111
|
1 /* { dg-do run } */
|
|
2
|
|
3 #include <stdlib.h>
|
|
4
|
|
5 int
|
|
6 main(int argc, char **argv)
|
|
7 {
|
|
8 int iexp, igot, imax, imin;
|
|
9 long long lexp, lgot;
|
|
10 int N = 32;
|
|
11 int i;
|
|
12 int idata[N];
|
|
13 long long ldata[N];
|
|
14 float fexp, fgot;
|
|
15 float fdata[N];
|
|
16
|
|
17 igot = 1234;
|
|
18 iexp = 31;
|
|
19
|
|
20 for (i = 0; i < N; i++)
|
|
21 idata[i] = i;
|
|
22
|
|
23 #pragma acc data copy (igot, idata[0:N])
|
|
24 {
|
|
25 #pragma acc parallel loop
|
|
26 for (i = 0; i < N; i++)
|
|
27 #pragma acc atomic capture
|
|
28 { idata[i] = igot; igot = i; }
|
|
29 }
|
|
30
|
|
31 imax = 0;
|
|
32 imin = N;
|
|
33
|
|
34 for (i = 0; i < N; i++)
|
|
35 {
|
|
36 imax = idata[i] > imax ? idata[i] : imax;
|
|
37 imin = idata[i] < imin ? idata[i] : imin;
|
|
38 }
|
|
39
|
|
40 if (imax != 1234 || imin != 0)
|
|
41 abort ();
|
|
42
|
|
43 return 0;
|
|
44
|
|
45 igot = 0;
|
|
46 iexp = 32;
|
|
47
|
|
48 #pragma acc data copy (igot, idata[0:N])
|
|
49 {
|
|
50 #pragma acc parallel loop
|
|
51 for (i = 0; i < N; i++)
|
|
52 #pragma acc atomic capture
|
|
53 { idata[i] = igot; igot++; }
|
|
54 }
|
|
55
|
|
56 if (iexp != igot)
|
|
57 abort ();
|
|
58
|
|
59 igot = 0;
|
|
60 iexp = 32;
|
|
61
|
|
62 #pragma acc data copy (igot, idata[0:N])
|
|
63 {
|
|
64 #pragma acc parallel loop
|
|
65 for (i = 0; i < N; i++)
|
|
66 #pragma acc atomic capture
|
|
67 { idata[i] = igot; ++igot; }
|
|
68 }
|
|
69
|
|
70 if (iexp != igot)
|
|
71 abort ();
|
|
72
|
|
73 igot = 0;
|
|
74 iexp = 32;
|
|
75
|
|
76 #pragma acc data copy (igot, idata[0:N])
|
|
77 {
|
|
78 #pragma acc parallel loop
|
|
79 for (i = 0; i < N; i++)
|
|
80 #pragma acc atomic capture
|
|
81 { ++igot; idata[i] = igot; }
|
|
82 }
|
|
83
|
|
84 if (iexp != igot)
|
|
85 abort ();
|
|
86
|
|
87 igot = 0;
|
|
88 iexp = 32;
|
|
89
|
|
90 #pragma acc data copy (igot, idata[0:N])
|
|
91 {
|
|
92 #pragma acc parallel loop
|
|
93 for (i = 0; i < N; i++)
|
|
94 #pragma acc atomic capture
|
|
95 { igot++; idata[i] = igot; }
|
|
96 }
|
|
97
|
|
98 if (iexp != igot)
|
|
99 abort ();
|
|
100
|
|
101 igot = 32;
|
|
102 iexp = 0;
|
|
103
|
|
104 #pragma acc data copy (igot, idata[0:N])
|
|
105 {
|
|
106 #pragma acc parallel loop
|
|
107 for (i = 0; i < N; i++)
|
|
108 #pragma acc atomic capture
|
|
109 { idata[i] = igot; igot--; }
|
|
110 }
|
|
111
|
|
112 if (iexp != igot)
|
|
113 abort ();
|
|
114
|
|
115 igot = 32;
|
|
116 iexp = 0;
|
|
117
|
|
118 #pragma acc data copy (igot, idata[0:N])
|
|
119 {
|
|
120 #pragma acc parallel loop
|
|
121 for (i = 0; i < N; i++)
|
|
122 #pragma acc atomic capture
|
|
123 { idata[i] = igot; --igot; }
|
|
124 }
|
|
125
|
|
126 if (iexp != igot)
|
|
127 abort ();
|
|
128
|
|
129 igot = 32;
|
|
130 iexp = 0;
|
|
131
|
|
132 #pragma acc data copy (igot, idata[0:N])
|
|
133 {
|
|
134 #pragma acc parallel loop
|
|
135 for (i = 0; i < N; i++)
|
|
136 #pragma acc atomic capture
|
|
137 { --igot; idata[i] = igot; }
|
|
138 }
|
|
139
|
|
140 if (iexp != igot)
|
|
141 abort ();
|
|
142
|
|
143 igot = 32;
|
|
144 iexp = 0;
|
|
145
|
|
146 #pragma acc data copy (igot, idata[0:N])
|
|
147 {
|
|
148 #pragma acc parallel loop
|
|
149 for (i = 0; i < N; i++)
|
|
150 #pragma acc atomic capture
|
|
151 { igot--; idata[i] = igot; }
|
|
152 }
|
|
153
|
|
154 if (iexp != igot)
|
|
155 abort ();
|
|
156
|
|
157 /* BINOP = + */
|
|
158 igot = 0;
|
|
159 iexp = 32;
|
|
160
|
|
161 #pragma acc data copy (igot, idata[0:N])
|
|
162 {
|
|
163 #pragma acc parallel loop
|
|
164 for (i = 0; i < N; i++)
|
|
165 {
|
|
166 int expr = 1;
|
|
167
|
|
168 #pragma acc atomic capture
|
|
169 { idata[i] = igot; igot += expr; }
|
|
170 }
|
|
171 }
|
|
172
|
|
173 if (iexp != igot)
|
|
174 abort ();
|
|
175
|
|
176 igot = 0;
|
|
177 iexp = 32;
|
|
178
|
|
179 #pragma acc data copy (igot, idata[0:N])
|
|
180 {
|
|
181 #pragma acc parallel loop
|
|
182 for (i = 0; i < N; i++)
|
|
183 {
|
|
184 int expr = 1;
|
|
185
|
|
186 #pragma acc atomic capture
|
|
187 { igot += expr; idata[i] = igot; }
|
|
188 }
|
|
189 }
|
|
190
|
|
191 if (iexp != igot)
|
|
192 abort ();
|
|
193
|
|
194 igot = 0;
|
|
195 iexp = 32;
|
|
196
|
|
197 #pragma acc data copy (igot, idata[0:N])
|
|
198 {
|
|
199 #pragma acc parallel loop
|
|
200 for (i = 0; i < N; i++)
|
|
201 {
|
|
202 int expr = 1;
|
|
203
|
|
204 #pragma acc atomic capture
|
|
205 { idata[i] = igot; igot = igot + expr; }
|
|
206 }
|
|
207 }
|
|
208
|
|
209 if (iexp != igot)
|
|
210 abort ();
|
|
211
|
|
212 igot = 0;
|
|
213 iexp = 32;
|
|
214
|
|
215 #pragma acc data copy (igot, idata[0:N])
|
|
216 {
|
|
217 #pragma acc parallel loop
|
|
218 for (i = 0; i < N; i++)
|
|
219 {
|
|
220 int expr = 1;
|
|
221
|
|
222 #pragma acc atomic capture
|
|
223 { idata[i] = igot; igot = expr + igot; }
|
|
224 }
|
|
225 }
|
|
226
|
|
227 if (iexp != igot)
|
|
228 abort ();
|
|
229
|
|
230 igot = 0;
|
|
231 iexp = 32;
|
|
232
|
|
233 #pragma acc data copy (igot, idata[0:N])
|
|
234 {
|
|
235 #pragma acc parallel loop
|
|
236 for (i = 0; i < N; i++)
|
|
237 {
|
|
238 int expr = 1;
|
|
239
|
|
240 #pragma acc atomic capture
|
|
241 { igot = igot + expr; idata[i] = igot; }
|
|
242 }
|
|
243 }
|
|
244
|
|
245 if (iexp != igot)
|
|
246 abort ();
|
|
247
|
|
248
|
|
249 igot = 0;
|
|
250 iexp = 32;
|
|
251
|
|
252 #pragma acc data copy (igot, idata[0:N])
|
|
253 {
|
|
254 #pragma acc parallel loop
|
|
255 for (i = 0; i < N; i++)
|
|
256 {
|
|
257 int expr = 1;
|
|
258
|
|
259 #pragma acc atomic capture
|
|
260 { igot = expr + igot; idata[i] = igot; }
|
|
261 }
|
|
262 }
|
|
263
|
|
264 if (iexp != igot)
|
|
265 abort ();
|
|
266
|
|
267 /* BINOP = * */
|
|
268 lgot = 1LL;
|
|
269 lexp = 1LL << 32;
|
|
270
|
|
271 #pragma acc data copy (lgot, ldata[0:N])
|
|
272 {
|
|
273 #pragma acc parallel loop
|
|
274 for (i = 0; i < N; i++)
|
|
275 {
|
|
276 long long expr = 2LL;
|
|
277
|
|
278 #pragma acc atomic capture
|
|
279 { ldata[i] = lgot; lgot *= expr; }
|
|
280 }
|
|
281 }
|
|
282
|
|
283 if (lexp != lgot)
|
|
284 abort ();
|
|
285
|
|
286 lgot = 1LL;
|
|
287 lexp = 1LL << 32;
|
|
288
|
|
289 #pragma acc data copy (lgot, ldata[0:N])
|
|
290 {
|
|
291 #pragma acc parallel loop
|
|
292 for (i = 0; i < N; i++)
|
|
293 {
|
|
294 long long expr = 2LL;
|
|
295
|
|
296 #pragma acc atomic capture
|
|
297 { lgot *= expr; ldata[i] = lgot; }
|
|
298 }
|
|
299 }
|
|
300
|
|
301 if (lexp != lgot)
|
|
302 abort ();
|
|
303
|
|
304 lgot = 1LL;
|
|
305 lexp = 1LL << 32;
|
|
306
|
|
307 #pragma acc data copy (lgot, ldata[0:N])
|
|
308 {
|
|
309 #pragma acc parallel loop
|
|
310 for (i = 0; i < N; i++)
|
|
311 {
|
|
312 long long expr = 2LL;
|
|
313
|
|
314 #pragma acc atomic capture
|
|
315 { ldata[i] = lgot; lgot = lgot * expr; }
|
|
316 }
|
|
317 }
|
|
318
|
|
319 if (lexp != lgot)
|
|
320 abort ();
|
|
321
|
|
322 lgot = 1LL;
|
|
323 lexp = 1LL << 32;
|
|
324
|
|
325 #pragma acc data copy (lgot, ldata[0:N])
|
|
326 {
|
|
327 #pragma acc parallel loop
|
|
328 for (i = 0; i < N; i++)
|
|
329 {
|
|
330 long long expr = 2LL;
|
|
331
|
|
332 #pragma acc atomic capture
|
|
333 { ldata[i] = lgot; lgot = expr * lgot; }
|
|
334 }
|
|
335 }
|
|
336
|
|
337 if (lexp != lgot)
|
|
338 abort ();
|
|
339
|
|
340 lgot = 1LL;
|
|
341 lexp = 1LL << 32;
|
|
342
|
|
343 #pragma acc data copy (lgot, ldata[0:N])
|
|
344 {
|
|
345 #pragma acc parallel loop
|
|
346 for (i = 0; i < N; i++)
|
|
347 {
|
|
348 long long expr = 2LL;
|
|
349
|
|
350 #pragma acc atomic capture
|
|
351 { lgot = lgot * expr; ldata[i] = lgot; }
|
|
352 }
|
|
353 }
|
|
354
|
|
355 if (lexp != lgot)
|
|
356 abort ();
|
|
357
|
|
358 lgot = 1LL;
|
|
359 lexp = 1LL << 32;
|
|
360
|
|
361 #pragma acc data copy (lgot, ldata[0:N])
|
|
362 {
|
|
363 #pragma acc parallel loop
|
|
364 for (i = 0; i < N; i++)
|
|
365 {
|
|
366 long long expr = 2;
|
|
367
|
|
368 #pragma acc atomic capture
|
|
369 { lgot = expr * lgot; ldata[i] = lgot; }
|
|
370 }
|
|
371 }
|
|
372
|
|
373 if (lexp != lgot)
|
|
374 abort ();
|
|
375
|
|
376 /* BINOP = - */
|
|
377 igot = 32;
|
|
378 iexp = 0;
|
|
379
|
|
380 #pragma acc data copy (igot, idata[0:N])
|
|
381 {
|
|
382 #pragma acc parallel loop
|
|
383 for (i = 0; i < N; i++)
|
|
384 {
|
|
385 int expr = 1;
|
|
386
|
|
387 #pragma acc atomic capture
|
|
388 { idata[i] = igot; igot -= expr; }
|
|
389 }
|
|
390 }
|
|
391
|
|
392 if (iexp != igot)
|
|
393 abort ();
|
|
394
|
|
395 igot = 32;
|
|
396 iexp = 0;
|
|
397
|
|
398 #pragma acc data copy (igot, idata[0:N])
|
|
399 {
|
|
400 #pragma acc parallel loop
|
|
401 for (i = 0; i < N; i++)
|
|
402 {
|
|
403 int expr = 1;
|
|
404
|
|
405 #pragma acc atomic capture
|
|
406 { igot -= expr; idata[i] = igot; }
|
|
407 }
|
|
408 }
|
|
409
|
|
410 if (iexp != igot)
|
|
411 abort ();
|
|
412
|
|
413 igot = 32;
|
|
414 iexp = 0;
|
|
415
|
|
416 #pragma acc data copy (igot, idata[0:N])
|
|
417 {
|
|
418 #pragma acc parallel loop
|
|
419 for (i = 0; i < N; i++)
|
|
420 {
|
|
421 int expr = 1;
|
|
422
|
|
423 #pragma acc atomic capture
|
|
424 { idata[i] = igot; igot = igot - expr; }
|
|
425 }
|
|
426 }
|
|
427
|
|
428 if (iexp != igot)
|
|
429 abort ();
|
|
430
|
|
431 igot = 1;
|
|
432 iexp = 1;
|
|
433
|
|
434 #pragma acc data copy (igot, idata[0:N])
|
|
435 {
|
|
436 #pragma acc parallel loop
|
|
437 for (i = 0; i < N; i++)
|
|
438 {
|
|
439 int expr = 1;
|
|
440
|
|
441 #pragma acc atomic capture
|
|
442 { idata[i] = igot; igot = expr - igot; }
|
|
443 }
|
|
444 }
|
|
445
|
|
446 for (i = 0; i < N; i++)
|
|
447 if (i % 2 == 0)
|
|
448 {
|
|
449 if (idata[i] != 1)
|
|
450 abort ();
|
|
451 }
|
|
452 else
|
|
453 {
|
|
454 if (idata[i] != 0)
|
|
455 abort ();
|
|
456 }
|
|
457
|
|
458 if (iexp != igot)
|
|
459 abort ();
|
|
460
|
|
461 igot = 1;
|
|
462 iexp = -31;
|
|
463
|
|
464 #pragma acc data copy (igot, idata[0:N])
|
|
465 {
|
|
466 #pragma acc parallel loop
|
|
467 for (i = 0; i < N; i++)
|
|
468 {
|
|
469 int expr = 1;
|
|
470
|
|
471 #pragma acc atomic capture
|
|
472 { igot = igot - expr; idata[i] = igot; }
|
|
473 }
|
|
474 }
|
|
475
|
|
476 if (iexp != igot)
|
|
477 abort ();
|
|
478
|
|
479 igot = 1;
|
|
480 iexp = 1;
|
|
481
|
|
482 #pragma acc data copy (igot, idata[0:N])
|
|
483 {
|
|
484 #pragma acc parallel loop
|
|
485 for (i = 0; i < N; i++)
|
|
486 {
|
|
487 int expr = 1;
|
|
488
|
|
489 #pragma acc atomic capture
|
|
490 { igot = expr - igot; idata[i] = igot; }
|
|
491 }
|
|
492 }
|
|
493
|
|
494 for (i = 0; i < N; i++)
|
|
495 if (i % 2 == 0)
|
|
496 {
|
|
497 if (idata[i] != 0)
|
|
498 abort ();
|
|
499 }
|
|
500 else
|
|
501 {
|
|
502 if (idata[i] != 1)
|
|
503 abort ();
|
|
504 }
|
|
505
|
|
506 if (iexp != igot)
|
|
507 abort ();
|
|
508
|
|
509 /* BINOP = / */
|
|
510 lgot = 1LL << 32;
|
|
511 lexp = 1LL;
|
|
512
|
|
513 #pragma acc data copy (lgot, ldata[0:N])
|
|
514 {
|
|
515 #pragma acc parallel loop
|
|
516 for (i = 0; i < N; i++)
|
|
517 {
|
|
518 long long expr = 2LL;
|
|
519
|
|
520 #pragma acc atomic capture
|
|
521 { ldata[i] = lgot; lgot /= expr; }
|
|
522 }
|
|
523 }
|
|
524
|
|
525 if (lexp != lgot)
|
|
526 abort ();
|
|
527
|
|
528 lgot = 1LL << 32;
|
|
529 lexp = 1LL;
|
|
530
|
|
531 #pragma acc data copy (lgot, ldata[0:N])
|
|
532 {
|
|
533 #pragma acc parallel loop
|
|
534 for (i = 0; i < N; i++)
|
|
535 {
|
|
536 long long expr = 2LL;
|
|
537
|
|
538 #pragma acc atomic capture
|
|
539 { lgot /= expr; ldata[i] = lgot; }
|
|
540 }
|
|
541 }
|
|
542
|
|
543 if (lexp != lgot)
|
|
544 abort ();
|
|
545
|
|
546 lgot = 1LL << 32;
|
|
547 lexp = 1LL;
|
|
548
|
|
549 #pragma acc data copy (lgot, ldata[0:N])
|
|
550 {
|
|
551 #pragma acc parallel loop
|
|
552 for (i = 0; i < N; i++)
|
|
553 {
|
|
554 long long expr = 2LL;
|
|
555
|
|
556 #pragma acc atomic capture
|
|
557 { ldata[i] = lgot; lgot = lgot / expr; }
|
|
558 }
|
|
559 }
|
|
560
|
|
561 if (lexp != lgot)
|
|
562 abort ();
|
|
563
|
|
564 lgot = 2LL;
|
|
565 lexp = 2LL;
|
|
566
|
|
567 #pragma acc data copy (lgot, ldata[0:N])
|
|
568 {
|
|
569 #pragma acc parallel loop
|
|
570 for (i = 0; i < N; i++)
|
|
571 {
|
|
572 long long expr = 1LL << N;
|
|
573
|
|
574 #pragma acc atomic capture
|
|
575 { ldata[i] = lgot; lgot = expr / lgot; }
|
|
576 }
|
|
577 }
|
|
578
|
|
579 if (lexp != lgot)
|
|
580 abort ();
|
|
581
|
|
582 lgot = 2LL;
|
|
583 lexp = 2LL;
|
|
584
|
|
585 #pragma acc data copy (lgot, ldata[0:N])
|
|
586 {
|
|
587 #pragma acc parallel loop
|
|
588 for (i = 0; i < N; i++)
|
|
589 {
|
|
590 long long expr = 1LL << N;
|
|
591
|
|
592 #pragma acc atomic capture
|
|
593 { lgot = lgot / expr; ldata[i] = lgot; }
|
|
594 }
|
|
595 }
|
|
596
|
|
597 if (lexp != lgot)
|
|
598 abort ();
|
|
599
|
|
600 lgot = 2LL;
|
|
601 lexp = 2LL;
|
|
602
|
|
603 #pragma acc data copy (lgot, ldata[0:N])
|
|
604 {
|
|
605 #pragma acc parallel loop
|
|
606 for (i = 0; i < N; i++)
|
|
607 {
|
|
608 long long expr = 1LL << N;
|
|
609
|
|
610 #pragma acc atomic capture
|
|
611 { lgot = expr / lgot; ldata[i] = lgot; }
|
|
612 }
|
|
613 }
|
|
614
|
|
615 if (lexp != lgot)
|
|
616 abort ();
|
|
617
|
|
618 /* BINOP = & */
|
|
619 lgot = ~0LL;
|
|
620 lexp = 0LL;
|
|
621
|
|
622 #pragma acc data copy (lgot, ldata[0:N])
|
|
623 {
|
|
624 #pragma acc parallel loop
|
|
625 for (i = 0; i < N; i++)
|
|
626 {
|
|
627 long long expr = ~(1 << i);
|
|
628
|
|
629 #pragma acc atomic capture
|
|
630 { ldata[i] = lgot; lgot &= expr; }
|
|
631 }
|
|
632 }
|
|
633
|
|
634 if (lexp != lgot)
|
|
635 abort ();
|
|
636
|
|
637 lgot = ~0LL;
|
|
638 iexp = 0LL;
|
|
639
|
|
640 #pragma acc data copy (lgot, ldata[0:N])
|
|
641 {
|
|
642 #pragma acc parallel loop
|
|
643 for (i = 0; i < N; i++)
|
|
644 {
|
|
645 long long expr = ~(1 << i);
|
|
646
|
|
647 #pragma acc atomic capture
|
|
648 { lgot &= expr; ldata[i] = lgot; }
|
|
649 }
|
|
650 }
|
|
651
|
|
652 if (lexp != lgot)
|
|
653 abort ();
|
|
654
|
|
655 lgot = ~0LL;
|
|
656 lexp = 0LL;
|
|
657
|
|
658 #pragma acc data copy (lgot, ldata[0:N])
|
|
659 {
|
|
660 #pragma acc parallel loop
|
|
661 for (i = 0; i < N; i++)
|
|
662 {
|
|
663 long long expr = ~(1 << i);
|
|
664
|
|
665 #pragma acc atomic capture
|
|
666 { ldata[i] = lgot; lgot = lgot & expr; }
|
|
667 }
|
|
668 }
|
|
669
|
|
670 if (lexp != lgot)
|
|
671 abort ();
|
|
672
|
|
673 lgot = ~0LL;
|
|
674 lexp = 0LL;
|
|
675
|
|
676 #pragma acc data copy (lgot, ldata[0:N])
|
|
677 {
|
|
678 #pragma acc parallel loop
|
|
679 for (i = 0; i < N; i++)
|
|
680 {
|
|
681 long long expr = ~(1 << i);
|
|
682
|
|
683 #pragma acc atomic capture
|
|
684 { ldata[i] = lgot; lgot = expr & lgot; }
|
|
685 }
|
|
686 }
|
|
687
|
|
688 if (lexp != lgot)
|
|
689 abort ();
|
|
690
|
|
691 lgot = ~0LL;
|
|
692 iexp = 0LL;
|
|
693
|
|
694 #pragma acc data copy (lgot, ldata[0:N])
|
|
695 {
|
|
696 #pragma acc parallel loop
|
|
697 for (i = 0; i < N; i++)
|
|
698 {
|
|
699 long long expr = ~(1 << i);
|
|
700
|
|
701 #pragma acc atomic capture
|
|
702 { lgot = lgot & expr; ldata[i] = lgot; }
|
|
703 }
|
|
704 }
|
|
705
|
|
706 if (lexp != lgot)
|
|
707 abort ();
|
|
708
|
|
709 lgot = ~0LL;
|
|
710 lexp = 0LL;
|
|
711
|
|
712 #pragma acc data copy (lgot, ldata[0:N])
|
|
713 {
|
|
714 #pragma acc parallel loop
|
|
715 for (i = 0; i < N; i++)
|
|
716 {
|
|
717 long long expr = ~(1 << i);
|
|
718
|
|
719 #pragma acc atomic capture
|
|
720 { lgot = expr & lgot; ldata[i] = lgot; }
|
|
721 }
|
|
722 }
|
|
723
|
|
724 if (lexp != lgot)
|
|
725 abort ();
|
|
726
|
|
727 /* BINOP = ^ */
|
|
728 lgot = ~0LL;
|
|
729 lexp = 0LL;
|
|
730
|
|
731 #pragma acc data copy (lgot, ldata[0:N])
|
|
732 {
|
|
733 #pragma acc parallel loop
|
|
734 for (i = 0; i < N; i++)
|
|
735 {
|
|
736 long long expr = 1 << i;
|
|
737
|
|
738 #pragma acc atomic capture
|
|
739 { ldata[i] = lgot; lgot ^= expr; }
|
|
740 }
|
|
741 }
|
|
742
|
|
743 if (lexp != lgot)
|
|
744 abort ();
|
|
745
|
|
746 lgot = ~0LL;
|
|
747 iexp = 0LL;
|
|
748
|
|
749 #pragma acc data copy (lgot, ldata[0:N])
|
|
750 {
|
|
751 #pragma acc parallel loop
|
|
752 for (i = 0; i < N; i++)
|
|
753 {
|
|
754 long long expr = ~(1 << i);
|
|
755
|
|
756 #pragma acc atomic capture
|
|
757 { lgot ^= expr; ldata[i] = lgot; }
|
|
758 }
|
|
759 }
|
|
760
|
|
761 if (lexp != lgot)
|
|
762 abort ();
|
|
763
|
|
764 lgot = ~0LL;
|
|
765 lexp = 0LL;
|
|
766
|
|
767 #pragma acc data copy (lgot, ldata[0:N])
|
|
768 {
|
|
769 #pragma acc parallel loop
|
|
770 for (i = 0; i < N; i++)
|
|
771 {
|
|
772 long long expr = ~(1 << i);
|
|
773
|
|
774 #pragma acc atomic capture
|
|
775 { ldata[i] = lgot; lgot = lgot ^ expr; }
|
|
776 }
|
|
777 }
|
|
778
|
|
779 if (lexp != lgot)
|
|
780 abort ();
|
|
781
|
|
782 lgot = ~0LL;
|
|
783 lexp = 0LL;
|
|
784
|
|
785 #pragma acc data copy (lgot, ldata[0:N])
|
|
786 {
|
|
787 #pragma acc parallel loop
|
|
788 for (i = 0; i < N; i++)
|
|
789 {
|
|
790 long long expr = ~(1 << i);
|
|
791
|
|
792 #pragma acc atomic capture
|
|
793 { ldata[i] = lgot; lgot = expr ^ lgot; }
|
|
794 }
|
|
795 }
|
|
796
|
|
797 if (lexp != lgot)
|
|
798 abort ();
|
|
799
|
|
800 lgot = ~0LL;
|
|
801 iexp = 0LL;
|
|
802
|
|
803 #pragma acc data copy (lgot, ldata[0:N])
|
|
804 {
|
|
805 #pragma acc parallel loop
|
|
806 for (i = 0; i < N; i++)
|
|
807 {
|
|
808 long long expr = ~(1 << i);
|
|
809
|
|
810 #pragma acc atomic capture
|
|
811 { lgot = lgot ^ expr; ldata[i] = lgot; }
|
|
812 }
|
|
813 }
|
|
814
|
|
815 if (lexp != lgot)
|
|
816 abort ();
|
|
817
|
|
818 lgot = ~0LL;
|
|
819 lexp = 0LL;
|
|
820
|
|
821 #pragma acc data copy (lgot, ldata[0:N])
|
|
822 {
|
|
823 #pragma acc parallel loop
|
|
824 for (i = 0; i < N; i++)
|
|
825 {
|
|
826 long long expr = ~(1 << i);
|
|
827
|
|
828 #pragma acc atomic capture
|
|
829 { lgot = expr ^ lgot; ldata[i] = lgot; }
|
|
830 }
|
|
831 }
|
|
832
|
|
833 if (lexp != lgot)
|
|
834 abort ();
|
|
835
|
|
836 /* BINOP = | */
|
|
837 lgot = 0LL;
|
|
838 lexp = ~0LL;
|
|
839
|
|
840 #pragma acc data copy (lgot, ldata[0:N])
|
|
841 {
|
|
842 #pragma acc parallel loop
|
|
843 for (i = 0; i < N; i++)
|
|
844 {
|
|
845 long long expr = 1 << i;
|
|
846
|
|
847 #pragma acc atomic capture
|
|
848 { ldata[i] = lgot; lgot |= expr; }
|
|
849 }
|
|
850 }
|
|
851
|
|
852 if (lexp != lgot)
|
|
853 abort ();
|
|
854
|
|
855 lgot = 0LL;
|
|
856 iexp = ~0LL;
|
|
857
|
|
858 #pragma acc data copy (lgot, ldata[0:N])
|
|
859 {
|
|
860 #pragma acc parallel loop
|
|
861 for (i = 0; i < N; i++)
|
|
862 {
|
|
863 long long expr = ~(1 << i);
|
|
864
|
|
865 #pragma acc atomic capture
|
|
866 { lgot |= expr; ldata[i] = lgot; }
|
|
867 }
|
|
868 }
|
|
869
|
|
870 if (lexp != lgot)
|
|
871 abort ();
|
|
872
|
|
873 lgot = 0LL;
|
|
874 lexp = ~0LL;
|
|
875
|
|
876 #pragma acc data copy (lgot, ldata[0:N])
|
|
877 {
|
|
878 #pragma acc parallel loop
|
|
879 for (i = 0; i < N; i++)
|
|
880 {
|
|
881 long long expr = ~(1 << i);
|
|
882
|
|
883 #pragma acc atomic capture
|
|
884 { ldata[i] = lgot; lgot = lgot | expr; }
|
|
885 }
|
|
886 }
|
|
887
|
|
888 if (lexp != lgot)
|
|
889 abort ();
|
|
890
|
|
891 lgot = 0LL;
|
|
892 lexp = ~0LL;
|
|
893
|
|
894 #pragma acc data copy (lgot, ldata[0:N])
|
|
895 {
|
|
896 #pragma acc parallel loop
|
|
897 for (i = 0; i < N; i++)
|
|
898 {
|
|
899 long long expr = ~(1 << i);
|
|
900
|
|
901 #pragma acc atomic capture
|
|
902 { ldata[i] = lgot; lgot = expr | lgot; }
|
|
903 }
|
|
904 }
|
|
905
|
|
906 if (lexp != lgot)
|
|
907 abort ();
|
|
908
|
|
909 lgot = 0LL;
|
|
910 iexp = ~0LL;
|
|
911
|
|
912 #pragma acc data copy (lgot, ldata[0:N])
|
|
913 {
|
|
914 #pragma acc parallel loop
|
|
915 for (i = 0; i < N; i++)
|
|
916 {
|
|
917 long long expr = ~(1 << i);
|
|
918
|
|
919 #pragma acc atomic capture
|
|
920 { lgot = lgot | expr; ldata[i] = lgot; }
|
|
921 }
|
|
922 }
|
|
923
|
|
924 if (lexp != lgot)
|
|
925 abort ();
|
|
926
|
|
927 lgot = 0LL;
|
|
928 lexp = ~0LL;
|
|
929
|
|
930 #pragma acc data copy (lgot, ldata[0:N])
|
|
931 {
|
|
932 #pragma acc parallel loop
|
|
933 for (i = 0; i < N; i++)
|
|
934 {
|
|
935 long long expr = ~(1 << i);
|
|
936
|
|
937 #pragma acc atomic capture
|
|
938 { lgot = expr | lgot; ldata[i] = lgot; }
|
|
939 }
|
|
940 }
|
|
941
|
|
942 if (lexp != lgot)
|
|
943 abort ();
|
|
944
|
|
945 /* BINOP = << */
|
|
946 lgot = 1LL;
|
|
947 lexp = 1LL << N;
|
|
948
|
|
949 #pragma acc data copy (lgot, ldata[0:N])
|
|
950 {
|
|
951 #pragma acc parallel loop
|
|
952 for (i = 0; i < N; i++)
|
|
953 {
|
|
954 long long expr = 1LL;
|
|
955
|
|
956 #pragma acc atomic capture
|
|
957 { ldata[i] = lgot; lgot <<= expr; }
|
|
958 }
|
|
959 }
|
|
960
|
|
961 if (lexp != lgot)
|
|
962 abort ();
|
|
963
|
|
964 lgot = 1LL;
|
|
965 iexp = 1LL << N;
|
|
966
|
|
967 #pragma acc data copy (lgot, ldata[0:N])
|
|
968 {
|
|
969 #pragma acc parallel loop
|
|
970 for (i = 0; i < N; i++)
|
|
971 {
|
|
972 long long expr = 1LL;
|
|
973
|
|
974 #pragma acc atomic capture
|
|
975 { lgot <<= expr; ldata[i] = lgot; }
|
|
976 }
|
|
977 }
|
|
978
|
|
979 if (lexp != lgot)
|
|
980 abort ();
|
|
981
|
|
982 lgot = 1LL;
|
|
983 lexp = 1LL << N;
|
|
984
|
|
985 #pragma acc data copy (lgot, ldata[0:N])
|
|
986 {
|
|
987 #pragma acc parallel loop
|
|
988 for (i = 0; i < N; i++)
|
|
989 {
|
|
990 long long expr = 1LL;
|
|
991
|
|
992 #pragma acc atomic capture
|
|
993 { ldata[i] = lgot; lgot = lgot << expr; }
|
|
994 }
|
|
995 }
|
|
996
|
|
997 if (lexp != lgot)
|
|
998 abort ();
|
|
999
|
|
1000 lgot = 1LL;
|
|
1001 lexp = 2LL;
|
|
1002
|
|
1003 #pragma acc data copy (lgot, ldata[0:N])
|
|
1004 {
|
|
1005 #pragma acc parallel loop
|
|
1006 for (i = 0; i < 1; i++)
|
|
1007 {
|
|
1008 long long expr = 1LL;
|
|
1009
|
|
1010 #pragma acc atomic capture
|
|
1011 { ldata[i] = lgot; lgot = expr << lgot; }
|
|
1012 }
|
|
1013 }
|
|
1014
|
|
1015 if (lexp != lgot)
|
|
1016 abort ();
|
|
1017
|
|
1018 lgot = 1LL;
|
|
1019 lexp = 2LL;
|
|
1020
|
|
1021 #pragma acc data copy (lgot, ldata[0:N])
|
|
1022 {
|
|
1023 #pragma acc parallel loop
|
|
1024 for (i = 0; i < 1; i++)
|
|
1025 {
|
|
1026 long long expr = 1LL;
|
|
1027
|
|
1028 #pragma acc atomic capture
|
|
1029 { lgot = lgot << expr; ldata[i] = lgot; }
|
|
1030 }
|
|
1031 }
|
|
1032
|
|
1033 if (lexp != lgot)
|
|
1034 abort ();
|
|
1035
|
|
1036 lgot = 1LL;
|
|
1037 lexp = 2LL;
|
|
1038
|
|
1039 #pragma acc data copy (lgot, ldata[0:N])
|
|
1040 {
|
|
1041 #pragma acc parallel loop
|
|
1042 for (i = 0; i < 1; i++)
|
|
1043 {
|
|
1044 long long expr = 1LL;
|
|
1045
|
|
1046 #pragma acc atomic capture
|
|
1047 { lgot = expr << lgot; ldata[i] = lgot; }
|
|
1048 }
|
|
1049 }
|
|
1050
|
|
1051 if (lexp != lgot)
|
|
1052 abort ();
|
|
1053
|
|
1054 /* BINOP = >> */
|
|
1055 lgot = 1LL << N;
|
|
1056 lexp = 1LL;
|
|
1057
|
|
1058 #pragma acc data copy (lgot, ldata[0:N])
|
|
1059 {
|
|
1060 #pragma acc parallel loop
|
|
1061 for (i = 0; i < N; i++)
|
|
1062 {
|
|
1063 long long expr = 1LL;
|
|
1064
|
|
1065 #pragma acc atomic capture
|
|
1066 { ldata[i] = lgot; lgot >>= expr; }
|
|
1067 }
|
|
1068 }
|
|
1069
|
|
1070 if (lexp != lgot)
|
|
1071 abort ();
|
|
1072
|
|
1073 lgot = 1LL << N;
|
|
1074 iexp = 1LL;
|
|
1075
|
|
1076 #pragma acc data copy (lgot, ldata[0:N])
|
|
1077 {
|
|
1078 #pragma acc parallel loop
|
|
1079 for (i = 0; i < N; i++)
|
|
1080 {
|
|
1081 long long expr = 1LL;
|
|
1082
|
|
1083 #pragma acc atomic capture
|
|
1084 { lgot >>= expr; ldata[i] = lgot; }
|
|
1085 }
|
|
1086 }
|
|
1087
|
|
1088 if (lexp != lgot)
|
|
1089 abort ();
|
|
1090
|
|
1091 lgot = 1LL << N;
|
|
1092 lexp = 1LL;
|
|
1093
|
|
1094 #pragma acc data copy (lgot, ldata[0:N])
|
|
1095 {
|
|
1096 #pragma acc parallel loop
|
|
1097 for (i = 0; i < N; i++)
|
|
1098 {
|
|
1099 long long expr = 1LL;
|
|
1100
|
|
1101 #pragma acc atomic capture
|
|
1102 { ldata[i] = lgot; lgot = lgot >> expr; }
|
|
1103 }
|
|
1104 }
|
|
1105
|
|
1106 if (lexp != lgot)
|
|
1107 abort ();
|
|
1108
|
|
1109 lgot = 1LL;
|
|
1110 lexp = 1LL << (N - 1);
|
|
1111
|
|
1112 #pragma acc data copy (lgot, ldata[0:N])
|
|
1113 {
|
|
1114 #pragma acc parallel loop
|
|
1115 for (i = 0; i < 1; i++)
|
|
1116 {
|
|
1117 long long expr = 1LL << N;
|
|
1118
|
|
1119 #pragma acc atomic capture
|
|
1120 { ldata[i] = lgot; lgot = expr >> lgot; }
|
|
1121 }
|
|
1122 }
|
|
1123
|
|
1124 if (lexp != lgot)
|
|
1125 abort ();
|
|
1126
|
|
1127 lgot = 1LL << N;
|
|
1128 lexp = 1LL;
|
|
1129
|
|
1130 #pragma acc data copy (lgot, ldata[0:N])
|
|
1131 {
|
|
1132 #pragma acc parallel loop
|
|
1133 for (i = 0; i < N; i++)
|
|
1134 {
|
|
1135 long long expr = 1LL;
|
|
1136
|
|
1137 #pragma acc atomic capture
|
|
1138 { lgot = lgot >> expr; ldata[i] = lgot; }
|
|
1139 }
|
|
1140 }
|
|
1141
|
|
1142 if (lexp != lgot)
|
|
1143 abort ();
|
|
1144
|
|
1145 lgot = 1LL;
|
|
1146 lexp = 1LL << (N - 1);
|
|
1147
|
|
1148 #pragma acc data copy (lgot, ldata[0:N])
|
|
1149 {
|
|
1150 #pragma acc parallel loop
|
|
1151 for (i = 0; i < 1; i++)
|
|
1152 {
|
|
1153 long long expr = 1LL << N;
|
|
1154
|
|
1155 #pragma acc atomic capture
|
|
1156 { lgot = expr >> lgot; ldata[i] = lgot; }
|
|
1157 }
|
|
1158 }
|
|
1159
|
|
1160 if (lexp != lgot)
|
|
1161 abort ();
|
|
1162
|
|
1163 // FLOAT FLOAT FLOAT
|
|
1164
|
|
1165 /* BINOP = + */
|
|
1166 fgot = 0.0;
|
|
1167 fexp = 32.0;
|
|
1168
|
|
1169 #pragma acc data copy (fgot, fdata[0:N])
|
|
1170 {
|
|
1171 #pragma acc parallel loop
|
|
1172 for (i = 0; i < N; i++)
|
|
1173 {
|
|
1174 float expr = 1.0;
|
|
1175
|
|
1176 #pragma acc atomic capture
|
|
1177 { fdata[i] = fgot; fgot += expr; }
|
|
1178 }
|
|
1179 }
|
|
1180
|
|
1181 if (fexp != fgot)
|
|
1182 abort ();
|
|
1183
|
|
1184 fgot = 0.0;
|
|
1185 fexp = 32.0;
|
|
1186
|
|
1187 #pragma acc data copy (fgot, fdata[0:N])
|
|
1188 {
|
|
1189 #pragma acc parallel loop
|
|
1190 for (i = 0; i < N; i++)
|
|
1191 {
|
|
1192 float expr = 1.0;
|
|
1193
|
|
1194 #pragma acc atomic capture
|
|
1195 { fgot += expr; fdata[i] = fgot; }
|
|
1196 }
|
|
1197 }
|
|
1198
|
|
1199 if (fexp != fgot)
|
|
1200 abort ();
|
|
1201
|
|
1202 fgot = 0.0;
|
|
1203 fexp = 32.0;
|
|
1204
|
|
1205 #pragma acc data copy (fgot, fdata[0:N])
|
|
1206 {
|
|
1207 #pragma acc parallel loop
|
|
1208 for (i = 0; i < N; i++)
|
|
1209 {
|
|
1210 float expr = 1.0;
|
|
1211
|
|
1212 #pragma acc atomic capture
|
|
1213 { idata[i] = fgot; fgot = fgot + expr; }
|
|
1214 }
|
|
1215 }
|
|
1216
|
|
1217 if (fexp != fgot)
|
|
1218 abort ();
|
|
1219
|
|
1220 fgot = 0.0;
|
|
1221 fexp = 32.0;
|
|
1222
|
|
1223 #pragma acc data copy (fgot, fdata[0:N])
|
|
1224 {
|
|
1225 #pragma acc parallel loop
|
|
1226 for (i = 0; i < N; i++)
|
|
1227 {
|
|
1228 float expr = 1.0;
|
|
1229
|
|
1230 #pragma acc atomic capture
|
|
1231 { fdata[i] = fgot; fgot = expr + fgot; }
|
|
1232 }
|
|
1233 }
|
|
1234
|
|
1235 if (fexp != fgot)
|
|
1236 abort ();
|
|
1237
|
|
1238 fgot = 0.0;
|
|
1239 fexp = 32.0;
|
|
1240
|
|
1241 #pragma acc data copy (fgot, fdata[0:N])
|
|
1242 {
|
|
1243 #pragma acc parallel loop
|
|
1244 for (i = 0; i < N; i++)
|
|
1245 {
|
|
1246 float expr = 1.0;
|
|
1247
|
|
1248 #pragma acc atomic capture
|
|
1249 { fgot = fgot + expr; fdata[i] = fgot; }
|
|
1250 }
|
|
1251 }
|
|
1252
|
|
1253 if (fexp != fgot)
|
|
1254 abort ();
|
|
1255
|
|
1256 fgot = 0.0;
|
|
1257 fexp = 32.0;
|
|
1258
|
|
1259 #pragma acc data copy (fgot, fdata[0:N])
|
|
1260 {
|
|
1261 #pragma acc parallel loop
|
|
1262 for (i = 0; i < N; i++)
|
|
1263 {
|
|
1264 float expr = 1.0;
|
|
1265
|
|
1266 #pragma acc atomic capture
|
|
1267 { fgot = expr + fgot; fdata[i] = fgot; }
|
|
1268 }
|
|
1269 }
|
|
1270
|
|
1271 if (fexp != fgot)
|
|
1272 abort ();
|
|
1273
|
|
1274 /* BINOP = * */
|
|
1275 fgot = 1.0;
|
|
1276 fexp = 8192.0*8192.0*64.0;
|
|
1277
|
|
1278 #pragma acc data copy (fgot, fdata[0:N])
|
|
1279 {
|
|
1280 #pragma acc parallel loop
|
|
1281 for (i = 0; i < N; i++)
|
|
1282 {
|
|
1283 float expr = 2.0;
|
|
1284
|
|
1285 #pragma acc atomic capture
|
|
1286 { fdata[i] = fgot; fgot *= expr; }
|
|
1287 }
|
|
1288 }
|
|
1289
|
|
1290 if (fexp != fgot)
|
|
1291 abort ();
|
|
1292
|
|
1293 fgot = 1.0;
|
|
1294 fexp = 8192.0*8192.0*64.0;
|
|
1295
|
|
1296 #pragma acc data copy (fgot, fdata[0:N])
|
|
1297 {
|
|
1298 #pragma acc parallel loop
|
|
1299 for (i = 0; i < N; i++)
|
|
1300 {
|
|
1301 float expr = 2.0;
|
|
1302
|
|
1303 #pragma acc atomic capture
|
|
1304 { fgot *= expr; fdata[i] = fgot; }
|
|
1305 }
|
|
1306 }
|
|
1307
|
|
1308 if (fexp != fgot)
|
|
1309 abort ();
|
|
1310
|
|
1311 fgot = 1.0;
|
|
1312 fexp = 8192.0*8192.0*64.0;
|
|
1313
|
|
1314 #pragma acc data copy (fgot, fdata[0:N])
|
|
1315 {
|
|
1316 #pragma acc parallel loop
|
|
1317 for (i = 0; i < N; i++)
|
|
1318 {
|
|
1319 float expr = 2.0;
|
|
1320
|
|
1321 #pragma acc atomic capture
|
|
1322 { fdata[i] = fgot; fgot = fgot * expr; }
|
|
1323 }
|
|
1324 }
|
|
1325
|
|
1326 if (fexp != fgot)
|
|
1327 abort ();
|
|
1328
|
|
1329 fgot = 1.0;
|
|
1330 fexp = 8192.0*8192.0*64.0;
|
|
1331
|
|
1332 #pragma acc data copy (fgot, fdata[0:N])
|
|
1333 {
|
|
1334 #pragma acc parallel loop
|
|
1335 for (i = 0; i < N; i++)
|
|
1336 {
|
|
1337 float expr = 2.0;
|
|
1338
|
|
1339 #pragma acc atomic capture
|
|
1340 { fdata[i] = fgot; fgot = expr * fgot; }
|
|
1341 }
|
|
1342 }
|
|
1343
|
|
1344 if (fexp != fgot)
|
|
1345 abort ();
|
|
1346
|
|
1347 lgot = 1LL;
|
|
1348 lexp = 1LL << 32;
|
|
1349
|
|
1350 #pragma acc data copy (lgot, ldata[0:N])
|
|
1351 {
|
|
1352 #pragma acc parallel loop
|
|
1353 for (i = 0; i < N; i++)
|
|
1354 {
|
|
1355 long long expr = 2LL;
|
|
1356
|
|
1357 #pragma acc atomic capture
|
|
1358 { lgot = lgot * expr; ldata[i] = lgot; }
|
|
1359 }
|
|
1360 }
|
|
1361
|
|
1362 if (lexp != lgot)
|
|
1363 abort ();
|
|
1364
|
|
1365 fgot = 1.0;
|
|
1366 fexp = 8192.0*8192.0*64.0;
|
|
1367
|
|
1368 #pragma acc data copy (fgot, fdata[0:N])
|
|
1369 {
|
|
1370 #pragma acc parallel loop
|
|
1371 for (i = 0; i < N; i++)
|
|
1372 {
|
|
1373 long long expr = 2;
|
|
1374
|
|
1375 #pragma acc atomic capture
|
|
1376 { fgot = expr * fgot; fdata[i] = fgot; }
|
|
1377 }
|
|
1378 }
|
|
1379
|
|
1380 if (fexp != fgot)
|
|
1381 abort ();
|
|
1382
|
|
1383 /* BINOP = - */
|
|
1384 fgot = 32.0;
|
|
1385 fexp = 0.0;
|
|
1386
|
|
1387 #pragma acc data copy (fgot, fdata[0:N])
|
|
1388 {
|
|
1389 #pragma acc parallel loop
|
|
1390 for (i = 0; i < N; i++)
|
|
1391 {
|
|
1392 float expr = 1.0;
|
|
1393
|
|
1394 #pragma acc atomic capture
|
|
1395 { fdata[i] = fgot; fgot -= expr; }
|
|
1396 }
|
|
1397 }
|
|
1398
|
|
1399 if (fexp != fgot)
|
|
1400 abort ();
|
|
1401
|
|
1402 fgot = 32.0;
|
|
1403 fexp = 0.0;
|
|
1404
|
|
1405 #pragma acc data copy (fgot, fdata[0:N])
|
|
1406 {
|
|
1407 #pragma acc parallel loop
|
|
1408 for (i = 0; i < N; i++)
|
|
1409 {
|
|
1410 float expr = 1.0;
|
|
1411
|
|
1412 #pragma acc atomic capture
|
|
1413 { fgot -= expr; fdata[i] = fgot; }
|
|
1414 }
|
|
1415 }
|
|
1416
|
|
1417 if (fexp != fgot)
|
|
1418 abort ();
|
|
1419
|
|
1420 fgot = 32.0;
|
|
1421 fexp = 0.0;
|
|
1422
|
|
1423 #pragma acc data copy (fgot, fdata[0:N])
|
|
1424 {
|
|
1425 #pragma acc parallel loop
|
|
1426 for (i = 0; i < N; i++)
|
|
1427 {
|
|
1428 float expr = 1.0;
|
|
1429
|
|
1430 #pragma acc atomic capture
|
|
1431 { fdata[i] = fgot; fgot = fgot - expr; }
|
|
1432 }
|
|
1433 }
|
|
1434
|
|
1435 if (fexp != fgot)
|
|
1436 abort ();
|
|
1437
|
|
1438 fgot = 1.0;
|
|
1439 fexp = 1.0;
|
|
1440
|
|
1441 #pragma acc data copy (fgot, fdata[0:N])
|
|
1442 {
|
|
1443 #pragma acc parallel loop
|
|
1444 for (i = 0; i < N; i++)
|
|
1445 {
|
|
1446 float expr = 1.0;
|
|
1447
|
|
1448 #pragma acc atomic capture
|
|
1449 { fdata[i] = fgot; fgot = expr - fgot; }
|
|
1450 }
|
|
1451 }
|
|
1452
|
|
1453 for (i = 0; i < N; i++)
|
|
1454 if (i % 2 == 0)
|
|
1455 {
|
|
1456 if (fdata[i] != 1.0)
|
|
1457 abort ();
|
|
1458 }
|
|
1459 else
|
|
1460 {
|
|
1461 if (fdata[i] != 0.0)
|
|
1462 abort ();
|
|
1463 }
|
|
1464
|
|
1465 if (fexp != fgot)
|
|
1466 abort ();
|
|
1467
|
|
1468 fgot = 1.0;
|
|
1469 fexp = -31.0;
|
|
1470
|
|
1471 #pragma acc data copy (fgot, fdata[0:N])
|
|
1472 {
|
|
1473 #pragma acc parallel loop
|
|
1474 for (i = 0; i < N; i++)
|
|
1475 {
|
|
1476 float expr = 1.0;
|
|
1477
|
|
1478 #pragma acc atomic capture
|
|
1479 { fgot = fgot - expr; fdata[i] = fgot; }
|
|
1480 }
|
|
1481 }
|
|
1482
|
|
1483 if (fexp != fgot)
|
|
1484 abort ();
|
|
1485
|
|
1486 fgot = 1.0;
|
|
1487 fexp = 1.0;
|
|
1488
|
|
1489 #pragma acc data copy (fgot, fdata[0:N])
|
|
1490 {
|
|
1491 #pragma acc parallel loop
|
|
1492 for (i = 0; i < N; i++)
|
|
1493 {
|
|
1494 float expr = 1.0;
|
|
1495
|
|
1496 #pragma acc atomic capture
|
|
1497 { fgot = expr - fgot; fdata[i] = fgot; }
|
|
1498 }
|
|
1499 }
|
|
1500
|
|
1501 for (i = 0; i < N; i++)
|
|
1502 if (i % 2 == 0)
|
|
1503 {
|
|
1504 if (fdata[i] != 0.0)
|
|
1505 abort ();
|
|
1506 }
|
|
1507 else
|
|
1508 {
|
|
1509 if (fdata[i] != 1.0)
|
|
1510 abort ();
|
|
1511 }
|
|
1512
|
|
1513 if (fexp != fgot)
|
|
1514 abort ();
|
|
1515
|
|
1516 /* BINOP = / */
|
|
1517 fgot = 8192.0*8192.0*64.0;
|
|
1518 fexp = 1.0;
|
|
1519
|
|
1520 #pragma acc data copy (fgot, fdata[0:N])
|
|
1521 {
|
|
1522 #pragma acc parallel loop
|
|
1523 for (i = 0; i < N; i++)
|
|
1524 {
|
|
1525 float expr = 2.0;
|
|
1526
|
|
1527 #pragma acc atomic capture
|
|
1528 { fdata[i] = fgot; fgot /= expr; }
|
|
1529 }
|
|
1530 }
|
|
1531
|
|
1532 if (fexp != fgot)
|
|
1533 abort ();
|
|
1534
|
|
1535 fgot = 8192.0*8192.0*64.0;
|
|
1536 fexp = 1.0;
|
|
1537
|
|
1538 #pragma acc data copy (fgot, fdata[0:N])
|
|
1539 {
|
|
1540 #pragma acc parallel loop
|
|
1541 for (i = 0; i < N; i++)
|
|
1542 {
|
|
1543 float expr = 2.0;
|
|
1544
|
|
1545 #pragma acc atomic capture
|
|
1546 { fgot /= expr; fdata[i] = fgot; }
|
|
1547 }
|
|
1548 }
|
|
1549
|
|
1550 if (fexp != fgot)
|
|
1551 abort ();
|
|
1552
|
|
1553 fgot = 8192.0*8192.0*64.0;
|
|
1554 fexp = 1.0;
|
|
1555
|
|
1556 #pragma acc data copy (fgot, fdata[0:N])
|
|
1557 {
|
|
1558 #pragma acc parallel loop
|
|
1559 for (i = 0; i < N; i++)
|
|
1560 {
|
|
1561 float expr = 2.0;
|
|
1562
|
|
1563 #pragma acc atomic capture
|
|
1564 { fdata[i] = fgot; fgot = fgot / expr; }
|
|
1565 }
|
|
1566 }
|
|
1567
|
|
1568 if (fexp != fgot)
|
|
1569 abort ();
|
|
1570
|
|
1571 fgot = 8192.0*8192.0*64.0;
|
|
1572 fexp = 1.0;
|
|
1573
|
|
1574 #pragma acc data copy (fgot, fdata[0:N])
|
|
1575 {
|
|
1576 #pragma acc parallel loop
|
|
1577 for (i = 0; i < N; i++)
|
|
1578 {
|
|
1579 float expr = 1.0;
|
|
1580
|
|
1581 #pragma acc atomic capture
|
|
1582 { fdata[i] = fgot; fgot = expr / fgot; }
|
|
1583 }
|
|
1584 }
|
|
1585
|
|
1586 if (fexp != fgot)
|
|
1587 abort ();
|
|
1588
|
|
1589 fgot = 4.0;
|
|
1590 fexp = 4.0;
|
|
1591
|
|
1592 #pragma acc data copy (fgot, fdata[0:N])
|
|
1593 {
|
|
1594 #pragma acc parallel loop
|
|
1595 for (i = 0; i < N; i++)
|
|
1596 {
|
|
1597 long long expr = 1LL << N;
|
|
1598
|
|
1599 #pragma acc atomic capture
|
|
1600 { fgot = fgot / expr; fdata[i] = fgot; }
|
|
1601 }
|
|
1602 }
|
|
1603
|
|
1604 if (fexp != fgot)
|
|
1605 abort ();
|
|
1606
|
|
1607 fgot = 4.0;
|
|
1608 fexp = 4.0;
|
|
1609
|
|
1610 #pragma acc data copy (fgot, fdata[0:N])
|
|
1611 {
|
|
1612 #pragma acc parallel loop
|
|
1613 for (i = 0; i < N; i++)
|
|
1614 {
|
|
1615 float expr = 2.0;
|
|
1616
|
|
1617 #pragma acc atomic capture
|
|
1618 { fgot = expr / fgot; fdata[i] = fgot; }
|
|
1619 }
|
|
1620 }
|
|
1621
|
|
1622 if (fexp != fgot)
|
|
1623 abort ();
|
|
1624
|
|
1625 return 0;
|
|
1626 }
|