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
|
/* Plugin for NVPTX execution.
Copyright (C) 2013-2018 Free Software Foundation, Inc.
Contributed by Mentor Embedded.
This file is part of the GNU Offloading and Multi Processing Library
(libgomp).
Libgomp is free software; you can redistribute it and/or modify it
under the terms of the GNU General Public License as published by
the Free Software Foundation; either version 3, or (at your option)
any later version.
Libgomp is distributed in the hope that it will be useful, but WITHOUT ANY
WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU General Public License for
more details.
Under Section 7 of GPL version 3, you are granted additional
permissions described in the GCC Runtime Library Exception, version
3.1, as published by the Free Software Foundation.
You should have received a copy of the GNU General Public License and
a copy of the GCC Runtime Library Exception along with this program;
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
<http://www.gnu.org/licenses/>. */
/* Nvidia PTX-specific parts of OpenACC support. The cuda driver
library appears to hold some implicit state, but the documentation
is not clear as to what that state might be. Or how one might
propagate it from one thread to another. */
#define _GNU_SOURCE
#include "openacc.h"
#include "config.h"
#include "libgomp-plugin.h"
#include "oacc-plugin.h"
#include "gomp-constants.h"
#include <pthread.h>
#include <cuda.h>
#include <stdbool.h>
#include <stdint.h>
#include <limits.h>
#include <string.h>
#include <stdio.h>
#include <unistd.h>
#include <assert.h>
#include <errno.h>
#if CUDA_VERSION < 6000
extern CUresult cuGetErrorString (CUresult, const char **);
#endif
#define DO_PRAGMA(x) _Pragma (#x)
#if PLUGIN_NVPTX_DYNAMIC
# include <dlfcn.h>
struct cuda_lib_s {
# define CUDA_ONE_CALL(call) \
__typeof (call) *call;
# define CUDA_ONE_CALL_MAYBE_NULL(call) \
CUDA_ONE_CALL (call)
#include "cuda-lib.def"
# undef CUDA_ONE_CALL
# undef CUDA_ONE_CALL_MAYBE_NULL
} cuda_lib;
/* -1 if init_cuda_lib has not been called yet, false
if it has been and failed, true if it has been and succeeded. */
static signed char cuda_lib_inited = -1;
/* Dynamically load the CUDA runtime library and initialize function
pointers, return false if unsuccessful, true if successful. */
static bool
init_cuda_lib (void)
{
if (cuda_lib_inited != -1)
return cuda_lib_inited;
const char *cuda_runtime_lib = "libcuda.so.1";
void *h = dlopen (cuda_runtime_lib, RTLD_LAZY);
cuda_lib_inited = false;
if (h == NULL)
return false;
# define CUDA_ONE_CALL(call) CUDA_ONE_CALL_1 (call, false)
# define CUDA_ONE_CALL_MAYBE_NULL(call) CUDA_ONE_CALL_1 (call, true)
# define CUDA_ONE_CALL_1(call, allow_null) \
cuda_lib.call = dlsym (h, #call); \
if (!allow_null && cuda_lib.call == NULL) \
return false;
#include "cuda-lib.def"
# undef CUDA_ONE_CALL
# undef CUDA_ONE_CALL_1
# undef CUDA_ONE_CALL_MAYBE_NULL
cuda_lib_inited = true;
return true;
}
# define CUDA_CALL_PREFIX cuda_lib.
#else
# define CUDA_ONE_CALL(call)
# define CUDA_ONE_CALL_MAYBE_NULL(call) DO_PRAGMA (weak call)
#include "cuda-lib.def"
#undef CUDA_ONE_CALL_MAYBE_NULL
#undef CUDA_ONE_CALL
# define CUDA_CALL_PREFIX
# define init_cuda_lib() true
#endif
#include "secure_getenv.h"
#undef MIN
#undef MAX
#define MIN(X,Y) ((X) < (Y) ? (X) : (Y))
#define MAX(X,Y) ((X) > (Y) ? (X) : (Y))
/* Convenience macros for the frequently used CUDA library call and
error handling sequence as well as CUDA library calls that
do the error checking themselves or don't do it at all. */
#define CUDA_CALL_ERET(ERET, FN, ...) \
do { \
unsigned __r \
= CUDA_CALL_PREFIX FN (__VA_ARGS__); \
if (__r != CUDA_SUCCESS) \
{ \
GOMP_PLUGIN_error (#FN " error: %s", \
cuda_error (__r)); \
return ERET; \
} \
} while (0)
#define CUDA_CALL(FN, ...) \
CUDA_CALL_ERET (false, FN, __VA_ARGS__)
#define CUDA_CALL_ASSERT(FN, ...) \
do { \
unsigned __r \
= CUDA_CALL_PREFIX FN (__VA_ARGS__); \
if (__r != CUDA_SUCCESS) \
{ \
GOMP_PLUGIN_fatal (#FN " error: %s", \
cuda_error (__r)); \
} \
} while (0)
#define CUDA_CALL_NOCHECK(FN, ...) \
CUDA_CALL_PREFIX FN (__VA_ARGS__)
#define CUDA_CALL_EXISTS(FN) \
CUDA_CALL_PREFIX FN
static const char *
cuda_error (CUresult r)
{
const char *desc;
r = CUDA_CALL_NOCHECK (cuGetErrorString, r, &desc);
if (r != CUDA_SUCCESS)
desc = "unknown cuda error";
return desc;
}
static unsigned int instantiated_devices = 0;
static pthread_mutex_t ptx_dev_lock = PTHREAD_MUTEX_INITIALIZER;
struct ptx_stream
{
CUstream stream;
pthread_t host_thread;
bool multithreaded;
CUdeviceptr d;
void *h;
void *h_begin;
void *h_end;
void *h_next;
void *h_prev;
void *h_tail;
struct ptx_stream *next;
};
/* Thread-specific data for PTX. */
struct nvptx_thread
{
struct ptx_stream *current_stream;
struct ptx_device *ptx_dev;
};
static bool
map_init (struct ptx_stream *s)
{
int size = getpagesize ();
assert (s);
assert (!s->d);
assert (!s->h);
CUDA_CALL (cuMemAllocHost, &s->h, size);
CUDA_CALL (cuMemHostGetDevicePointer, &s->d, s->h, 0);
assert (s->h);
s->h_begin = s->h;
s->h_end = s->h_begin + size;
s->h_next = s->h_prev = s->h_tail = s->h_begin;
assert (s->h_next);
assert (s->h_end);
return true;
}
static bool
map_fini (struct ptx_stream *s)
{
CUDA_CALL (cuMemFreeHost, s->h);
return true;
}
static void
map_pop (struct ptx_stream *s)
{
assert (s != NULL);
assert (s->h_next);
assert (s->h_prev);
assert (s->h_tail);
s->h_tail = s->h_next;
if (s->h_tail >= s->h_end)
s->h_tail = s->h_begin + (int) (s->h_tail - s->h_end);
if (s->h_next == s->h_tail)
s->h_prev = s->h_next;
assert (s->h_next >= s->h_begin);
assert (s->h_tail >= s->h_begin);
assert (s->h_prev >= s->h_begin);
assert (s->h_next <= s->h_end);
assert (s->h_tail <= s->h_end);
assert (s->h_prev <= s->h_end);
}
static void
map_push (struct ptx_stream *s, size_t size, void **h, void **d)
{
int left;
int offset;
assert (s != NULL);
left = s->h_end - s->h_next;
assert (s->h_prev);
assert (s->h_next);
if (size >= left)
{
assert (s->h_next == s->h_prev);
s->h_next = s->h_prev = s->h_tail = s->h_begin;
}
assert (s->h_next);
offset = s->h_next - s->h;
*d = (void *)(s->d + offset);
*h = (void *)(s->h + offset);
s->h_prev = s->h_next;
s->h_next += size;
assert (s->h_prev);
assert (s->h_next);
assert (s->h_next >= s->h_begin);
assert (s->h_tail >= s->h_begin);
assert (s->h_prev >= s->h_begin);
assert (s->h_next <= s->h_end);
assert (s->h_tail <= s->h_end);
assert (s->h_prev <= s->h_end);
return;
}
/* Target data function launch information. */
struct targ_fn_launch
{
const char *fn;
unsigned short dim[GOMP_DIM_MAX];
};
/* Target PTX object information. */
struct targ_ptx_obj
{
const char *code;
size_t size;
};
/* Target data image information. */
typedef struct nvptx_tdata
{
const struct targ_ptx_obj *ptx_objs;
unsigned ptx_num;
const char *const *var_names;
unsigned var_num;
const struct targ_fn_launch *fn_descs;
unsigned fn_num;
} nvptx_tdata_t;
/* Descriptor of a loaded function. */
struct targ_fn_descriptor
{
CUfunction fn;
const struct targ_fn_launch *launch;
int regs_per_thread;
int max_threads_per_block;
};
/* A loaded PTX image. */
struct ptx_image_data
{
const void *target_data;
CUmodule module;
struct targ_fn_descriptor *fns; /* Array of functions. */
struct ptx_image_data *next;
};
struct ptx_device
{
CUcontext ctx;
bool ctx_shared;
CUdevice dev;
struct ptx_stream *null_stream;
/* All non-null streams associated with this device (actually context),
either created implicitly or passed in from the user (via
acc_set_cuda_stream). */
struct ptx_stream *active_streams;
struct {
struct ptx_stream **arr;
int size;
} async_streams;
/* A lock for use when manipulating the above stream list and array. */
pthread_mutex_t stream_lock;
int ord;
bool overlap;
bool map;
bool concur;
bool mkern;
int mode;
int clock_khz;
int num_sms;
int regs_per_block;
int regs_per_sm;
int warp_size;
int max_threads_per_block;
int max_threads_per_multiprocessor;
int default_dims[GOMP_DIM_MAX];
struct ptx_image_data *images; /* Images loaded on device. */
pthread_mutex_t image_lock; /* Lock for above list. */
struct ptx_device *next;
};
enum ptx_event_type
{
PTX_EVT_MEM,
PTX_EVT_KNL,
PTX_EVT_SYNC,
PTX_EVT_ASYNC_CLEANUP
};
struct ptx_event
{
CUevent *evt;
int type;
void *addr;
int ord;
int val;
struct ptx_event *next;
};
static pthread_mutex_t ptx_event_lock;
static struct ptx_event *ptx_events;
static struct ptx_device **ptx_devices;
static inline struct nvptx_thread *
nvptx_thread (void)
{
return (struct nvptx_thread *) GOMP_PLUGIN_acc_thread ();
}
static bool
init_streams_for_device (struct ptx_device *ptx_dev, int concurrency)
{
int i;
struct ptx_stream *null_stream
= GOMP_PLUGIN_malloc (sizeof (struct ptx_stream));
null_stream->stream = NULL;
null_stream->host_thread = pthread_self ();
null_stream->multithreaded = true;
null_stream->d = (CUdeviceptr) NULL;
null_stream->h = NULL;
if (!map_init (null_stream))
return false;
ptx_dev->null_stream = null_stream;
ptx_dev->active_streams = NULL;
pthread_mutex_init (&ptx_dev->stream_lock, NULL);
if (concurrency < 1)
concurrency = 1;
/* This is just a guess -- make space for as many async streams as the
current device is capable of concurrently executing. This can grow
later as necessary. No streams are created yet. */
ptx_dev->async_streams.arr
= GOMP_PLUGIN_malloc (concurrency * sizeof (struct ptx_stream *));
ptx_dev->async_streams.size = concurrency;
for (i = 0; i < concurrency; i++)
ptx_dev->async_streams.arr[i] = NULL;
return true;
}
static bool
fini_streams_for_device (struct ptx_device *ptx_dev)
{
free (ptx_dev->async_streams.arr);
bool ret = true;
while (ptx_dev->active_streams != NULL)
{
struct ptx_stream *s = ptx_dev->active_streams;
ptx_dev->active_streams = ptx_dev->active_streams->next;
ret &= map_fini (s);
CUresult r = CUDA_CALL_NOCHECK (cuStreamDestroy, s->stream);
if (r != CUDA_SUCCESS)
{
GOMP_PLUGIN_error ("cuStreamDestroy error: %s", cuda_error (r));
ret = false;
}
free (s);
}
ret &= map_fini (ptx_dev->null_stream);
free (ptx_dev->null_stream);
return ret;
}
/* Select a stream for (OpenACC-semantics) ASYNC argument for the current
thread THREAD (and also current device/context). If CREATE is true, create
the stream if it does not exist (or use EXISTING if it is non-NULL), and
associate the stream with the same thread argument. Returns stream to use
as result. */
static struct ptx_stream *
select_stream_for_async (int async, pthread_t thread, bool create,
CUstream existing)
{
struct nvptx_thread *nvthd = nvptx_thread ();
/* Local copy of TLS variable. */
struct ptx_device *ptx_dev = nvthd->ptx_dev;
struct ptx_stream *stream = NULL;
int orig_async = async;
/* The special value acc_async_noval (-1) maps (for now) to an
implicitly-created stream, which is then handled the same as any other
numbered async stream. Other options are available, e.g. using the null
stream for anonymous async operations, or choosing an idle stream from an
active set. But, stick with this for now. */
if (async > acc_async_sync)
async++;
if (create)
pthread_mutex_lock (&ptx_dev->stream_lock);
/* NOTE: AFAICT there's no particular need for acc_async_sync to map to the
null stream, and in fact better performance may be obtainable if it doesn't
(because the null stream enforces overly-strict synchronisation with
respect to other streams for legacy reasons, and that's probably not
needed with OpenACC). Maybe investigate later. */
if (async == acc_async_sync)
stream = ptx_dev->null_stream;
else if (async >= 0 && async < ptx_dev->async_streams.size
&& ptx_dev->async_streams.arr[async] && !(create && existing))
stream = ptx_dev->async_streams.arr[async];
else if (async >= 0 && create)
{
if (async >= ptx_dev->async_streams.size)
{
int i, newsize = ptx_dev->async_streams.size * 2;
if (async >= newsize)
newsize = async + 1;
ptx_dev->async_streams.arr
= GOMP_PLUGIN_realloc (ptx_dev->async_streams.arr,
newsize * sizeof (struct ptx_stream *));
for (i = ptx_dev->async_streams.size; i < newsize; i++)
ptx_dev->async_streams.arr[i] = NULL;
ptx_dev->async_streams.size = newsize;
}
/* Create a new stream on-demand if there isn't one already, or if we're
setting a particular async value to an existing (externally-provided)
stream. */
if (!ptx_dev->async_streams.arr[async] || existing)
{
CUresult r;
struct ptx_stream *s
= GOMP_PLUGIN_malloc (sizeof (struct ptx_stream));
if (existing)
s->stream = existing;
else
{
r = CUDA_CALL_NOCHECK (cuStreamCreate, &s->stream,
CU_STREAM_DEFAULT);
if (r != CUDA_SUCCESS)
{
pthread_mutex_unlock (&ptx_dev->stream_lock);
GOMP_PLUGIN_fatal ("cuStreamCreate error: %s",
cuda_error (r));
}
}
/* If CREATE is true, we're going to be queueing some work on this
stream. Associate it with the current host thread. */
s->host_thread = thread;
s->multithreaded = false;
s->d = (CUdeviceptr) NULL;
s->h = NULL;
if (!map_init (s))
{
pthread_mutex_unlock (&ptx_dev->stream_lock);
GOMP_PLUGIN_fatal ("map_init fail");
}
s->next = ptx_dev->active_streams;
ptx_dev->active_streams = s;
ptx_dev->async_streams.arr[async] = s;
}
stream = ptx_dev->async_streams.arr[async];
}
else if (async < 0)
{
if (create)
pthread_mutex_unlock (&ptx_dev->stream_lock);
GOMP_PLUGIN_fatal ("bad async %d", async);
}
if (create)
{
assert (stream != NULL);
/* If we're trying to use the same stream from different threads
simultaneously, set stream->multithreaded to true. This affects the
behaviour of acc_async_test_all and acc_wait_all, which are supposed to
only wait for asynchronous launches from the same host thread they are
invoked on. If multiple threads use the same async value, we make note
of that here and fall back to testing/waiting for all threads in those
functions. */
if (thread != stream->host_thread)
stream->multithreaded = true;
pthread_mutex_unlock (&ptx_dev->stream_lock);
}
else if (stream && !stream->multithreaded
&& !pthread_equal (stream->host_thread, thread))
GOMP_PLUGIN_fatal ("async %d used on wrong thread", orig_async);
return stream;
}
/* Initialize the device. Return TRUE on success, else FALSE. PTX_DEV_LOCK
should be locked on entry and remains locked on exit. */
static bool
nvptx_init (void)
{
int ndevs;
if (instantiated_devices != 0)
return true;
ptx_events = NULL;
pthread_mutex_init (&ptx_event_lock, NULL);
if (!init_cuda_lib ())
return false;
CUDA_CALL (cuInit, 0);
CUDA_CALL (cuDeviceGetCount, &ndevs);
ptx_devices = GOMP_PLUGIN_malloc_cleared (sizeof (struct ptx_device *)
* ndevs);
return true;
}
/* Select the N'th PTX device for the current host thread. The device must
have been previously opened before calling this function. */
static bool
nvptx_attach_host_thread_to_device (int n)
{
CUdevice dev;
CUresult r;
struct ptx_device *ptx_dev;
CUcontext thd_ctx;
r = CUDA_CALL_NOCHECK (cuCtxGetDevice, &dev);
if (r != CUDA_SUCCESS && r != CUDA_ERROR_INVALID_CONTEXT)
{
GOMP_PLUGIN_error ("cuCtxGetDevice error: %s", cuda_error (r));
return false;
}
if (r != CUDA_ERROR_INVALID_CONTEXT && dev == n)
return true;
else
{
CUcontext old_ctx;
ptx_dev = ptx_devices[n];
if (!ptx_dev)
{
GOMP_PLUGIN_error ("device %d not found", n);
return false;
}
CUDA_CALL (cuCtxGetCurrent, &thd_ctx);
/* We don't necessarily have a current context (e.g. if it has been
destroyed. Pop it if we do though. */
if (thd_ctx != NULL)
CUDA_CALL (cuCtxPopCurrent, &old_ctx);
CUDA_CALL (cuCtxPushCurrent, ptx_dev->ctx);
}
return true;
}
static struct ptx_device *
nvptx_open_device (int n)
{
struct ptx_device *ptx_dev;
CUdevice dev, ctx_dev;
CUresult r;
int async_engines, pi;
CUDA_CALL_ERET (NULL, cuDeviceGet, &dev, n);
ptx_dev = GOMP_PLUGIN_malloc (sizeof (struct ptx_device));
ptx_dev->ord = n;
ptx_dev->dev = dev;
ptx_dev->ctx_shared = false;
r = CUDA_CALL_NOCHECK (cuCtxGetDevice, &ctx_dev);
if (r != CUDA_SUCCESS && r != CUDA_ERROR_INVALID_CONTEXT)
{
GOMP_PLUGIN_error ("cuCtxGetDevice error: %s", cuda_error (r));
return NULL;
}
if (r != CUDA_ERROR_INVALID_CONTEXT && ctx_dev != dev)
{
/* The current host thread has an active context for a different device.
Detach it. */
CUcontext old_ctx;
CUDA_CALL_ERET (NULL, cuCtxPopCurrent, &old_ctx);
}
CUDA_CALL_ERET (NULL, cuCtxGetCurrent, &ptx_dev->ctx);
if (!ptx_dev->ctx)
CUDA_CALL_ERET (NULL, cuCtxCreate, &ptx_dev->ctx, CU_CTX_SCHED_AUTO, dev);
else
ptx_dev->ctx_shared = true;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_GPU_OVERLAP, dev);
ptx_dev->overlap = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, dev);
ptx_dev->map = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_CONCURRENT_KERNELS, dev);
ptx_dev->concur = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_COMPUTE_MODE, dev);
ptx_dev->mode = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_INTEGRATED, dev);
ptx_dev->mkern = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_CLOCK_RATE, dev);
ptx_dev->clock_khz = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_MULTIPROCESSOR_COUNT, dev);
ptx_dev->num_sms = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_BLOCK, dev);
ptx_dev->regs_per_block = pi;
/* CU_DEVICE_ATTRIBUTE_MAX_REGISTERS_PER_MULTIPROCESSOR = 82 is defined only
in CUDA 6.0 and newer. */
r = CUDA_CALL_NOCHECK (cuDeviceGetAttribute, &pi, 82, dev);
/* Fallback: use limit of registers per block, which is usually equal. */
if (r == CUDA_ERROR_INVALID_VALUE)
pi = ptx_dev->regs_per_block;
else if (r != CUDA_SUCCESS)
{
GOMP_PLUGIN_error ("cuDeviceGetAttribute error: %s", cuda_error (r));
return NULL;
}
ptx_dev->regs_per_sm = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute,
&pi, CU_DEVICE_ATTRIBUTE_WARP_SIZE, dev);
if (pi != 32)
{
GOMP_PLUGIN_error ("Only warp size 32 is supported");
return NULL;
}
ptx_dev->warp_size = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute, &pi,
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_BLOCK, dev);
ptx_dev->max_threads_per_block = pi;
CUDA_CALL_ERET (NULL, cuDeviceGetAttribute, &pi,
CU_DEVICE_ATTRIBUTE_MAX_THREADS_PER_MULTIPROCESSOR, dev);
ptx_dev->max_threads_per_multiprocessor = pi;
r = CUDA_CALL_NOCHECK (cuDeviceGetAttribute, &async_engines,
CU_DEVICE_ATTRIBUTE_ASYNC_ENGINE_COUNT, dev);
if (r != CUDA_SUCCESS)
async_engines = 1;
for (int i = 0; i != GOMP_DIM_MAX; i++)
ptx_dev->default_dims[i] = 0;
ptx_dev->images = NULL;
pthread_mutex_init (&ptx_dev->image_lock, NULL);
if (!init_streams_for_device (ptx_dev, async_engines))
return NULL;
return ptx_dev;
}
static bool
nvptx_close_device (struct ptx_device *ptx_dev)
{
if (!ptx_dev)
return true;
if (!fini_streams_for_device (ptx_dev))
return false;
pthread_mutex_destroy (&ptx_dev->image_lock);
if (!ptx_dev->ctx_shared)
CUDA_CALL (cuCtxDestroy, ptx_dev->ctx);
free (ptx_dev);
return true;
}
static int
nvptx_get_num_devices (void)
{
int n;
/* PR libgomp/65099: Currently, we only support offloading in 64-bit
configurations. */
if (sizeof (void *) != 8)
{
GOMP_PLUGIN_debug (0, "Disabling nvptx offloading;"
" only 64-bit configurations are supported\n");
return 0;
}
/* This function will be called before the plugin has been initialized in
order to enumerate available devices, but CUDA API routines can't be used
until cuInit has been called. Just call it now (but don't yet do any
further initialization). */
if (instantiated_devices == 0)
{
if (!init_cuda_lib ())
return 0;
CUresult r = CUDA_CALL_NOCHECK (cuInit, 0);
/* This is not an error: e.g. we may have CUDA libraries installed but
no devices available. */
if (r != CUDA_SUCCESS)
{
GOMP_PLUGIN_debug (0, "Disabling nvptx offloading; cuInit: %s\n",
cuda_error (r));
return 0;
}
}
CUDA_CALL_ERET (-1, cuDeviceGetCount, &n);
return n;
}
static void
notify_var (const char *var_name, const char *env_var)
{
if (env_var == NULL)
GOMP_PLUGIN_debug (0, "%s: <Not defined>\n", var_name);
else
GOMP_PLUGIN_debug (0, "%s: '%s'\n", var_name, env_var);
}
static void
process_GOMP_NVPTX_JIT (intptr_t *gomp_nvptx_o)
{
const char *var_name = "GOMP_NVPTX_JIT";
const char *env_var = secure_getenv (var_name);
notify_var (var_name, env_var);
if (env_var == NULL)
return;
const char *c = env_var;
while (*c != '\0')
{
while (*c == ' ')
c++;
if (c[0] == '-' && c[1] == 'O'
&& '0' <= c[2] && c[2] <= '4'
&& (c[3] == '\0' || c[3] == ' '))
{
*gomp_nvptx_o = c[2] - '0';
c += 3;
continue;
}
GOMP_PLUGIN_error ("Error parsing %s", var_name);
break;
}
}
static bool
link_ptx (CUmodule *module, const struct targ_ptx_obj *ptx_objs,
unsigned num_objs)
{
CUjit_option opts[7];
void *optvals[7];
float elapsed = 0.0;
char elog[1024];
char ilog[16384];
CUlinkState linkstate;
CUresult r;
void *linkout;
size_t linkoutsize __attribute__ ((unused));
opts[0] = CU_JIT_WALL_TIME;
optvals[0] = &elapsed;
opts[1] = CU_JIT_INFO_LOG_BUFFER;
optvals[1] = &ilog[0];
opts[2] = CU_JIT_INFO_LOG_BUFFER_SIZE_BYTES;
optvals[2] = (void *) sizeof ilog;
opts[3] = CU_JIT_ERROR_LOG_BUFFER;
optvals[3] = &elog[0];
opts[4] = CU_JIT_ERROR_LOG_BUFFER_SIZE_BYTES;
optvals[4] = (void *) sizeof elog;
opts[5] = CU_JIT_LOG_VERBOSE;
optvals[5] = (void *) 1;
static intptr_t gomp_nvptx_o = -1;
static bool init_done = false;
if (!init_done)
{
process_GOMP_NVPTX_JIT (&gomp_nvptx_o);
init_done = true;
}
int nopts = 6;
if (gomp_nvptx_o != -1)
{
opts[nopts] = CU_JIT_OPTIMIZATION_LEVEL;
optvals[nopts] = (void *) gomp_nvptx_o;
nopts++;
}
CUDA_CALL (cuLinkCreate, nopts, opts, optvals, &linkstate);
for (; num_objs--; ptx_objs++)
{
/* cuLinkAddData's 'data' argument erroneously omits the const
qualifier. */
GOMP_PLUGIN_debug (0, "Loading:\n---\n%s\n---\n", ptx_objs->code);
r = CUDA_CALL_NOCHECK (cuLinkAddData, linkstate, CU_JIT_INPUT_PTX,
(char *) ptx_objs->code, ptx_objs->size,
0, 0, 0, 0);
if (r != CUDA_SUCCESS)
{
GOMP_PLUGIN_error ("Link error log %s\n", &elog[0]);
GOMP_PLUGIN_error ("cuLinkAddData (ptx_code) error: %s",
cuda_error (r));
return false;
}
}
GOMP_PLUGIN_debug (0, "Linking\n");
r = CUDA_CALL_NOCHECK (cuLinkComplete, linkstate, &linkout, &linkoutsize);
GOMP_PLUGIN_debug (0, "Link complete: %fms\n", elapsed);
GOMP_PLUGIN_debug (0, "Link log %s\n", &ilog[0]);
if (r != CUDA_SUCCESS)
{
GOMP_PLUGIN_error ("cuLinkComplete error: %s", cuda_error (r));
return false;
}
CUDA_CALL (cuModuleLoadData, module, linkout);
CUDA_CALL (cuLinkDestroy, linkstate);
return true;
}
static void
event_gc (bool memmap_lockable)
{
struct ptx_event *ptx_event = ptx_events;
struct ptx_event *async_cleanups = NULL;
struct nvptx_thread *nvthd = nvptx_thread ();
pthread_mutex_lock (&ptx_event_lock);
while (ptx_event != NULL)
{
CUresult r;
struct ptx_event *e = ptx_event;
ptx_event = ptx_event->next;
if (e->ord != nvthd->ptx_dev->ord)
continue;
r = CUDA_CALL_NOCHECK (cuEventQuery, *e->evt);
if (r == CUDA_SUCCESS)
{
bool append_async = false;
CUevent *te;
te = e->evt;
switch (e->type)
{
case PTX_EVT_MEM:
case PTX_EVT_SYNC:
break;
case PTX_EVT_KNL:
map_pop (e->addr);
break;
case PTX_EVT_ASYNC_CLEANUP:
{
/* The function gomp_plugin_async_unmap_vars needs to claim the
memory-map splay tree lock for the current device, so we
can't call it when one of our callers has already claimed
the lock. In that case, just delay the GC for this event
until later. */
if (!memmap_lockable)
continue;
append_async = true;
}
break;
}
CUDA_CALL_NOCHECK (cuEventDestroy, *te);
free ((void *)te);
/* Unlink 'e' from ptx_events list. */
if (ptx_events == e)
ptx_events = ptx_events->next;
else
{
struct ptx_event *e_ = ptx_events;
while (e_->next != e)
e_ = e_->next;
e_->next = e_->next->next;
}
if (append_async)
{
e->next = async_cleanups;
async_cleanups = e;
}
else
free (e);
}
}
pthread_mutex_unlock (&ptx_event_lock);
/* We have to do these here, after ptx_event_lock is released. */
while (async_cleanups)
{
struct ptx_event *e = async_cleanups;
async_cleanups = async_cleanups->next;
GOMP_PLUGIN_async_unmap_vars (e->addr, e->val);
free (e);
}
}
static void
event_add (enum ptx_event_type type, CUevent *e, void *h, int val)
{
struct ptx_event *ptx_event;
struct nvptx_thread *nvthd = nvptx_thread ();
assert (type == PTX_EVT_MEM || type == PTX_EVT_KNL || type == PTX_EVT_SYNC
|| type == PTX_EVT_ASYNC_CLEANUP);
ptx_event = GOMP_PLUGIN_malloc (sizeof (struct ptx_event));
ptx_event->type = type;
ptx_event->evt = e;
ptx_event->addr = h;
ptx_event->ord = nvthd->ptx_dev->ord;
ptx_event->val = val;
pthread_mutex_lock (&ptx_event_lock);
ptx_event->next = ptx_events;
ptx_events = ptx_event;
pthread_mutex_unlock (&ptx_event_lock);
}
static void
nvptx_exec (void (*fn), size_t mapnum, void **hostaddrs, void **devaddrs,
int async, unsigned *dims, void *targ_mem_desc)
{
struct targ_fn_descriptor *targ_fn = (struct targ_fn_descriptor *) fn;
CUfunction function;
CUresult r;
int i;
struct ptx_stream *dev_str;
void *kargs[1];
void *hp, *dp;
struct nvptx_thread *nvthd = nvptx_thread ();
int warp_size = nvthd->ptx_dev->warp_size;
const char *maybe_abort_msg = "(perhaps abort was called)";
function = targ_fn->fn;
dev_str = select_stream_for_async (async, pthread_self (), false, NULL);
assert (dev_str == nvthd->current_stream);
/* Initialize the launch dimensions. Typically this is constant,
provided by the device compiler, but we must permit runtime
values. */
int seen_zero = 0;
for (i = 0; i != GOMP_DIM_MAX; i++)
{
if (targ_fn->launch->dim[i])
dims[i] = targ_fn->launch->dim[i];
if (!dims[i])
seen_zero = 1;
}
if (seen_zero)
{
pthread_mutex_lock (&ptx_dev_lock);
static int gomp_openacc_dims[GOMP_DIM_MAX];
if (!gomp_openacc_dims[0])
{
/* See if the user provided GOMP_OPENACC_DIM environment
variable to specify runtime defaults. */
for (int i = 0; i < GOMP_DIM_MAX; ++i)
gomp_openacc_dims[i] = GOMP_PLUGIN_acc_default_dim (i);
}
if (!nvthd->ptx_dev->default_dims[0])
{
int default_dims[GOMP_DIM_MAX];
for (int i = 0; i < GOMP_DIM_MAX; ++i)
default_dims[i] = gomp_openacc_dims[i];
int gang, worker, vector;
{
int block_size = nvthd->ptx_dev->max_threads_per_block;
int cpu_size = nvthd->ptx_dev->max_threads_per_multiprocessor;
int dev_size = nvthd->ptx_dev->num_sms;
GOMP_PLUGIN_debug (0, " warp_size=%d, block_size=%d,"
" dev_size=%d, cpu_size=%d\n",
warp_size, block_size, dev_size, cpu_size);
gang = (cpu_size / block_size) * dev_size;
worker = block_size / warp_size;
vector = warp_size;
}
/* There is no upper bound on the gang size. The best size
matches the hardware configuration. Logical gangs are
scheduled onto physical hardware. To maximize usage, we
should guess a large number. */
if (default_dims[GOMP_DIM_GANG] < 1)
default_dims[GOMP_DIM_GANG] = gang ? gang : 1024;
/* The worker size must not exceed the hardware. */
if (default_dims[GOMP_DIM_WORKER] < 1
|| (default_dims[GOMP_DIM_WORKER] > worker && gang))
default_dims[GOMP_DIM_WORKER] = worker;
/* The vector size must exactly match the hardware. */
if (default_dims[GOMP_DIM_VECTOR] < 1
|| (default_dims[GOMP_DIM_VECTOR] != vector && gang))
default_dims[GOMP_DIM_VECTOR] = vector;
GOMP_PLUGIN_debug (0, " default dimensions [%d,%d,%d]\n",
default_dims[GOMP_DIM_GANG],
default_dims[GOMP_DIM_WORKER],
default_dims[GOMP_DIM_VECTOR]);
for (i = 0; i != GOMP_DIM_MAX; i++)
nvthd->ptx_dev->default_dims[i] = default_dims[i];
}
pthread_mutex_unlock (&ptx_dev_lock);
{
bool default_dim_p[GOMP_DIM_MAX];
for (i = 0; i != GOMP_DIM_MAX; i++)
{
default_dim_p[i] = !dims[i];
if (default_dim_p[i])
dims[i] = nvthd->ptx_dev->default_dims[i];
}
if (default_dim_p[GOMP_DIM_VECTOR])
dims[GOMP_DIM_VECTOR]
= MIN (dims[GOMP_DIM_VECTOR],
(targ_fn->max_threads_per_block / warp_size * warp_size));
if (default_dim_p[GOMP_DIM_WORKER])
dims[GOMP_DIM_WORKER]
= MIN (dims[GOMP_DIM_WORKER],
targ_fn->max_threads_per_block / dims[GOMP_DIM_VECTOR]);
}
}
/* Check if the accelerator has sufficient hardware resources to
launch the offloaded kernel. */
if (dims[GOMP_DIM_WORKER] * dims[GOMP_DIM_VECTOR]
> targ_fn->max_threads_per_block)
{
int suggest_workers
= targ_fn->max_threads_per_block / dims[GOMP_DIM_VECTOR];
GOMP_PLUGIN_fatal ("The Nvidia accelerator has insufficient resources to"
" launch '%s' with num_workers = %d; recompile the"
" program with 'num_workers = %d' on that offloaded"
" region or '-fopenacc-dim=:%d'",
targ_fn->launch->fn, dims[GOMP_DIM_WORKER],
suggest_workers, suggest_workers);
}
/* This reserves a chunk of a pre-allocated page of memory mapped on both
the host and the device. HP is a host pointer to the new chunk, and DP is
the corresponding device pointer. */
map_push (dev_str, mapnum * sizeof (void *), &hp, &dp);
GOMP_PLUGIN_debug (0, " %s: prepare mappings\n", __FUNCTION__);
/* Copy the array of arguments to the mapped page. */
for (i = 0; i < mapnum; i++)
((void **) hp)[i] = devaddrs[i];
/* Copy the (device) pointers to arguments to the device (dp and hp might in
fact have the same value on a unified-memory system). */
CUDA_CALL_ASSERT (cuMemcpy, (CUdeviceptr) dp, (CUdeviceptr) hp,
mapnum * sizeof (void *));
GOMP_PLUGIN_debug (0, " %s: kernel %s: launch"
" gangs=%u, workers=%u, vectors=%u\n",
__FUNCTION__, targ_fn->launch->fn, dims[GOMP_DIM_GANG],
dims[GOMP_DIM_WORKER], dims[GOMP_DIM_VECTOR]);
// OpenACC CUDA
//
// num_gangs nctaid.x
// num_workers ntid.y
// vector length ntid.x
kargs[0] = &dp;
CUDA_CALL_ASSERT (cuLaunchKernel, function,
dims[GOMP_DIM_GANG], 1, 1,
dims[GOMP_DIM_VECTOR], dims[GOMP_DIM_WORKER], 1,
0, dev_str->stream, kargs, 0);
#ifndef DISABLE_ASYNC
if (async < acc_async_noval)
{
r = CUDA_CALL_NOCHECK (cuStreamSynchronize, dev_str->stream);
if (r == CUDA_ERROR_LAUNCH_FAILED)
GOMP_PLUGIN_fatal ("cuStreamSynchronize error: %s %s\n", cuda_error (r),
maybe_abort_msg);
else if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuStreamSynchronize error: %s", cuda_error (r));
}
else
{
CUevent *e;
e = (CUevent *)GOMP_PLUGIN_malloc (sizeof (CUevent));
r = CUDA_CALL_NOCHECK (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
if (r == CUDA_ERROR_LAUNCH_FAILED)
GOMP_PLUGIN_fatal ("cuEventCreate error: %s %s\n", cuda_error (r),
maybe_abort_msg);
else if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuEventCreate error: %s", cuda_error (r));
event_gc (true);
CUDA_CALL_ASSERT (cuEventRecord, *e, dev_str->stream);
event_add (PTX_EVT_KNL, e, (void *)dev_str, 0);
}
#else
r = CUDA_CALL_NOCHECK (cuCtxSynchronize, );
if (r == CUDA_ERROR_LAUNCH_FAILED)
GOMP_PLUGIN_fatal ("cuCtxSynchronize error: %s %s\n", cuda_error (r),
maybe_abort_msg);
else if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuCtxSynchronize error: %s", cuda_error (r));
#endif
GOMP_PLUGIN_debug (0, " %s: kernel %s: finished\n", __FUNCTION__,
targ_fn->launch->fn);
#ifndef DISABLE_ASYNC
if (async < acc_async_noval)
#endif
map_pop (dev_str);
}
void * openacc_get_current_cuda_context (void);
static void *
nvptx_alloc (size_t s)
{
CUdeviceptr d;
CUDA_CALL_ERET (NULL, cuMemAlloc, &d, s);
return (void *) d;
}
static bool
nvptx_free (void *p)
{
CUdeviceptr pb;
size_t ps;
CUDA_CALL (cuMemGetAddressRange, &pb, &ps, (CUdeviceptr) p);
if ((CUdeviceptr) p != pb)
{
GOMP_PLUGIN_error ("invalid device address");
return false;
}
CUDA_CALL (cuMemFree, (CUdeviceptr) p);
return true;
}
static bool
nvptx_host2dev (void *d, const void *h, size_t s)
{
CUdeviceptr pb;
size_t ps;
struct nvptx_thread *nvthd = nvptx_thread ();
if (!s)
return true;
if (!d)
{
GOMP_PLUGIN_error ("invalid device address");
return false;
}
CUDA_CALL (cuMemGetAddressRange, &pb, &ps, (CUdeviceptr) d);
if (!pb)
{
GOMP_PLUGIN_error ("invalid device address");
return false;
}
if (!h)
{
GOMP_PLUGIN_error ("invalid host address");
return false;
}
if (d == h)
{
GOMP_PLUGIN_error ("invalid host or device address");
return false;
}
if ((void *)(d + s) > (void *)(pb + ps))
{
GOMP_PLUGIN_error ("invalid size");
return false;
}
#ifndef DISABLE_ASYNC
if (nvthd && nvthd->current_stream != nvthd->ptx_dev->null_stream)
{
CUevent *e = (CUevent *)GOMP_PLUGIN_malloc (sizeof (CUevent));
CUDA_CALL (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
event_gc (false);
CUDA_CALL (cuMemcpyHtoDAsync,
(CUdeviceptr) d, h, s, nvthd->current_stream->stream);
CUDA_CALL (cuEventRecord, *e, nvthd->current_stream->stream);
event_add (PTX_EVT_MEM, e, (void *)h, 0);
}
else
#endif
CUDA_CALL (cuMemcpyHtoD, (CUdeviceptr) d, h, s);
return true;
}
static bool
nvptx_dev2host (void *h, const void *d, size_t s)
{
CUdeviceptr pb;
size_t ps;
struct nvptx_thread *nvthd = nvptx_thread ();
if (!s)
return true;
if (!d)
{
GOMP_PLUGIN_error ("invalid device address");
return false;
}
CUDA_CALL (cuMemGetAddressRange, &pb, &ps, (CUdeviceptr) d);
if (!pb)
{
GOMP_PLUGIN_error ("invalid device address");
return false;
}
if (!h)
{
GOMP_PLUGIN_error ("invalid host address");
return false;
}
if (d == h)
{
GOMP_PLUGIN_error ("invalid host or device address");
return false;
}
if ((void *)(d + s) > (void *)(pb + ps))
{
GOMP_PLUGIN_error ("invalid size");
return false;
}
#ifndef DISABLE_ASYNC
if (nvthd && nvthd->current_stream != nvthd->ptx_dev->null_stream)
{
CUevent *e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
CUDA_CALL (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
event_gc (false);
CUDA_CALL (cuMemcpyDtoHAsync,
h, (CUdeviceptr) d, s, nvthd->current_stream->stream);
CUDA_CALL (cuEventRecord, *e, nvthd->current_stream->stream);
event_add (PTX_EVT_MEM, e, (void *)h, 0);
}
else
#endif
CUDA_CALL (cuMemcpyDtoH, h, (CUdeviceptr) d, s);
return true;
}
static void
nvptx_set_async (int async)
{
struct nvptx_thread *nvthd = nvptx_thread ();
nvthd->current_stream
= select_stream_for_async (async, pthread_self (), true, NULL);
}
static int
nvptx_async_test (int async)
{
CUresult r;
struct ptx_stream *s;
s = select_stream_for_async (async, pthread_self (), false, NULL);
if (!s)
GOMP_PLUGIN_fatal ("unknown async %d", async);
r = CUDA_CALL_NOCHECK (cuStreamQuery, s->stream);
if (r == CUDA_SUCCESS)
{
/* The oacc-parallel.c:goacc_wait function calls this hook to determine
whether all work has completed on this stream, and if so omits the call
to the wait hook. If that happens, event_gc might not get called
(which prevents variables from getting unmapped and their associated
device storage freed), so call it here. */
event_gc (true);
return 1;
}
else if (r == CUDA_ERROR_NOT_READY)
return 0;
GOMP_PLUGIN_fatal ("cuStreamQuery error: %s", cuda_error (r));
return 0;
}
static int
nvptx_async_test_all (void)
{
struct ptx_stream *s;
pthread_t self = pthread_self ();
struct nvptx_thread *nvthd = nvptx_thread ();
pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
for (s = nvthd->ptx_dev->active_streams; s != NULL; s = s->next)
{
if ((s->multithreaded || pthread_equal (s->host_thread, self))
&& CUDA_CALL_NOCHECK (cuStreamQuery,
s->stream) == CUDA_ERROR_NOT_READY)
{
pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
return 0;
}
}
pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
event_gc (true);
return 1;
}
static void
nvptx_wait (int async)
{
struct ptx_stream *s;
s = select_stream_for_async (async, pthread_self (), false, NULL);
if (!s)
GOMP_PLUGIN_fatal ("unknown async %d", async);
CUDA_CALL_ASSERT (cuStreamSynchronize, s->stream);
event_gc (true);
}
static void
nvptx_wait_async (int async1, int async2)
{
CUevent *e;
struct ptx_stream *s1, *s2;
pthread_t self = pthread_self ();
/* The stream that is waiting (rather than being waited for) doesn't
necessarily have to exist already. */
s2 = select_stream_for_async (async2, self, true, NULL);
s1 = select_stream_for_async (async1, self, false, NULL);
if (!s1)
GOMP_PLUGIN_fatal ("invalid async 1\n");
if (s1 == s2)
GOMP_PLUGIN_fatal ("identical parameters");
e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
CUDA_CALL_ASSERT (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
event_gc (true);
CUDA_CALL_ASSERT (cuEventRecord, *e, s1->stream);
event_add (PTX_EVT_SYNC, e, NULL, 0);
CUDA_CALL_ASSERT (cuStreamWaitEvent, s2->stream, *e, 0);
}
static void
nvptx_wait_all (void)
{
CUresult r;
struct ptx_stream *s;
pthread_t self = pthread_self ();
struct nvptx_thread *nvthd = nvptx_thread ();
pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
/* Wait for active streams initiated by this thread (or by multiple threads)
to complete. */
for (s = nvthd->ptx_dev->active_streams; s != NULL; s = s->next)
{
if (s->multithreaded || pthread_equal (s->host_thread, self))
{
r = CUDA_CALL_NOCHECK (cuStreamQuery, s->stream);
if (r == CUDA_SUCCESS)
continue;
else if (r != CUDA_ERROR_NOT_READY)
GOMP_PLUGIN_fatal ("cuStreamQuery error: %s", cuda_error (r));
CUDA_CALL_ASSERT (cuStreamSynchronize, s->stream);
}
}
pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
event_gc (true);
}
static void
nvptx_wait_all_async (int async)
{
struct ptx_stream *waiting_stream, *other_stream;
CUevent *e;
struct nvptx_thread *nvthd = nvptx_thread ();
pthread_t self = pthread_self ();
/* The stream doing the waiting. This could be the first mention of the
stream, so create it if necessary. */
waiting_stream
= select_stream_for_async (async, pthread_self (), true, NULL);
/* Launches on the null stream already block on other streams in the
context. */
if (!waiting_stream || waiting_stream == nvthd->ptx_dev->null_stream)
return;
event_gc (true);
pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
for (other_stream = nvthd->ptx_dev->active_streams;
other_stream != NULL;
other_stream = other_stream->next)
{
if (!other_stream->multithreaded
&& !pthread_equal (other_stream->host_thread, self))
continue;
e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
CUDA_CALL_ASSERT (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
/* Record an event on the waited-for stream. */
CUDA_CALL_ASSERT (cuEventRecord, *e, other_stream->stream);
event_add (PTX_EVT_SYNC, e, NULL, 0);
CUDA_CALL_ASSERT (cuStreamWaitEvent, waiting_stream->stream, *e, 0);
}
pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
}
static void *
nvptx_get_current_cuda_device (void)
{
struct nvptx_thread *nvthd = nvptx_thread ();
if (!nvthd || !nvthd->ptx_dev)
return NULL;
return &nvthd->ptx_dev->dev;
}
static void *
nvptx_get_current_cuda_context (void)
{
struct nvptx_thread *nvthd = nvptx_thread ();
if (!nvthd || !nvthd->ptx_dev)
return NULL;
return nvthd->ptx_dev->ctx;
}
static void *
nvptx_get_cuda_stream (int async)
{
struct ptx_stream *s;
struct nvptx_thread *nvthd = nvptx_thread ();
if (!nvthd || !nvthd->ptx_dev)
return NULL;
s = select_stream_for_async (async, pthread_self (), false, NULL);
return s ? s->stream : NULL;
}
static int
nvptx_set_cuda_stream (int async, void *stream)
{
struct ptx_stream *oldstream;
pthread_t self = pthread_self ();
struct nvptx_thread *nvthd = nvptx_thread ();
if (async < 0)
GOMP_PLUGIN_fatal ("bad async %d", async);
pthread_mutex_lock (&nvthd->ptx_dev->stream_lock);
/* We have a list of active streams and an array mapping async values to
entries of that list. We need to take "ownership" of the passed-in stream,
and add it to our list, removing the previous entry also (if there was one)
in order to prevent resource leaks. Note the potential for surprise
here: maybe we should keep track of passed-in streams and leave it up to
the user to tidy those up, but that doesn't work for stream handles
returned from acc_get_cuda_stream above... */
oldstream = select_stream_for_async (async, self, false, NULL);
if (oldstream)
{
if (nvthd->ptx_dev->active_streams == oldstream)
nvthd->ptx_dev->active_streams = nvthd->ptx_dev->active_streams->next;
else
{
struct ptx_stream *s = nvthd->ptx_dev->active_streams;
while (s->next != oldstream)
s = s->next;
s->next = s->next->next;
}
CUDA_CALL_ASSERT (cuStreamDestroy, oldstream->stream);
if (!map_fini (oldstream))
GOMP_PLUGIN_fatal ("error when freeing host memory");
free (oldstream);
}
pthread_mutex_unlock (&nvthd->ptx_dev->stream_lock);
(void) select_stream_for_async (async, self, true, (CUstream) stream);
return 1;
}
/* Plugin entry points. */
const char *
GOMP_OFFLOAD_get_name (void)
{
return "nvptx";
}
unsigned int
GOMP_OFFLOAD_get_caps (void)
{
return GOMP_OFFLOAD_CAP_OPENACC_200 | GOMP_OFFLOAD_CAP_OPENMP_400;
}
int
GOMP_OFFLOAD_get_type (void)
{
return OFFLOAD_TARGET_TYPE_NVIDIA_PTX;
}
int
GOMP_OFFLOAD_get_num_devices (void)
{
return nvptx_get_num_devices ();
}
bool
GOMP_OFFLOAD_init_device (int n)
{
struct ptx_device *dev;
pthread_mutex_lock (&ptx_dev_lock);
if (!nvptx_init () || ptx_devices[n] != NULL)
{
pthread_mutex_unlock (&ptx_dev_lock);
return false;
}
dev = nvptx_open_device (n);
if (dev)
{
ptx_devices[n] = dev;
instantiated_devices++;
}
pthread_mutex_unlock (&ptx_dev_lock);
return dev != NULL;
}
bool
GOMP_OFFLOAD_fini_device (int n)
{
pthread_mutex_lock (&ptx_dev_lock);
if (ptx_devices[n] != NULL)
{
if (!nvptx_attach_host_thread_to_device (n)
|| !nvptx_close_device (ptx_devices[n]))
{
pthread_mutex_unlock (&ptx_dev_lock);
return false;
}
ptx_devices[n] = NULL;
instantiated_devices--;
}
pthread_mutex_unlock (&ptx_dev_lock);
return true;
}
/* Return the libgomp version number we're compatible with. There is
no requirement for cross-version compatibility. */
unsigned
GOMP_OFFLOAD_version (void)
{
return GOMP_VERSION;
}
/* Initialize __nvptx_clocktick, if present in MODULE. */
static void
nvptx_set_clocktick (CUmodule module, struct ptx_device *dev)
{
CUdeviceptr dptr;
CUresult r = CUDA_CALL_NOCHECK (cuModuleGetGlobal, &dptr, NULL,
module, "__nvptx_clocktick");
if (r == CUDA_ERROR_NOT_FOUND)
return;
if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuModuleGetGlobal error: %s", cuda_error (r));
double __nvptx_clocktick = 1e-3 / dev->clock_khz;
r = CUDA_CALL_NOCHECK (cuMemcpyHtoD, dptr, &__nvptx_clocktick,
sizeof (__nvptx_clocktick));
if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuMemcpyHtoD error: %s", cuda_error (r));
}
/* Load the (partial) program described by TARGET_DATA to device
number ORD. Allocate and return TARGET_TABLE. */
int
GOMP_OFFLOAD_load_image (int ord, unsigned version, const void *target_data,
struct addr_pair **target_table)
{
CUmodule module;
const char *const *var_names;
const struct targ_fn_launch *fn_descs;
unsigned int fn_entries, var_entries, i, j;
struct targ_fn_descriptor *targ_fns;
struct addr_pair *targ_tbl;
const nvptx_tdata_t *img_header = (const nvptx_tdata_t *) target_data;
struct ptx_image_data *new_image;
struct ptx_device *dev;
if (GOMP_VERSION_DEV (version) > GOMP_VERSION_NVIDIA_PTX)
{
GOMP_PLUGIN_error ("Offload data incompatible with PTX plugin"
" (expected %u, received %u)",
GOMP_VERSION_NVIDIA_PTX, GOMP_VERSION_DEV (version));
return -1;
}
if (!nvptx_attach_host_thread_to_device (ord)
|| !link_ptx (&module, img_header->ptx_objs, img_header->ptx_num))
return -1;
dev = ptx_devices[ord];
/* The mkoffload utility emits a struct of pointers/integers at the
start of each offload image. The array of kernel names and the
functions addresses form a one-to-one correspondence. */
var_entries = img_header->var_num;
var_names = img_header->var_names;
fn_entries = img_header->fn_num;
fn_descs = img_header->fn_descs;
targ_tbl = GOMP_PLUGIN_malloc (sizeof (struct addr_pair)
* (fn_entries + var_entries));
targ_fns = GOMP_PLUGIN_malloc (sizeof (struct targ_fn_descriptor)
* fn_entries);
*target_table = targ_tbl;
new_image = GOMP_PLUGIN_malloc (sizeof (struct ptx_image_data));
new_image->target_data = target_data;
new_image->module = module;
new_image->fns = targ_fns;
pthread_mutex_lock (&dev->image_lock);
new_image->next = dev->images;
dev->images = new_image;
pthread_mutex_unlock (&dev->image_lock);
for (i = 0; i < fn_entries; i++, targ_fns++, targ_tbl++)
{
CUfunction function;
int nregs, mthrs;
CUDA_CALL_ERET (-1, cuModuleGetFunction, &function, module,
fn_descs[i].fn);
CUDA_CALL_ERET (-1, cuFuncGetAttribute, &nregs,
CU_FUNC_ATTRIBUTE_NUM_REGS, function);
CUDA_CALL_ERET (-1, cuFuncGetAttribute, &mthrs,
CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, function);
targ_fns->fn = function;
targ_fns->launch = &fn_descs[i];
targ_fns->regs_per_thread = nregs;
targ_fns->max_threads_per_block = mthrs;
targ_tbl->start = (uintptr_t) targ_fns;
targ_tbl->end = targ_tbl->start + 1;
}
for (j = 0; j < var_entries; j++, targ_tbl++)
{
CUdeviceptr var;
size_t bytes;
CUDA_CALL_ERET (-1, cuModuleGetGlobal,
&var, &bytes, module, var_names[j]);
targ_tbl->start = (uintptr_t) var;
targ_tbl->end = targ_tbl->start + bytes;
}
nvptx_set_clocktick (module, dev);
return fn_entries + var_entries;
}
/* Unload the program described by TARGET_DATA. DEV_DATA is the
function descriptors allocated by G_O_load_image. */
bool
GOMP_OFFLOAD_unload_image (int ord, unsigned version, const void *target_data)
{
struct ptx_image_data *image, **prev_p;
struct ptx_device *dev = ptx_devices[ord];
if (GOMP_VERSION_DEV (version) > GOMP_VERSION_NVIDIA_PTX)
{
GOMP_PLUGIN_error ("Offload data incompatible with PTX plugin"
" (expected %u, received %u)",
GOMP_VERSION_NVIDIA_PTX, GOMP_VERSION_DEV (version));
return false;
}
bool ret = true;
pthread_mutex_lock (&dev->image_lock);
for (prev_p = &dev->images; (image = *prev_p) != 0; prev_p = &image->next)
if (image->target_data == target_data)
{
*prev_p = image->next;
if (CUDA_CALL_NOCHECK (cuModuleUnload, image->module) != CUDA_SUCCESS)
ret = false;
free (image->fns);
free (image);
break;
}
pthread_mutex_unlock (&dev->image_lock);
return ret;
}
void *
GOMP_OFFLOAD_alloc (int ord, size_t size)
{
if (!nvptx_attach_host_thread_to_device (ord))
return NULL;
return nvptx_alloc (size);
}
bool
GOMP_OFFLOAD_free (int ord, void *ptr)
{
return (nvptx_attach_host_thread_to_device (ord)
&& nvptx_free (ptr));
}
bool
GOMP_OFFLOAD_dev2host (int ord, void *dst, const void *src, size_t n)
{
return (nvptx_attach_host_thread_to_device (ord)
&& nvptx_dev2host (dst, src, n));
}
bool
GOMP_OFFLOAD_host2dev (int ord, void *dst, const void *src, size_t n)
{
return (nvptx_attach_host_thread_to_device (ord)
&& nvptx_host2dev (dst, src, n));
}
bool
GOMP_OFFLOAD_dev2dev (int ord, void *dst, const void *src, size_t n)
{
struct ptx_device *ptx_dev = ptx_devices[ord];
CUDA_CALL (cuMemcpyDtoDAsync, (CUdeviceptr) dst, (CUdeviceptr) src, n,
ptx_dev->null_stream->stream);
return true;
}
void (*device_run) (int n, void *fn_ptr, void *vars) = NULL;
void
GOMP_OFFLOAD_openacc_exec (void (*fn) (void *), size_t mapnum,
void **hostaddrs, void **devaddrs,
int async, unsigned *dims, void *targ_mem_desc)
{
nvptx_exec (fn, mapnum, hostaddrs, devaddrs, async, dims, targ_mem_desc);
}
void
GOMP_OFFLOAD_openacc_register_async_cleanup (void *targ_mem_desc, int async)
{
struct nvptx_thread *nvthd = nvptx_thread ();
CUevent *e = (CUevent *) GOMP_PLUGIN_malloc (sizeof (CUevent));
CUDA_CALL_ASSERT (cuEventCreate, e, CU_EVENT_DISABLE_TIMING);
CUDA_CALL_ASSERT (cuEventRecord, *e, nvthd->current_stream->stream);
event_add (PTX_EVT_ASYNC_CLEANUP, e, targ_mem_desc, async);
}
int
GOMP_OFFLOAD_openacc_async_test (int async)
{
return nvptx_async_test (async);
}
int
GOMP_OFFLOAD_openacc_async_test_all (void)
{
return nvptx_async_test_all ();
}
void
GOMP_OFFLOAD_openacc_async_wait (int async)
{
nvptx_wait (async);
}
void
GOMP_OFFLOAD_openacc_async_wait_async (int async1, int async2)
{
nvptx_wait_async (async1, async2);
}
void
GOMP_OFFLOAD_openacc_async_wait_all (void)
{
nvptx_wait_all ();
}
void
GOMP_OFFLOAD_openacc_async_wait_all_async (int async)
{
nvptx_wait_all_async (async);
}
void
GOMP_OFFLOAD_openacc_async_set_async (int async)
{
nvptx_set_async (async);
}
void *
GOMP_OFFLOAD_openacc_create_thread_data (int ord)
{
struct ptx_device *ptx_dev;
struct nvptx_thread *nvthd
= GOMP_PLUGIN_malloc (sizeof (struct nvptx_thread));
CUcontext thd_ctx;
ptx_dev = ptx_devices[ord];
assert (ptx_dev);
CUDA_CALL_ASSERT (cuCtxGetCurrent, &thd_ctx);
assert (ptx_dev->ctx);
if (!thd_ctx)
CUDA_CALL_ASSERT (cuCtxPushCurrent, ptx_dev->ctx);
nvthd->current_stream = ptx_dev->null_stream;
nvthd->ptx_dev = ptx_dev;
return (void *) nvthd;
}
void
GOMP_OFFLOAD_openacc_destroy_thread_data (void *data)
{
free (data);
}
void *
GOMP_OFFLOAD_openacc_cuda_get_current_device (void)
{
return nvptx_get_current_cuda_device ();
}
void *
GOMP_OFFLOAD_openacc_cuda_get_current_context (void)
{
return nvptx_get_current_cuda_context ();
}
/* NOTE: This returns a CUstream, not a ptx_stream pointer. */
void *
GOMP_OFFLOAD_openacc_cuda_get_stream (int async)
{
return nvptx_get_cuda_stream (async);
}
/* NOTE: This takes a CUstream, not a ptx_stream pointer. */
int
GOMP_OFFLOAD_openacc_cuda_set_stream (int async, void *stream)
{
return nvptx_set_cuda_stream (async, stream);
}
/* Adjust launch dimensions: pick good values for number of blocks and warps
and ensure that number of warps does not exceed CUDA limits as well as GCC's
own limits. */
static void
nvptx_adjust_launch_bounds (struct targ_fn_descriptor *fn,
struct ptx_device *ptx_dev,
int *teams_p, int *threads_p)
{
int max_warps_block = fn->max_threads_per_block / 32;
/* Maximum 32 warps per block is an implementation limit in NVPTX backend
and libgcc, which matches documented limit of all GPUs as of 2015. */
if (max_warps_block > 32)
max_warps_block = 32;
if (*threads_p <= 0)
*threads_p = 8;
if (*threads_p > max_warps_block)
*threads_p = max_warps_block;
int regs_per_block = fn->regs_per_thread * 32 * *threads_p;
/* This is an estimate of how many blocks the device can host simultaneously.
Actual limit, which may be lower, can be queried with "occupancy control"
driver interface (since CUDA 6.0). */
int max_blocks = ptx_dev->regs_per_sm / regs_per_block * ptx_dev->num_sms;
if (*teams_p <= 0 || *teams_p > max_blocks)
*teams_p = max_blocks;
}
/* Return the size of per-warp stacks (see gcc -msoft-stack) to use for OpenMP
target regions. */
static size_t
nvptx_stacks_size ()
{
return 128 * 1024;
}
/* Return contiguous storage for NUM stacks, each SIZE bytes. */
static void *
nvptx_stacks_alloc (size_t size, int num)
{
CUdeviceptr stacks;
CUresult r = CUDA_CALL_NOCHECK (cuMemAlloc, &stacks, size * num);
if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuMemAlloc error: %s", cuda_error (r));
return (void *) stacks;
}
/* Release storage previously allocated by nvptx_stacks_alloc. */
static void
nvptx_stacks_free (void *p, int num)
{
CUresult r = CUDA_CALL_NOCHECK (cuMemFree, (CUdeviceptr) p);
if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuMemFree error: %s", cuda_error (r));
}
void
GOMP_OFFLOAD_run (int ord, void *tgt_fn, void *tgt_vars, void **args)
{
CUfunction function = ((struct targ_fn_descriptor *) tgt_fn)->fn;
CUresult r;
struct ptx_device *ptx_dev = ptx_devices[ord];
const char *maybe_abort_msg = "(perhaps abort was called)";
int teams = 0, threads = 0;
if (!args)
GOMP_PLUGIN_fatal ("No target arguments provided");
while (*args)
{
intptr_t id = (intptr_t) *args++, val;
if (id & GOMP_TARGET_ARG_SUBSEQUENT_PARAM)
val = (intptr_t) *args++;
else
val = id >> GOMP_TARGET_ARG_VALUE_SHIFT;
if ((id & GOMP_TARGET_ARG_DEVICE_MASK) != GOMP_TARGET_ARG_DEVICE_ALL)
continue;
val = val > INT_MAX ? INT_MAX : val;
id &= GOMP_TARGET_ARG_ID_MASK;
if (id == GOMP_TARGET_ARG_NUM_TEAMS)
teams = val;
else if (id == GOMP_TARGET_ARG_THREAD_LIMIT)
threads = val;
}
nvptx_adjust_launch_bounds (tgt_fn, ptx_dev, &teams, &threads);
size_t stack_size = nvptx_stacks_size ();
void *stacks = nvptx_stacks_alloc (stack_size, teams * threads);
void *fn_args[] = {tgt_vars, stacks, (void *) stack_size};
size_t fn_args_size = sizeof fn_args;
void *config[] = {
CU_LAUNCH_PARAM_BUFFER_POINTER, fn_args,
CU_LAUNCH_PARAM_BUFFER_SIZE, &fn_args_size,
CU_LAUNCH_PARAM_END
};
r = CUDA_CALL_NOCHECK (cuLaunchKernel, function, teams, 1, 1,
32, threads, 1, 0, ptx_dev->null_stream->stream,
NULL, config);
if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuLaunchKernel error: %s", cuda_error (r));
r = CUDA_CALL_NOCHECK (cuCtxSynchronize, );
if (r == CUDA_ERROR_LAUNCH_FAILED)
GOMP_PLUGIN_fatal ("cuCtxSynchronize error: %s %s\n", cuda_error (r),
maybe_abort_msg);
else if (r != CUDA_SUCCESS)
GOMP_PLUGIN_fatal ("cuCtxSynchronize error: %s", cuda_error (r));
nvptx_stacks_free (stacks, teams * threads);
}
void
GOMP_OFFLOAD_async_run (int ord, void *tgt_fn, void *tgt_vars, void **args,
void *async_data)
{
GOMP_PLUGIN_fatal ("GOMP_OFFLOAD_async_run unimplemented");
}
|