), "lib")) # type: ignore[arg-type]
else:
warnings.warn("environment variable `OMP_PREFIX` is invalid.")
omp_available = omp_available or valid_env
libs = [] if omp_available else ["omp"]
# prefer to use openmp from `conda install llvm-openmp`
if not omp_available and os.getenv("CONDA_PREFIX") is not None:
omp_available = is_conda_llvm_openmp_installed()
if omp_available:
conda_lib_path = os.path.join(os.getenv("CONDA_PREFIX"), "lib") # type: ignore[arg-type]
ipaths.append(os.path.join(os.getenv("CONDA_PREFIX"), "include")) # type: ignore[arg-type]
lpaths.append(conda_lib_path)
# Prefer Intel OpenMP on x86 machine
if os.uname().machine == "x86_64" and os.path.exists(
os.path.join(conda_lib_path, "libiomp5.dylib")
):
libs = ["iomp5"]
# next, try to use openmp from `brew install libomp`
if not omp_available:
omp_available, libomp_path = homebrew_libomp()
if omp_available:
ipaths.append(os.path.join(libomp_path, "include"))
lpaths.append(os.path.join(libomp_path, "lib"))
# if openmp is still not available, we let the compiler to have a try,
# and raise error together with instructions at compilation error later
else:
libs = ["omp"] if config.is_fbcode() else ["gomp"]
# Unconditionally import c10 for non-abi-compatible mode to use TORCH_CHECK - See PyTorch #108690
if not config.abi_compatible:
libs += ["c10"]
lpaths += [cpp_extension.TORCH_LIB_PATH]
# third party libs
if config.is_fbcode():
ipaths.append(build_paths.sleef())
ipaths.append(build_paths.openmp())
ipaths.append(build_paths.cc_include())
ipaths.append(build_paths.libgcc())
ipaths.append(build_paths.libgcc_arch())
ipaths.append(build_paths.libgcc_backward())
ipaths.append(build_paths.glibc())
ipaths.append(build_paths.linux_kernel())
ipaths.append(build_paths.cuda())
# We also need to bundle includes with absolute paths into a remote directory
# (later on, we copy the include paths from cpp_extensions into our remote dir)
ipaths.append("include")
static_link_libs = []
if aot_mode and cuda and config.is_fbcode():
# For Meta internal cuda-12, it is recommended to static link cudart
static_link_libs = ["-Wl,-Bstatic", "-lcudart_static", "-Wl,-Bdynamic"]
lpaths_str = " ".join(["-L" + p for p in lpaths])
libs_str = " ".join(static_link_libs + ["-l" + p for p in libs])
return ipaths, lpaths_str, libs_str, macros, build_arch_flags
def cpp_compile_command(
input: Union[str, List[str]],
output: str,
warning_all: bool = True,
shared: bool = True,
include_pytorch: bool = False,
vec_isa: VecISA = invalid_vec_isa,
cuda: bool = False,
aot_mode: bool = False,
compile_only: bool = False,
use_absolute_path: bool = False,
) -> str:
ipaths, lpaths, libs, macros, build_arch_flags = get_include_and_linking_paths(
include_pytorch, vec_isa, cuda, aot_mode
)
if isinstance(input, str):
input = [input]
ipaths_str = " ".join(["-I" + p for p in ipaths])
clang_flags = ""
if config.is_fbcode():
if aot_mode and not use_absolute_path:
inp_name = input
out_name = output
linker_script = _LINKER_SCRIPT
else:
# We need to copy any absolute-path torch includes
inp_name = [os.path.basename(i) for i in input]
out_name = os.path.basename(output)
linker_script = os.path.basename(_LINKER_SCRIPT)
assert is_clang()
# Use clang runtime instead of libgcc
clang_flags += " --rtlib=compiler-rt"
clang_flags += " -fuse-ld=lld"
clang_flags += f" -Wl,--script={linker_script}"
linker_paths = "-B" + build_paths.glibc_lib()
linker_paths += " -L" + build_paths.glibc_lib()
else:
inp_name = input
out_name = output
linker_paths = "" # let the compiler pick
if compile_only:
libs, lpaths = "", ""
inp_name_str = " ".join(inp_name)
return re.sub(
r"[ \n]+",
" ",
f"""
{cpp_compiler()} {inp_name_str} {get_shared(shared, compile_only)}
{get_warning_all_flag(warning_all)} {cpp_flags()}
{get_glibcxx_abi_build_flags()}
{ipaths_str} {lpaths} {libs} {build_arch_flags}
{macros} {linker_paths} {clang_flags}
{optimization_flags()}
{use_custom_generated_macros()}
{use_fb_internal_macros()}
{use_standard_sys_dir_headers()}
{get_compile_only(compile_only)}
-o {out_name}
""",
).strip()
def run_command_and_check(cmd: str):
cmd = shlex.split(cmd)
try:
subprocess.check_call(cmd)
except subprocess.CalledProcessError as e:
raise exc.CppCompileError(cmd, e.output) from e
@functools.lru_cache(None)
def split_aot_inductor_output_path(path: str) -> Tuple[str, str]:
"""Returns the path where the AOT Inductor compiled kernels are stored."""
if path.endswith(".so"):
return os.path.split(path)
else:
return path, ""
class CudaKernelParamCache:
cache: Dict[str, Dict[str, str]] = dict()
clear = staticmethod(cache.clear)
@classmethod
def set(cls, key: str, params: Dict[str, str], cubin: str) -> None:
bin_type = "cubin" if torch.version.hip is None else "hsaco"
_, path = write(
cubin,
bin_type,
hash_type=bin_type,
specified_dir=split_aot_inductor_output_path(
config.aot_inductor.output_path
)[0],
)
params[get_cpp_wrapper_cubin_path_name()] = path
cls.cache[key] = params
@classmethod
def get(cls, key: str) -> Optional[Dict[str, str]]:
return cls.cache.get(key, None)
@classmethod
def get_keys(cls):
return cls.cache.keys()
class AotCodeCompiler:
@classmethod
def compile(
cls,
graph: GraphLowering,
source_code: str,
serialized_extern_kernel_nodes: Optional[str],
cuda: bool,
) -> str:
picked_vec_isa = pick_vec_isa()
cpp_command = repr(
cpp_compile_command(
"i", "o", vec_isa=picked_vec_isa, cuda=cuda, aot_mode=graph.aot_mode
)
)
fbcode_aot_cpu_re = False
use_absolute_path = False
if config.is_fbcode():
ld_command = build_paths.ld()
if not cuda and graph.aot_mode: # Meta internal AOTInductor CPU
objcopy_command = build_paths.objcopy_fallback()
fbcode_aot_cpu_re = True
use_absolute_path = True
else:
objcopy_command = build_paths.objcopy()
else:
ld_command = "ld"
objcopy_command = "objcopy"
(
specified_output_path,
specified_so_name,
) = split_aot_inductor_output_path(config.aot_inductor.output_path)
key, input_path = write(
source_code,
"cpp",
extra=cpp_command,
specified_dir=specified_output_path,
)
def _compile_consts_linux(consts: bytes) -> str:
_, consts_path = write(
consts,
"bin",
specified_dir=specified_output_path,
)
consts_o = os.path.splitext(consts_path)[0] + ".o"
if fbcode_aot_cpu_re:
cmd = f"{ld_command} -r -b binary -o {os.path.basename(consts_o)} {os.path.basename(consts_path)}"
compile_file(consts_path, consts_o, cmd.split())
os.chmod(consts_o, 0o644)
else:
cmd = f"{ld_command} -r -b binary -o {consts_o} {consts_path}"
run_command_and_check(cmd)
log.debug("aot constant binary command: %s", cmd)
cmd = (
f"{objcopy_command} --rename-section"
" .data=.lrodata,alloc,load,readonly,data,contents"
f" {consts_o} {consts_o}"
)
log.debug("aot constant obj command: %s", cmd)
run_command_and_check(cmd)
cmd = f"rm {consts_path}"
log.debug("aot constant bin removal command: %s", cmd)
run_command_and_check(cmd)
if fbcode_aot_cpu_re:
body = re.sub(r"[\W]", "_", os.path.basename(consts_path))
else:
body = re.sub(r"[\W]", "_", consts_path)
symbol_list = []
symbol_list.append(
f"{objcopy_command} --redefine-sym _binary_{body}_start=_binary_constants_bin_start {consts_o}"
)
symbol_list.append(
f"{objcopy_command} --redefine-sym _binary_{body}_size=_binary_constants_bin_size {consts_o}"
)
symbol_list.append(
f"{objcopy_command} --redefine-sym _binary_{body}_end=_binary_constants_bin_end {consts_o}"
)
log.debug("aot constant binary redefine symbol: %s", " ".join(symbol_list))
for cmd in symbol_list:
run_command_and_check(cmd)
return consts_o
def _compile_consts_darwin(consts: bytes) -> str:
is_large_consts = len(consts) > 1024
consts_asm = "\t.section\t__TEXT,__const\n"
consts_asm += "\t.globl\t__binary_constants_bin_start\n"
consts_asm += "__binary_constants_bin_start:\n"
if not is_large_consts:
for c in consts:
consts_asm += f"\t.byte {c}\n"
# Add one element even if constants are empty
# Otherwise assembler will not put them in data section
if not consts:
consts_asm += "\t.space 1\n"
else:
consts_asm += "\t.quad 0x1234567899abcdef\n"
consts_asm += f"\t.space {len(consts) - 8}\n"
consts_asm += ".globl\t__binary_constants_bin_end\n"
consts_asm += "__binary_constants_bin_end:\n"
_, consts_path = write(
consts_asm,
"S",
specified_dir=specified_output_path,
)
consts_o = os.path.splitext(consts_path)[0] + ".o"
cmd = f"{cpp_compiler()} -c -o {consts_o} {consts_path}"
run_command_and_check(cmd)
if is_large_consts:
with open(consts_o, "r+b") as f:
f.seek(0)
hdr = f.read(1024)
# Search for magic number and write the actual data over it
start_idx = hdr.find(b"\xef\xcd\xab\x99\x78\x56\x34\x12")
assert start_idx != -1
f.seek(start_idx)
pos = 0
while pos < len(consts):
rc = f.write(consts[pos:])
pos += rc
return consts_o
from filelock import FileLock
lock_dir = get_lock_dir()
lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT)
with lock:
# Currently, this only support serializing extern nodes in fbcode
# Eventually, we should also have a serializer for OSS.
if config.is_fbcode() and serialized_extern_kernel_nodes:
output_json = os.path.splitext(input_path)[0] + ".json"
with open(output_json, "w") as f:
f.write(serialized_extern_kernel_nodes)
output_so = (
config.aot_inductor.output_path
if specified_so_name
else os.path.splitext(input_path)[0] + ".so"
)
output_o = os.path.splitext(input_path)[0] + ".o"
cmd = cpp_compile_command(
input=input_path,
output=output_o,
vec_isa=picked_vec_isa,
cuda=cuda,
aot_mode=graph.aot_mode,
compile_only=True,
use_absolute_path=use_absolute_path,
)
log.debug("aot compilation command: %s", cmd)
if fbcode_aot_cpu_re:
compile_file(input_path, output_o, cmd.split())
os.chmod(output_o, 0o644)
else:
run_command_and_check(cmd)
def _to_bytes(t: torch.Tensor) -> bytes:
# This serializes the tensor's untyped_storage to bytes by accessing
# the raw data of the underlying structure.
import ctypes
if t.numel() == 0:
return b""
t_cpu = t.untyped_storage().cpu()
raw_array = ctypes.cast(
t_cpu.data_ptr(),
ctypes.POINTER(ctypes.c_ubyte * t_cpu.nbytes()),
)
return bytes(raw_array.contents)
aot_constants = b"".join(
_to_bytes(tensor)
for name, tensor in graph.constants.items()
if name not in graph.folded_constants
)
consts_o = {
"linux": _compile_consts_linux,
"darwin": _compile_consts_darwin,
}[sys.platform](aot_constants)
cmd = cpp_compile_command(
input=[output_o, consts_o],
output=output_so,
vec_isa=picked_vec_isa,
cuda=cuda,
aot_mode=graph.aot_mode,
use_absolute_path=use_absolute_path,
)
log.debug("aot linkage command: %s", cmd)
if fbcode_aot_cpu_re:
compile_file([output_o, consts_o], output_so, cmd.split())
os.chmod(output_so, 0o755)
else:
run_command_and_check(cmd)
return output_so
# Putting this fn in cpp.py (unfortunately) causes a deadlock, which is why it's in codecache.py.
# Why? importing from cpp.py invokes codecache.pick_vec_isa(), which takes out a lock.
# Cycle goes:
# - CppCodeCache.load()
# - pick_vec_isa()
# - valid_vec_isa_list()
# - VecISA.__bool__() <-- takes out a lock
# - compile_file() <-- imports cpp_prefix_path from cpp, which causes us to try to take out the same lock.
@functools.lru_cache
def cpp_prefix_path() -> str:
path = Path(__file__).parent / "codegen/cpp_prefix.h"
with path.open() as f:
content = f.read()
_, filename = write(
content,
"h",
)
return filename
def cpp_prefix() -> str:
filename = cpp_prefix_path()
if config.is_fbcode():
# We need relative paths, since we bundle up
# everything that we compile into a folder for remote compilation.
return f'#include "{os.path.basename(filename)}"'
else:
return f'#include "{filename}"'
# Given a path to an input cpp file and an output path,
# Attempts to compile the file, storing the output in "output_path"
@dynamo_timed
def compile_file(
input_path: Union[str, List[str]], output_path: str, cmd: List[str]
) -> None:
input_paths = [input_path] if isinstance(input_path, str) else input_path
input_files = [
os.path.basename(ip) if config.is_fbcode() else ip for ip in input_paths
]
try:
if config.is_fbcode():
# Need to copy our header into the same folder as the sourcecode.
header_path = cpp_prefix_path()
header_name = os.path.basename(header_path)
output_name = os.path.basename(output_path)
# When we build remotely, we need to make sure to carefully copy any files
# that are required during the compilation process into our build directly.
# This is where all of the ATen/c10/Torch includes come from.
torch_includes_path = os.path.join(_TORCH_PATH, "include")
with tempfile.TemporaryDirectory() as tmp_dir:
# Copy everything to tmp compilation folder
shutil.copy(header_path, os.path.join(tmp_dir, header_name))
shutil.copy(_LINKER_SCRIPT, os.path.join(tmp_dir, "script.ld"))
for p, f in zip(input_paths, input_files):
shutil.copy(p, os.path.join(tmp_dir, f))
dest_include_path = os.path.join(tmp_dir, "include")
shutil.copytree(torch_includes_path, dest_include_path)
# Run the build
output_file_path = _run_build_command(cmd, tmp_dir, output_name)
# Copy output from the build
if os.path.exists(output_path):
os.remove(output_path)
shutil.copy(output_file_path, output_path)
else:
subprocess.check_output(cmd, stderr=subprocess.STDOUT)
except subprocess.CalledProcessError as e:
output = e.output.decode("utf-8")
openmp_problem = "'omp.h' file not found" in output or "libomp" in output
if openmp_problem and sys.platform == "darwin":
instruction = (
"\n\nOpenMP support not found. Please try one of the following solutions:\n"
"(1) Set the `CXX` environment variable to a compiler other than Apple clang++/g++ "
"that has builtin OpenMP support;\n"
"(2) install OpenMP via conda: `conda install llvm-openmp`;\n"
"(3) install libomp via brew: `brew install libomp`;\n"
"(4) manually setup OpenMP and set the `OMP_PREFIX` environment variable to point to a path"
" with `include/omp.h` under it."
)
output += instruction
raise exc.CppCompileError(cmd, output) from e
_libgomp: Optional[CDLL] = None
class CppCodeCache:
cache: Dict[str, Union[CDLL, ModuleType]] = {}
clear = staticmethod(cache.clear)
cpp_compile_command_flags: Dict[str, Any] = {}
@staticmethod
def _load_library_inner(path: str, key: str) -> Union[CDLL, ModuleType]:
return cdll.LoadLibrary(path)
@classmethod
def _load_library(cls, path: str, key: str) -> Union[CDLL, ModuleType]:
try:
return cls._load_library_inner(path, key)
except (ImportError, OSError) as e:
if "gomp" in str(e) and os.path.exists("/usr/lib64/libgomp.so.1"):
# hacky workaround for fbcode/buck
global _libgomp
_libgomp = cdll.LoadLibrary("/usr/lib64/libgomp.so.1")
return cls._load_library_inner(path, key)
if "failed to map segment from shared object" in str(e):
raise OSError(
f"{e}. The most common reason this may occur is if the {tempfile.gettempdir()} folder "
"is mounted with noexec (e.g., by default Docker mounts tmp file systems "
f"as noexec). Please remount {tempfile.gettempdir()} with exec enabled, or set another "
"temporary directory with TORCHINDUCTOR_CACHE_DIR environment variable."
) from e
raise
@classmethod
def load(cls, source_code: str, cuda: bool = False) -> Union[CDLL, ModuleType]:
cls.cpp_compile_command_flags.update({"cuda": cuda})
picked_vec_isa = pick_vec_isa()
cpp_command = repr(
cpp_compile_command(
"i", "o", vec_isa=picked_vec_isa, **cls.cpp_compile_command_flags
)
)
key, input_path = write(source_code, "cpp", extra=cpp_command)
if key not in cls.cache:
from filelock import FileLock
lock_dir = get_lock_dir()
lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT)
with lock:
output_path = input_path[:-3] + "so"
if not os.path.exists(output_path):
cmd = shlex.split(
cpp_compile_command(
input=input_path,
output=output_path,
vec_isa=picked_vec_isa,
**cls.cpp_compile_command_flags,
)
)
compile_file(input_path, output_path, cmd)
cls.cache[key] = cls._load_library(output_path, key)
cls.cache[key].key = key # type: ignore[union-attr]
return cls.cache[key]
# Customized Python binding for cpp kernels
class CppPythonBindingsCodeCache(CppCodeCache):
cache: Dict[str, Union[CDLL, ModuleType]] = {}
clear = staticmethod(cache.clear)
cpp_compile_command_flags = {
# kernels have no dependency on libtorch
"include_pytorch": False,
"shared": True,
}
entry_function = "kernel"
call_entry_function = "kernel(%s);Py_RETURN_NONE;"
extra_parse_arg = ""
suffix_template = textwrap.dedent(
"""
// Python bindings to call %s():
#define PY_SSIZE_T_CLEAN
#include <Python.h>
#include <sstream>
#include <cstdlib>
// This is defined in guards.cpp so we don't need to import PyTorch headers that are slooow.
// We manually link it below to workaround issues with fbcode build.
static void* (*_torchinductor_pyobject_tensor_data_ptr)(PyObject* obj);
template <typename T> static inline T parse_arg(PyObject* args, size_t n) {
static_assert(std::is_pointer<T>::value, "arg type must be pointer or long");
return static_cast<T>(_torchinductor_pyobject_tensor_data_ptr(PyTuple_GET_ITEM(args, n)));
}
template <> inline long parse_arg<long>(PyObject* args, size_t n) {
auto result = PyLong_AsSsize_t(PyTuple_GET_ITEM(args, n));
if(result == -1 && PyErr_Occurred())
[[unlikely]] throw std::runtime_error("expected int arg");
return result;
}
%s
static PyObject* %s_py(PyObject* self, PyObject* args) {
try {
if(!PyTuple_CheckExact(args))
[[unlikely]] throw std::runtime_error("tuple args required");
if(PyTuple_GET_SIZE(args) != %s)
[[unlikely]] throw std::runtime_error("requires %s args");
%s
} catch(std::exception const& e) {
PyErr_SetString(PyExc_RuntimeError, e.what());
return nullptr;
} catch(...) {
PyErr_SetString(PyExc_RuntimeError, "unhandled error");
return nullptr;
}
}
static PyMethodDef py_methods[] = {
{"%s", %s_py, METH_VARARGS, ""},
{NULL, NULL, 0, NULL}};
static struct PyModuleDef py_module =
{PyModuleDef_HEAD_INIT, "%s", NULL, -1, py_methods};
PyMODINIT_FUNC PyInit_%s(void) {
const char* str_addr = std::getenv("_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR");
if(!str_addr) {
PyErr_SetString(PyExc_RuntimeError, "_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR must be set");
return nullptr;
}
std::istringstream iss(str_addr);
uintptr_t addr = 0;
iss >> addr;
_torchinductor_pyobject_tensor_data_ptr =
reinterpret_cast<decltype(_torchinductor_pyobject_tensor_data_ptr)>(addr);
return PyModule_Create(&py_module);
}
"""
)
@classmethod
def _load_library_inner(cls, path: str, key: str) -> ModuleType:
os.environ["_TORCHINDUCTOR_PYOBJECT_TENSOR_DATA_PTR"] = str(
torch._C._dynamo.guards._torchinductor_pyobject_tensor_data_ptr # type: ignore[attr-defined]
)
return importlib.machinery.ExtensionFileLoader(
f"{key}.{cls.entry_function}", path
).load_module() # type: ignore[call-arg]
@classmethod
def load_pybinding(
cls,
argtypes: List[str],
source_code: str,
cuda: bool = False,
num_outputs: int = -1,
) -> Any:
"""
Wrap a C++ function in fast Python bindings.
Args:
argtypes: The types of args to ENTRY_FUNCTION(), e.g. ["float*", "long"]
source_code: C++ source code containing a ENTRY_FUNCTION() function
Returns:
A python version of ENTRY_FUNCTION()
"""
parseargs = ", ".join(
f"parse_arg<{argtype.replace('const ', '')}>(args, {n})"
for n, argtype in enumerate(argtypes)
)
suffix = cls.suffix_template % (
cls.entry_function,
cls.extra_parse_arg % num_outputs if cls.extra_parse_arg else "",
cls.entry_function,
len(argtypes),
len(argtypes),
cls.call_entry_function % parseargs,
cls.entry_function,
cls.entry_function,
cls.entry_function,
cls.entry_function,
)
result = cls.load(source_code + suffix, cuda)
assert isinstance(result, ModuleType)
return getattr(result, cls.entry_function)
class CppWrapperCodeCache(CppPythonBindingsCodeCache):
cache: Dict[str, Union[CDLL, ModuleType]] = {}
clear = staticmethod(cache.clear)
cpp_compile_command_flags = {
"include_pytorch": True,
"shared": True,
}
entry_function = "inductor_entry_cpp"
call_entry_function = "return THPVariable_WrapList(inductor_entry_cpp(%s));"
extra_parse_arg = textwrap.dedent(
"""
#include <torch/csrc/autograd/python_variable.h>
#include <torch/csrc/inductor/aoti_torch/tensor_converter.h>
template <> inline std::vector<at::Tensor> parse_arg<std::vector<at::Tensor>>(PyObject* args, size_t n) {
return THPVariable_UnpackList(PyTuple_GET_ITEM(args, n));
}
std::vector<at::Tensor> inductor_entry_cpp(std::vector<at::Tensor>&& inputs) {
auto input_handles = unsafe_alloc_new_handles_from_tensors(inputs);
// For outputs, we only allocate a vector to hold returned tensor handles,
// not allocating the actual output tensor storage here
std::vector<AtenTensorHandle> output_handles(%s);
try {
inductor_entry_impl(input_handles.data(), output_handles.data());
} catch(std::exception const& e) {
PyErr_SetString(PyExc_RuntimeError, e.what());
return {};
} catch(...) {
PyErr_SetString(PyExc_RuntimeError, "unhandled error");
return {};
}
return alloc_tensors_by_stealing_from_handles(output_handles.data(), output_handles.size());
}
"""
)
class PyCodeCache:
cache: Dict[str, ModuleType] = dict()
linemaps: Dict[str, List[Tuple[Any, ...]]] = dict()
clear = staticmethod(cache.clear)
@classmethod
def write(cls, source_code: str, extra: str = "") -> Tuple[str, str]:
return write(source_code, "py", extra=extra)
@classmethod
def load(
cls,
source_code: str,
extra: str = "",
linemap: Optional[List[Tuple[int, str]]] = None,
attrs: Optional[Dict[str, Any]] = None,
) -> ModuleType:
key, path = write(source_code, "py", extra=extra)
return cls.load_by_key_path(key, path, linemap, attrs)
@classmethod
def load_by_key_path(
cls,
key: str,
path: str,
linemap: Optional[List[Tuple[int, str]]] = None,
attrs: Optional[Dict[str, Any]] = None,
) -> ModuleType:
if linemap is None:
linemap = []
if key not in cls.cache:
with open(path) as f:
try:
code = compile(f.read(), path, "exec")
except Exception as e:
raise RuntimeError(
f"Failed to import {path}\n{type(e).__name__}: {e}"
) from None
mod = ModuleType(f"{__name__}.{key}")
mod.__file__ = path
mod.key = key # type: ignore[attr-defined]
exec(code, mod.__dict__, mod.__dict__)
sys.modules[mod.__name__] = mod
# another thread might set this first
cls.cache.setdefault(key, mod)
# unzip into separate lines/nodes lists
cls.linemaps[path] = list(zip(*linemap))
if attrs is not None:
for k, v in attrs.items():
setattr(mod, k, v)
return cls.cache[key]
@classmethod
@functools.lru_cache(None)
def stack_frames_for_code(
cls, path: str, lineno: int
) -> Optional[List[Dict[str, Any]]]:
if path not in cls.linemaps:
return None
# [(starting_line, <fx node>), ...]
lines, nodes = cls.linemaps[path]
p = bisect_right(lines, lineno)
if p == 0:
return None
entry = nodes[p - 1]
if not entry:
return None
def parse_stack_trace(stack_trace: str) -> List[Dict[str, Any]]:
# ideally fx stores stack traces as data rather than a string
# but this is not along a performance critical path
regex = r'File "(.+)", line (\d+), in (.+)\n'
matches = re.findall(regex, stack_trace)
return [
{"filename": f, "line": int(l), "name": n}
for f, l, n in reversed(matches)
]
return parse_stack_trace(entry)
class TritonCodeCache:
@classmethod
def load(cls, kernel_name: str, source_code: str) -> ModuleType:
mod = PyCodeCache.load(source_code)
return getattr(mod, kernel_name)
def _cuda_compiler() -> Optional[str]:
if cuda_env.nvcc_exist(config.cuda.cuda_cxx):
return config.cuda.cuda_cxx
if cuda_env.nvcc_exist(os.getenv("CUDACXX")):
return os.getenv("CUDACXX", "")
if cuda_env.nvcc_exist(os.getenv("CUDA_HOME")):
return os.path.join(os.getenv("CUDA_HOME", ""), "bin/nvcc")
return "nvcc"
def _cutlass_include_paths() -> List[str]:
cutlass_path = config.cuda.cutlass_dir
return [
os.path.join(cutlass_path, "include"),
os.path.join(cutlass_path, "tools/library/include"),
os.path.join(cutlass_path, "tools/library/src"),
os.path.join(cutlass_path, "tools/util/include"),
]
def _cuda_lib_options() -> List[str]:
from torch.utils import cpp_extension
extra_ldflags: List[str] = []
if is_linux():
extra_lib_dir = "lib64"
if not os.path.exists(
cpp_extension._join_cuda_home(extra_lib_dir)
) and os.path.exists(cpp_extension._join_cuda_home("lib")):
# 64-bit CUDA may be installed in "lib"
# Note that it's also possible both don't exist (see _find_cuda_home) - in that case we stay with "lib64"
extra_lib_dir = "lib"
extra_ldflags.append(f"-L{cpp_extension._join_cuda_home(extra_lib_dir)}")
extra_ldflags.append(
f'-L{cpp_extension._join_cuda_home(extra_lib_dir, "stubs")}'
)
extra_ldflags.append("-lcuda")
extra_ldflags.append("-lcudart")
else:
raise NotImplementedError(
"Unsupported env, failed to find cuda libs! Currently only Linux is supported."
)
return extra_ldflags
def _nvcc_host_compiler_options() -> List[str]:
return [
"-fPIC",
"-fno-strict-aliasing",
"-fvisibility=hidden",
"-Wconversion",
]
def _nvcc_compiler_options() -> List[str]:
arch = cuda_env.get_cuda_arch()
if arch == "90":
# Required by cutlass compilation.
arch = "90a"
code = [f"sm_{arch}", f"compute_{arch}"]
if config.cuda.enable_cuda_lto:
code += [f"lto_{arch}"]
options = [
"-t=0",
"-DCUTLASS_ENABLE_TENSOR_CORE_MMA=1",
"-w",
f"-gencode=arch=compute_{arch},code=[{','.join(code)}]",
config.cuda.compile_opt_level,
"-std=c++17",
"--expt-relaxed-constexpr",
"-DNDEBUG",
]
if config.cuda.enable_debug_info:
options.extend(["-lineinfo", "-g", "-DCUTLASS_DEBUG_TRACE_LEVEL=1"])
if config.cuda.enable_ptxas_info:
options.extend(
[
"--keep", # Keep the intermediate files for debugging (including ptx, sass, cubin etc.)
"--ptxas-options=--warn-on-local-memory-usage", # warn us if local memory is used in CUDA Kernels
"--ptxas-options=--warn-on-spills", # warn us if register spilling happens in CUDA Kernels
"--resource-usage", # Report on CUDA resource usage (shared mem, registers etc.)
"--source-in-ptx",
]
) # Annotate the ptx file with source information
if config.cuda.use_fast_math:
options.extend(
[
"--use_fast_math",
"-DCUTLASS_USE_TANH_FOR_SIGMOID=1",
]
)
return options
def cuda_compile_command(
src_files: List[str],
dst_file: str,
dst_file_ext: str,
) -> str:
include_paths = _cutlass_include_paths()
cuda_lib_options = _cuda_lib_options()
nvcc_host_compiler_options = _nvcc_host_compiler_options()
nvcc_compiler_options = _nvcc_compiler_options()
options = (
nvcc_compiler_options
+ [
f"-Xcompiler {opt}" if "=" in opt else f"-Xcompiler={opt}"
for opt in nvcc_host_compiler_options
]
+ ["-I" + path for path in include_paths]
+ cuda_lib_options
)
src_file = " ".join(src_files)
res = ""
if dst_file_ext == "o":
res = f"{_cuda_compiler()} {' '.join(options)} -c -o {dst_file} {src_file}"
elif dst_file_ext == "so":
options.append("-shared")
res = f"{_cuda_compiler()} {' '.join(options)} -o {dst_file} {src_file}"
else:
raise NotImplementedError(f"Unsupported output file suffix {dst_file_ext}!")
log.debug("CUDA command: %s", res)
return res
class DLLWrapper:
"""A wrapper for a dynamic library."""
def __init__(
self,
lib_path: str,
):
self.lib_path = lib_path
self.DLL = cdll.LoadLibrary(lib_path)
self.is_open = True
def close(self):
if self.is_open:
self._dlclose()
self.is_open = False
def _dlclose(self):
f_dlclose = None
if is_linux():
syms = CDLL(None)
if not hasattr(syms, "dlclose"):
# Apline Linux
syms = CDLL("libc.so")
if hasattr(syms, "dlclose"):
f_dlclose = syms.dlclose
else:
raise NotImplementedError("Unsupported env, failed to do dlclose!")
if f_dlclose is not None:
f_dlclose.argtypes = [c_void_p]
f_dlclose(self.DLL._handle)
else:
log.warning(
"dll unloading function was not found, library may not be unloaded properly!"
)
def __getattr__(self, name):
if not self.is_open:
raise RuntimeError(f"Cannot use closed DLL library: {self.lib_path}")
method = getattr(self.DLL, name)
def _wrapped_func(*args):
err = method(*args)
if err:
raise RuntimeError(f"Error in function: {method.__name__}")
return _wrapped_func
def __enter__(self):
return self
def __exit__(self, *args):
self.close()
def __del__(self):
self.close()
class CUDACodeCache:
@dataclasses.dataclass
class CacheEntry:
input_path: str
output_path: str
cache: Dict[str, CacheEntry] = dict()
clear = staticmethod(cache.clear)
_SOURCE_CODE_SUFFIX = "cu"
@classmethod
def write(cls, source_code, dst_file_ext) -> Tuple[str, str]:
"""
Writes source code into a file with dst_file_ext as the file extension.
Returns the hash key of source code, and the path to the file.
"""
cuda_command = repr(
cuda_compile_command(["dummy_input"], "dummy_output", dst_file_ext)
)
key, input_path = write(
source_code, cls._SOURCE_CODE_SUFFIX, extra=cuda_command
)
return key, input_path
@classmethod
def compile(cls, source_code, dst_file_ext) -> Tuple[str, str, str]:
"""
Compiles CUDA source_code into a file with dst_file_ext extension.
Returns a tuple of dst_file_path, hash_key, source_code_path
"""
key, input_path = cls.write(source_code, dst_file_ext)
if key not in cls.cache:
from filelock import FileLock
lock_dir = get_lock_dir()
lock = FileLock(os.path.join(lock_dir, key + ".lock"), timeout=LOCK_TIMEOUT)
with lock:
output_path = input_path[: -len(cls._SOURCE_CODE_SUFFIX)] + dst_file_ext
if not os.path.exists(output_path):
cmd = cuda_compile_command(
[input_path], output_path, dst_file_ext
).split(" ")
try:
subprocess.check_output(
cmd, stderr=subprocess.STDOUT, env=os.environ
)
except subprocess.CalledProcessError as error:
raise exc.CUDACompileError(cmd, error.output) from error
cls.cache[key] = CUDACodeCache.CacheEntry(input_path, output_path)
return (cls.cache[key].output_path, key, input_path)
@classmethod
def load(cls, source_code, dst_file_ext) -> Tuple[DLLWrapper, str, str]:
"""
Compiles source code and loads the generated .so file.
Returns a tuple of DLLWrapper, hash_key, source_code_path
"""
if dst_file_ext != "so":
raise RuntimeError(
f"Only support loading a .so file for now. "
f"Requested file extension: {dst_file_ext}. Source code: {source_code}"
)
dst_file_path, hash_key, source_code_path = cls.compile(
source_code, dst_file_ext
)
return (DLLWrapper(dst_file_path), hash_key, source_code_path)
def caching_device_properties():
for _, device_interface in get_registered_device_interfaces():
if device_interface.is_available():
device_interface.Worker.get_device_properties()
def _set_triton_ptxas_path() -> None:
if os.environ.get("TRITON_PTXAS_PATH") is not None:
return
ptxas_path = os.path.abspath(
os.path.join(os.path.dirname(__file__), "..", "bin", "ptxas")
)
if not os.path.exists(ptxas_path):
return
if os.path.isfile(ptxas_path) and os.access(ptxas_path, os.X_OK):
os.environ["TRITON_PTXAS_PATH"] = ptxas_path
else:
warnings.warn(f"{ptxas_path} exists but is not an executable")
def _worker_compile(
kernel_name: str, source_code: str, cc: int, device: torch.device
) -> None:
device_interface = get_interface_for_device(device.type)
device_interface.Worker.set_device(device.index)
kernel = TritonCodeCache.load(kernel_name, source_code)
kernel.precompile(warm_cache_only_with_cc=cc)
def _load_kernel(kernel_name: str, source_code: str) -> ModuleType:
_set_triton_ptxas_path()
kernel = TritonCodeCache.load(kernel_name, source_code)
kernel.precompile()
return kernel
class TritonFuture:
kernel: ModuleType
def __init__(
self,
kernel_name: str,
source_code: str,
future: Future[Any],
) -> None:
self.kernel_name = kernel_name
self.source_code = source_code
self.future = future
# @dynamo_utils.dynamo_timed
def result(self) -> ModuleType:
t0 = time()
if hasattr(self, "kernel"):
return self.kernel
# If the worker failed this will throw an exception.
self.future.result()
kernel = self.kernel = _load_kernel(self.kernel_name, self.source_code)
latency = time() - t0
if latency > 50:
developer_warning(
f"Detected long compilation time of {latency} seconds for kernel name {self.kernel_name}"
)
developer_warning(self.source_code)
del self.kernel_name, self.source_code, self.future
return kernel
# If this process dies abnormally (e.g. segfault)
# it will not shut down the workers. Instead
# the workers will have their parent reassigned to the
# init process. This launches a separate thread to
# watch for the worker getting reassigned,
# and cleans it up in this case.
#
# This function cannot be an inner function since otherwise mp_context="spawn" would
# not work for ProcessPoolExecutor since inner functions cannot be pickled.
def _async_compile_initializer(orig_ppid) -> None:
def run() -> None:
while True:
sleep(1)
if orig_ppid != os.getppid():
os.kill(os.getpid(), signal.SIGKILL)
global _watchdog_thread
_watchdog_thread = Thread(target=run, daemon=True)
_watchdog_thread.start()
# Ignore Ctrl-C (i.e. SIGINT) sent to pool workers to avoid meaningless log spam.
signal.signal(signal.SIGINT, signal.SIG_IGN)
_watchdog_thread: Optional[Thread] = None
# Used to keep track of all process pools invoked so far.
_pool_set: Set[ProcessPoolExecutor] = set()
def shutdown_compile_workers() -> None:
"""Shut down all outstanding compile-worker pools."""
global _pool_set
for pool in _pool_set:
pool.shutdown()
_pool_set.clear()
class AsyncCompile:
def __init__(self) -> None:
pass
@staticmethod
@functools.lru_cache(1)
def pool() -> ThreadPoolExecutor:
assert config.compile_threads > 1
return ThreadPoolExecutor(config.compile_threads)
@staticmethod
@functools.lru_cache(1)
def process_pool() -> ProcessPoolExecutor:
# ensure properties have been calculated before processes
# are forked
caching_device_properties()
assert config.compile_threads > 1
orig_ppid = os.getpid()
ctx = multiprocessing.get_context(config.worker_start_method)
pool = ProcessPoolExecutor(
config.compile_threads,
mp_context=ctx,
initializer=partial(_async_compile_initializer, orig_ppid),
)
global _pool_set
_pool_set.add(pool)
# when this pool is created in a subprocess object, the normal exit handler
# doesn't run, and we need to register our own handler.
# exitpriority has to be high, because another one of the finalizers will
# kill the worker thread that sends the shutdown message to the workers...
multiprocessing.util.Finalize(None, pool.shutdown, exitpriority=sys.maxsize)
return pool
@classmethod
def warm_pool(cls) -> None:
if config.compile_threads <= 1:
return
_compile_start()
pool = cls.process_pool()
# We have to fork processes for compiler workers, but the more memory and other resources that are loaded, the
# slower the os.fork time is, quite drastically. It also holds the GIL so we can't put it on another thread.
# Examples:
# A simple x + x + x script: 10ms seconds in the middle of the program, 2ms at startup
# tf_efficientnet_b0 benchmark: 50ms! in the middle of the program , 3ms at startup
# So we want to start the workers early when it is still cheap, and also to allow the workers to get
# ready before we have work for them.
# ProcessPoolExecutor also does not launch the workers until it finds a point when all the workers are idle.
# But if we waited until then fork time will be long and we will be waiting for the processes to initialize.
# We force them to start here with some YOLOing of the internal methods.
if hasattr(pool, "_start_queue_management_thread"):
pool._start_queue_management_thread()
else:
for _ in range(config.compile_threads):
pool._adjust_process_count()
if hasattr(pool, "_start_executor_manager_thread"):
pool._start_executor_manager_thread()
_compile_end()
@classmethod
def submit(cls, task: Callable[..., Any]) -> Any:
if config.compile_threads <= 1:
return task()
return cls.pool().submit(task)
@classmethod
def map(cls, fn: Callable[..., Any], seq: List[Any]) -> List[Any]:
if config.compile_threads <= 1 or len(seq) <= 1:
return list(map(fn, seq))
return [t.result() for t in [cls.pool().submit(fn, x) for x in seq]]
def triton(
self, kernel_name: str, source_code: str, device_str: str = "cuda"
) -> Union[TritonFuture, ModuleType]:
_compile_start()
if config.compile_threads > 1:
device_interface = get_interface_for_device(device_str)
device = torch.device(device_str, device_interface.current_device())
cc = device_interface.get_compute_capability(device)
future = self.process_pool().submit(
_worker_compile, kernel_name, source_code, cc, device
)
return TritonFuture(kernel_name, source_code, future)
else:
return _load_kernel(kernel_name, source_code)
def multi_kernel(self, *args, **kwargs) -> ModuleType:
"""
Async compile the python shim for multi-kernel.
"""
def task():
from torch._inductor.codegen.multi_kernel import MultiKernelCall
return MultiKernelCall(*args, **kwargs)
return self.submit(task)
def cpp(self, source_code: str) -> ModuleType:
def task():
return CppCodeCache.load(source_code).kernel
return self.submit(task)
def cpp_pybinding(self, argtypes: List[str], source_code: str) -> ModuleType:
return self.submit(
functools.partial(
CppPythonBindingsCodeCache.load_pybinding, argtypes, source_code
)
)
def cuda(self, source_code, dst_file_ext):
def task():
return CUDACodeCache.load(source_code, dst_file_ext)[0]
return self.submit(task)
def wait(self, scope: Dict[str, Any]) -> None:
num_kernels = len(
[
value
for key, value in scope.items()
if isinstance(value, (Future, TritonFuture))
]
)
pbar = tqdm(
total=num_kernels,
desc="Inductor Compilation",
disable=config.disable_progress,
delay=0,
)
if config.compile_threads > 1:
for key, result in scope.items():
if config.verbose_progress and not isinstance(pbar, _Faketqdm):
pbar.set_postfix_str(key)
if isinstance(result, (Future, TritonFuture)):
scope[key] = result.result()
pbar.update(1)
_compile_end()
if os.environ.get("TORCH_TNT_IN_USE", "0") == "1":
# When TorchTNT is used, calling warm_pool() here will cause the
# compile workers created not being able to be shut down inside
# shutdown_compile_workers(). This may cause significant QPS drop.
log.info("Do not call AsyncCompile.warm_pool() because TorchTNT is in use.")
else:
AsyncCompile.warm_pool()
|