145
|
1 /* Test cases of nested reduction loops that should compile cleanly. */
|
|
2
|
|
3 void acc_parallel (void)
|
|
4 {
|
|
5 int i, j, k, sum, diff;
|
|
6
|
|
7 #pragma acc parallel
|
|
8 {
|
|
9 #pragma acc loop reduction(+:sum)
|
|
10 for (i = 0; i < 10; i++)
|
|
11 for (j = 0; j < 10; j++)
|
|
12 for (k = 0; k < 10; k++)
|
|
13 sum = 1;
|
|
14
|
|
15 #pragma acc loop collapse(2) reduction(+:sum)
|
|
16 for (i = 0; i < 10; i++)
|
|
17 for (j = 0; j < 10; j++)
|
|
18 for (k = 0; k < 10; k++)
|
|
19 sum = 1;
|
|
20
|
|
21 #pragma acc loop reduction(+:sum)
|
|
22 for (i = 0; i < 10; i++)
|
|
23 #pragma acc loop reduction(+:sum)
|
|
24 for (j = 0; j < 10; j++)
|
|
25 for (k = 0; k < 10; k++)
|
|
26 sum = 1;
|
|
27
|
|
28 #pragma acc loop reduction(+:sum)
|
|
29 for (i = 0; i < 10; i++)
|
|
30 #pragma acc loop collapse(2) reduction(+:sum)
|
|
31 for (j = 0; j < 10; j++)
|
|
32 for (k = 0; k < 10; k++)
|
|
33 sum = 1;
|
|
34
|
|
35 #pragma acc loop reduction(+:sum)
|
|
36 for (i = 0; i < 10; i++)
|
|
37 for (j = 0; j < 10; j++)
|
|
38 #pragma acc loop reduction(+:sum)
|
|
39 for (k = 0; k < 10; k++)
|
|
40 sum = 1;
|
|
41
|
|
42 #pragma acc loop reduction(+:sum)
|
|
43 for (i = 0; i < 10; i++)
|
|
44 #pragma acc loop reduction(+:sum)
|
|
45 for (j = 0; j < 10; j++)
|
|
46 #pragma acc loop reduction(+:sum)
|
|
47 for (k = 0; k < 10; k++)
|
|
48 sum = 1;
|
|
49
|
|
50 #pragma acc loop reduction(+:sum) reduction(-:diff)
|
|
51 for (i = 0; i < 10; i++)
|
|
52 {
|
|
53 #pragma acc loop reduction(+:sum)
|
|
54 for (j = 0; j < 10; j++)
|
|
55 #pragma acc loop reduction(+:sum)
|
|
56 for (k = 0; k < 10; k++)
|
|
57 sum = 1;
|
|
58
|
|
59 #pragma acc loop reduction(-:diff)
|
|
60 for (j = 0; j < 10; j++)
|
|
61 #pragma acc loop reduction(-:diff)
|
|
62 for (k = 0; k < 10; k++)
|
|
63 diff = 1;
|
|
64 }
|
|
65 }
|
|
66 }
|
|
67
|
|
68 /* The same tests as above, but using a combined parallel loop construct. */
|
|
69
|
|
70 void acc_parallel_loop (void)
|
|
71 {
|
|
72 int i, j, k, l, sum, diff;
|
|
73
|
|
74 #pragma acc parallel loop
|
|
75 for (int h = 0; h < 10; ++h)
|
|
76 {
|
|
77 #pragma acc loop reduction(+:sum)
|
|
78 for (i = 0; i < 10; i++)
|
|
79 for (j = 0; j < 10; j++)
|
|
80 for (k = 0; k < 10; k++)
|
|
81 sum = 1;
|
|
82
|
|
83 #pragma acc loop collapse(2) reduction(+:sum)
|
|
84 for (i = 0; i < 10; i++)
|
|
85 for (j = 0; j < 10; j++)
|
|
86 for (k = 0; k < 10; k++)
|
|
87 sum = 1;
|
|
88
|
|
89 #pragma acc loop reduction(+:sum)
|
|
90 for (i = 0; i < 10; i++)
|
|
91 #pragma acc loop reduction(+:sum)
|
|
92 for (j = 0; j < 10; j++)
|
|
93 for (k = 0; k < 10; k++)
|
|
94 sum = 1;
|
|
95
|
|
96 #pragma acc loop reduction(+:sum)
|
|
97 for (i = 0; i < 10; i++)
|
|
98 #pragma acc loop collapse(2) reduction(+:sum)
|
|
99 for (j = 0; j < 10; j++)
|
|
100 for (k = 0; k < 10; k++)
|
|
101 sum = 1;
|
|
102
|
|
103 #pragma acc loop reduction(+:sum)
|
|
104 for (i = 0; i < 10; i++)
|
|
105 for (j = 0; j < 10; j++)
|
|
106 #pragma acc loop reduction(+:sum)
|
|
107 for (k = 0; k < 10; k++)
|
|
108 sum = 1;
|
|
109
|
|
110 #pragma acc loop reduction(+:sum)
|
|
111 for (i = 0; i < 10; i++)
|
|
112 #pragma acc loop reduction(+:sum) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
113 for (j = 0; j < 10; j++)
|
|
114 #pragma acc loop reduction(+:sum)
|
|
115 for (k = 0; k < 10; k++)
|
|
116 sum = 1;
|
|
117
|
|
118 #pragma acc loop reduction(+:sum) reduction(-:diff)
|
|
119 for (i = 0; i < 10; i++)
|
|
120 {
|
|
121 #pragma acc loop reduction(+:sum) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
122 for (j = 0; j < 10; j++)
|
|
123 #pragma acc loop reduction(+:sum)
|
|
124 for (k = 0; k < 10; k++)
|
|
125 sum = 1;
|
|
126
|
|
127 #pragma acc loop reduction(-:diff) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
128 for (j = 0; j < 10; j++)
|
|
129 #pragma acc loop reduction(-:diff)
|
|
130 for (k = 0; k < 10; k++)
|
|
131 diff = 1;
|
|
132 }
|
|
133 }
|
|
134 }
|
|
135
|
|
136 /* The same tests as above, but now the outermost reduction clause is on
|
|
137 the parallel region, not the outermost loop. */
|
|
138
|
|
139 void acc_parallel_reduction (void)
|
|
140 {
|
|
141 int i, j, k, sum, diff;
|
|
142
|
|
143 #pragma acc parallel reduction(+:sum)
|
|
144 {
|
|
145 for (i = 0; i < 10; i++)
|
|
146 for (j = 0; j < 10; j++)
|
|
147 for (k = 0; k < 10; k++)
|
|
148 sum = 1;
|
|
149
|
|
150 for (i = 0; i < 10; i++)
|
|
151 #pragma acc loop
|
|
152 for (j = 0; j < 10; j++)
|
|
153 for (k = 0; k < 10; k++)
|
|
154 sum = 1;
|
|
155
|
|
156 #pragma acc loop reduction(+:sum)
|
|
157 for (i = 0; i < 10; i++)
|
|
158 for (j = 0; j < 10; j++)
|
|
159 #pragma acc loop reduction(+:sum)
|
|
160 for (k = 0; k < 10; k++)
|
|
161 sum = 1;
|
|
162
|
|
163 for (i = 0; i < 10; i++)
|
|
164 for (j = 0; j < 10; j++)
|
|
165 #pragma acc loop
|
|
166 for (k = 0; k < 10; k++)
|
|
167 sum = 1;
|
|
168
|
|
169 #pragma acc loop reduction(+:sum)
|
|
170 for (i = 0; i < 10; i++)
|
|
171 #pragma acc loop reduction(+:sum)
|
|
172 for (j = 0; j < 10; j++)
|
|
173 #pragma acc loop reduction(+:sum)
|
|
174 for (k = 0; k < 10; k++)
|
|
175 sum = 1;
|
|
176
|
|
177 #pragma acc loop reduction(+:sum) reduction(-:diff)
|
|
178 for (i = 0; i < 10; i++)
|
|
179 {
|
|
180 #pragma acc loop reduction(+:sum)
|
|
181 for (j = 0; j < 10; j++)
|
|
182 #pragma acc loop reduction(+:sum)
|
|
183 for (k = 0; k < 10; k++)
|
|
184 sum = 1;
|
|
185
|
|
186 #pragma acc loop reduction(-:diff)
|
|
187 for (j = 0; j < 10; j++)
|
|
188 #pragma acc loop reduction(-:diff)
|
|
189 for (k = 0; k < 10; k++)
|
|
190 diff = 1;
|
|
191 }
|
|
192
|
|
193 #pragma acc loop reduction(+:sum)
|
|
194 for (i = 0; i < 10; i++)
|
|
195 {
|
|
196 #pragma acc loop reduction(+:sum)
|
|
197 for (j = 0; j < 10; j++)
|
|
198 #pragma acc loop reduction(+:sum)
|
|
199 for (k = 0; k < 10; k++)
|
|
200 sum = 1;
|
|
201
|
|
202 #pragma acc loop reduction(-:diff)
|
|
203 for (j = 0; j < 10; j++)
|
|
204 #pragma acc loop reduction(-:diff)
|
|
205 for (k = 0; k < 10; k++)
|
|
206 diff = 1;
|
|
207 }
|
|
208
|
|
209 #pragma acc loop reduction(+:sum)
|
|
210 for (i = 0; i < 10; i++)
|
|
211 {
|
|
212 #pragma acc loop reduction(+:sum)
|
|
213 for (j = 0; j < 10; j++)
|
|
214 #pragma acc loop reduction(+:sum)
|
|
215 for (k = 0; k < 10; k++)
|
|
216 sum = 1;
|
|
217
|
|
218 #pragma acc loop
|
|
219 for (j = 0; j < 10; j++)
|
|
220 #pragma acc loop reduction(-:diff)
|
|
221 for (k = 0; k < 10; k++)
|
|
222 diff = 1;
|
|
223 }
|
|
224 }
|
|
225 }
|
|
226
|
|
227 /* The same tests as above, but using a combined parallel loop construct, and
|
|
228 the outermost reduction clause is on that one, not the outermost loop. */
|
|
229 void acc_parallel_loop_reduction (void)
|
|
230 {
|
|
231 int i, j, k, sum, diff;
|
|
232
|
|
233 #pragma acc parallel loop reduction(+:sum)
|
|
234 for (int h = 0; h < 10; ++h)
|
|
235 {
|
|
236 for (i = 0; i < 10; i++)
|
|
237 for (j = 0; j < 10; j++)
|
|
238 for (k = 0; k < 10; k++)
|
|
239 sum = 1;
|
|
240
|
|
241 for (i = 0; i < 10; i++)
|
|
242 #pragma acc loop
|
|
243 for (j = 0; j < 10; j++)
|
|
244 for (k = 0; k < 10; k++)
|
|
245 sum = 1;
|
|
246
|
|
247 #pragma acc loop reduction(+:sum)
|
|
248 for (i = 0; i < 10; i++)
|
|
249 for (j = 0; j < 10; j++)
|
|
250 #pragma acc loop reduction(+:sum)
|
|
251 for (k = 0; k < 10; k++)
|
|
252 sum = 1;
|
|
253
|
|
254 for (i = 0; i < 10; i++)
|
|
255 for (j = 0; j < 10; j++)
|
|
256 #pragma acc loop
|
|
257 for (k = 0; k < 10; k++)
|
|
258 sum = 1;
|
|
259
|
|
260 #pragma acc loop reduction(+:sum)
|
|
261 for (i = 0; i < 10; i++)
|
|
262 #pragma acc loop reduction(+:sum) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
263 for (j = 0; j < 10; j++)
|
|
264 #pragma acc loop reduction(+:sum)
|
|
265 for (k = 0; k < 10; k++)
|
|
266 sum = 1;
|
|
267
|
|
268 #pragma acc loop reduction(+:sum) reduction(-:diff)
|
|
269 for (i = 0; i < 10; i++)
|
|
270 {
|
|
271 #pragma acc loop reduction(+:sum) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
272 for (j = 0; j < 10; j++)
|
|
273 #pragma acc loop reduction(+:sum)
|
|
274 for (k = 0; k < 10; k++)
|
|
275 sum = 1;
|
|
276
|
|
277 #pragma acc loop reduction(-:diff) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
278 for (j = 0; j < 10; j++)
|
|
279 #pragma acc loop reduction(-:diff)
|
|
280 for (k = 0; k < 10; k++)
|
|
281 diff = 1;
|
|
282 }
|
|
283
|
|
284 #pragma acc loop reduction(+:sum)
|
|
285 for (i = 0; i < 10; i++)
|
|
286 {
|
|
287 #pragma acc loop reduction(+:sum) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
288 for (j = 0; j < 10; j++)
|
|
289 #pragma acc loop reduction(+:sum)
|
|
290 for (k = 0; k < 10; k++)
|
|
291 sum = 1;
|
|
292
|
|
293 #pragma acc loop reduction(-:diff) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
294 for (j = 0; j < 10; j++)
|
|
295 #pragma acc loop reduction(-:diff)
|
|
296 for (k = 0; k < 10; k++)
|
|
297 diff = 1;
|
|
298 }
|
|
299
|
|
300 #pragma acc loop reduction(+:sum)
|
|
301 for (i = 0; i < 10; i++)
|
|
302 {
|
|
303 #pragma acc loop reduction(+:sum) // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
304 for (j = 0; j < 10; j++)
|
|
305 #pragma acc loop reduction(+:sum)
|
|
306 for (k = 0; k < 10; k++)
|
|
307 sum = 1;
|
|
308
|
|
309 #pragma acc loop // { dg-warning "insufficient partitioning available to parallelize loop" }
|
|
310 for (j = 0; j < 10; j++)
|
|
311 #pragma acc loop reduction(-:diff)
|
|
312 for (k = 0; k < 10; k++)
|
|
313 diff = 1;
|
|
314 }
|
|
315 }
|
|
316 }
|
|
317
|
|
318 /* The same tests as above, but inside a routine construct. */
|
|
319 #pragma acc routine gang
|
|
320 void acc_routine (void)
|
|
321 {
|
|
322 int i, j, k, sum, diff;
|
|
323
|
|
324 {
|
|
325 #pragma acc loop reduction(+:sum)
|
|
326 for (i = 0; i < 10; i++)
|
|
327 for (j = 0; j < 10; j++)
|
|
328 for (k = 0; k < 10; k++)
|
|
329 sum = 1;
|
|
330
|
|
331 #pragma acc loop collapse(2) reduction(+:sum)
|
|
332 for (i = 0; i < 10; i++)
|
|
333 for (j = 0; j < 10; j++)
|
|
334 for (k = 0; k < 10; k++)
|
|
335 sum = 1;
|
|
336
|
|
337 #pragma acc loop reduction(+:sum)
|
|
338 for (i = 0; i < 10; i++)
|
|
339 #pragma acc loop reduction(+:sum)
|
|
340 for (j = 0; j < 10; j++)
|
|
341 for (k = 0; k < 10; k++)
|
|
342 sum = 1;
|
|
343
|
|
344 #pragma acc loop reduction(+:sum)
|
|
345 for (i = 0; i < 10; i++)
|
|
346 #pragma acc loop collapse(2) reduction(+:sum)
|
|
347 for (j = 0; j < 10; j++)
|
|
348 for (k = 0; k < 10; k++)
|
|
349 sum = 1;
|
|
350
|
|
351 #pragma acc loop reduction(+:sum)
|
|
352 for (i = 0; i < 10; i++)
|
|
353 for (j = 0; j < 10; j++)
|
|
354 #pragma acc loop reduction(+:sum)
|
|
355 for (k = 0; k < 10; k++)
|
|
356 sum = 1;
|
|
357
|
|
358 #pragma acc loop reduction(+:sum)
|
|
359 for (i = 0; i < 10; i++)
|
|
360 #pragma acc loop reduction(+:sum)
|
|
361 for (j = 0; j < 10; j++)
|
|
362 #pragma acc loop reduction(+:sum)
|
|
363 for (k = 0; k < 10; k++)
|
|
364 sum = 1;
|
|
365
|
|
366 #pragma acc loop reduction(+:sum) reduction(-:diff)
|
|
367 for (i = 0; i < 10; i++)
|
|
368 {
|
|
369 #pragma acc loop reduction(+:sum)
|
|
370 for (j = 0; j < 10; j++)
|
|
371 #pragma acc loop reduction(+:sum)
|
|
372 for (k = 0; k < 10; k++)
|
|
373 sum = 1;
|
|
374
|
|
375 #pragma acc loop reduction(-:diff)
|
|
376 for (j = 0; j < 10; j++)
|
|
377 #pragma acc loop reduction(-:diff)
|
|
378 for (k = 0; k < 10; k++)
|
|
379 diff = 1;
|
|
380 }
|
|
381 }
|
|
382 }
|
|
383
|
|
384 void acc_kernels (void)
|
|
385 {
|
|
386 int i, j, k, sum, diff;
|
|
387
|
|
388 /* FIXME: These tests are not meaningful yet because reductions in
|
|
389 kernels regions are not supported yet. */
|
|
390 #pragma acc kernels
|
|
391 {
|
|
392 #pragma acc loop reduction(+:sum)
|
|
393 for (i = 0; i < 10; i++)
|
|
394 for (j = 0; j < 10; j++)
|
|
395 for (k = 0; k < 10; k++)
|
|
396 sum = 1;
|
|
397
|
|
398 #pragma acc loop reduction(+:sum)
|
|
399 for (i = 0; i < 10; i++)
|
|
400 #pragma acc loop reduction(+:sum)
|
|
401 for (j = 0; j < 10; j++)
|
|
402 for (k = 0; k < 10; k++)
|
|
403 sum = 1;
|
|
404
|
|
405 #pragma acc loop reduction(+:sum)
|
|
406 for (i = 0; i < 10; i++)
|
|
407 for (j = 0; j < 10; j++)
|
|
408 #pragma acc loop reduction(+:sum)
|
|
409 for (k = 0; k < 10; k++)
|
|
410 sum = 1;
|
|
411
|
|
412 #pragma acc loop reduction(+:sum)
|
|
413 for (i = 0; i < 10; i++)
|
|
414 #pragma acc loop reduction(+:sum)
|
|
415 for (j = 0; j < 10; j++)
|
|
416 #pragma acc loop reduction(+:sum)
|
|
417 for (k = 0; k < 10; k++)
|
|
418 sum = 1;
|
|
419 }
|
|
420 }
|