1 | // RUN: %libomp-compile && env LIBOMP_USE_HIDDEN_HELPER_TASK=0 LIBOMP_NUM_HIDDEN_HELPER_THREADS=0 %libomp-run |
2 | /* |
3 | Test for the 'schedule(simd:guided)' clause. |
4 | Compiler needs to generate a dynamic dispatching and pass the schedule |
5 | value 46 to the OpenMP RTL. Test uses numerous loop parameter combinations. |
6 | */ |
7 | #include <stdio.h> |
8 | #include <stdlib.h> |
9 | #include <omp.h> |
10 | |
11 | #if defined(WIN32) || defined(_WIN32) |
12 | #include <windows.h> |
13 | #define delay() Sleep(1); |
14 | #else |
15 | #include <unistd.h> |
16 | #define delay() usleep(10); |
17 | #endif |
18 | |
19 | // uncomment for debug diagnostics: |
20 | //#define DEBUG |
21 | |
22 | #define SIMD_LEN 4 |
23 | |
24 | // --------------------------------------------------------------------------- |
25 | // Various definitions copied from OpenMP RTL |
26 | enum sched { |
27 | kmp_sch_static_balanced_chunked = 45, |
28 | kmp_sch_guided_simd = 46, |
29 | kmp_sch_runtime_simd = 47, |
30 | }; |
31 | typedef unsigned u32; |
32 | typedef long long i64; |
33 | typedef unsigned long long u64; |
34 | typedef struct { |
35 | int reserved_1; |
36 | int flags; |
37 | int reserved_2; |
38 | int reserved_3; |
39 | char *psource; |
40 | } id; |
41 | |
42 | extern int __kmpc_global_thread_num(id*); |
43 | extern void __kmpc_barrier(id*, int gtid); |
44 | extern void __kmpc_dispatch_init_4(id*, int, enum sched, int, int, int, int); |
45 | extern void __kmpc_dispatch_init_8(id*, int, enum sched, i64, i64, i64, i64); |
46 | extern int __kmpc_dispatch_next_4(id*, int, void*, void*, void*, void*); |
47 | extern int __kmpc_dispatch_next_8(id*, int, void*, void*, void*, void*); |
48 | // End of definitions copied from OpenMP RTL. |
49 | // --------------------------------------------------------------------------- |
50 | static id loc = {.reserved_1: 0, .flags: 2, .reserved_2: 0, .reserved_3: 0, .psource: ";file;func;0;0;;" }; |
51 | // This variable is defined in OpenMP RTL but we can't have it exposed so we |
52 | // need to redefine it here. |
53 | static int __kmp_hidden_helper_threads_num = 0; |
54 | |
55 | // --------------------------------------------------------------------------- |
56 | int run_loop_64(i64 loop_lb, i64 loop_ub, i64 loop_st, int loop_chunk) { |
57 | int err = 0; |
58 | static int volatile loop_sync = 0; |
59 | i64 lb; // Chunk lower bound |
60 | i64 ub; // Chunk upper bound |
61 | i64 st; // Chunk stride |
62 | int rc; |
63 | int tid = omp_get_thread_num(); |
64 | int gtid = tid; |
65 | if (gtid) { |
66 | gtid += __kmp_hidden_helper_threads_num; |
67 | } |
68 | int last; |
69 | #if DEBUG |
70 | printf("run_loop_<%d>(lb=%d, ub=%d, st=%d, ch=%d)\n" , |
71 | (int)sizeof(i64), gtid, tid, |
72 | (int)loop_lb, (int)loop_ub, (int)loop_st, loop_chunk); |
73 | #endif |
74 | // Don't test degenerate cases that should have been discovered by codegen |
75 | if (loop_st == 0) |
76 | return 0; |
77 | if (loop_st > 0 ? loop_lb > loop_ub : loop_lb < loop_ub) |
78 | return 0; |
79 | |
80 | __kmpc_dispatch_init_8(&loc, gtid, kmp_sch_guided_simd, |
81 | loop_lb, loop_ub, loop_st, loop_chunk); |
82 | if (tid == 0) { |
83 | // Let the master thread handle the chunks alone |
84 | int chunk; // No of current chunk |
85 | i64 next_lb; // Lower bound of the next chunk |
86 | i64 last_ub; // Upper bound of the last processed chunk |
87 | u64 cur; // Number of interations in current chunk |
88 | u64 max; // Max allowed iterations for current chunk |
89 | int undersized = 0; |
90 | |
91 | chunk = 0; |
92 | next_lb = loop_lb; |
93 | max = (loop_ub - loop_lb) / loop_st + 1; |
94 | // The first chunk can consume all iterations |
95 | while (__kmpc_dispatch_next_8(&loc, gtid, &last, &lb, &ub, &st)) { |
96 | ++ chunk; |
97 | #if DEBUG |
98 | printf("chunk=%d, lb=%d, ub=%d\n" , chunk, (int)lb, (int)ub); |
99 | #endif |
100 | // Check if previous chunk (it is not the final chunk) is undersized |
101 | if (undersized) { |
102 | printf(format: "Error with chunk %d\n" , chunk); |
103 | err++; |
104 | } |
105 | // Check lower and upper bounds |
106 | if (lb != next_lb) { |
107 | printf(format: "Error with lb %d, %d, ch %d\n" , (int)lb, (int)next_lb, chunk); |
108 | err++; |
109 | } |
110 | if (loop_st > 0) { |
111 | if (!(ub <= loop_ub)) { |
112 | printf(format: "Error with ub %d, %d, ch %d\n" , (int)ub, (int)loop_ub, chunk); |
113 | err++; |
114 | } |
115 | if (!(lb <= ub)) { |
116 | printf(format: "Error with bounds %d, %d, %d\n" , (int)lb, (int)ub, chunk); |
117 | err++; |
118 | } |
119 | } else { |
120 | if (!(ub >= loop_ub)) { |
121 | printf(format: "Error with ub %d, %d, %d\n" , (int)ub, (int)loop_ub, chunk); |
122 | err++; |
123 | } |
124 | if (!(lb >= ub)) { |
125 | printf(format: "Error with bounds %d, %d, %d\n" , (int)lb, (int)ub, chunk); |
126 | err++; |
127 | } |
128 | }; // if |
129 | // Stride should not change |
130 | if (!(st == loop_st)) { |
131 | printf(format: "Error with st %d, %d, ch %d\n" , (int)st, (int)loop_st, chunk); |
132 | err++; |
133 | } |
134 | cur = (ub - lb) / loop_st + 1; |
135 | // Guided scheduling uses FP computations, so current chunk may |
136 | // be a bit bigger (+1) than allowed maximum |
137 | if (!(cur <= max + 1)) { |
138 | printf(format: "Error with iter %llu, %llu\n" , cur, max); |
139 | err++; |
140 | } |
141 | // Update maximum for the next chunk |
142 | if (cur < max) |
143 | max = cur; |
144 | next_lb = ub + loop_st; |
145 | last_ub = ub; |
146 | undersized = (cur < loop_chunk); |
147 | }; // while |
148 | // Must have at least one chunk |
149 | if (!(chunk > 0)) { |
150 | printf(format: "Error with chunk %d\n" , chunk); |
151 | err++; |
152 | } |
153 | // Must have the right last iteration index |
154 | if (loop_st > 0) { |
155 | if (!(last_ub <= loop_ub)) { |
156 | printf(format: "Error with last1 %d, %d, ch %d\n" , |
157 | (int)last_ub, (int)loop_ub, chunk); |
158 | err++; |
159 | } |
160 | if (!(last_ub + loop_st > loop_ub)) { |
161 | printf(format: "Error with last2 %d, %d, %d, ch %d\n" , |
162 | (int)last_ub, (int)loop_st, (int)loop_ub, chunk); |
163 | err++; |
164 | } |
165 | } else { |
166 | if (!(last_ub >= loop_ub)) { |
167 | printf(format: "Error with last1 %d, %d, ch %d\n" , |
168 | (int)last_ub, (int)loop_ub, chunk); |
169 | err++; |
170 | } |
171 | if (!(last_ub + loop_st < loop_ub)) { |
172 | printf(format: "Error with last2 %d, %d, %d, ch %d\n" , |
173 | (int)last_ub, (int)loop_st, (int)loop_ub, chunk); |
174 | err++; |
175 | } |
176 | }; // if |
177 | // Let non-master threads go |
178 | loop_sync = 1; |
179 | } else { |
180 | int i; |
181 | // Workers wait for master thread to finish, then call __kmpc_dispatch_next |
182 | for (i = 0; i < 1000000; ++ i) { |
183 | if (loop_sync != 0) { |
184 | break; |
185 | }; // if |
186 | }; // for i |
187 | while (loop_sync == 0) { |
188 | delay(); |
189 | }; // while |
190 | // At this moment we do not have any more chunks -- all the chunks already |
191 | // processed by master thread |
192 | rc = __kmpc_dispatch_next_8(&loc, gtid, &last, &lb, &ub, &st); |
193 | if (rc) { |
194 | printf(format: "Error return value\n" ); |
195 | err++; |
196 | } |
197 | }; // if |
198 | |
199 | __kmpc_barrier(&loc, gtid); |
200 | if (tid == 0) { |
201 | loop_sync = 0; // Restore original state |
202 | #if DEBUG |
203 | printf("run_loop_64(): at the end\n" ); |
204 | #endif |
205 | }; // if |
206 | __kmpc_barrier(&loc, gtid); |
207 | return err; |
208 | } // run_loop |
209 | |
210 | // --------------------------------------------------------------------------- |
211 | int run_loop_32(int loop_lb, int loop_ub, int loop_st, int loop_chunk) { |
212 | int err = 0; |
213 | static int volatile loop_sync = 0; |
214 | int lb; // Chunk lower bound |
215 | int ub; // Chunk upper bound |
216 | int st; // Chunk stride |
217 | int rc; |
218 | int tid = omp_get_thread_num(); |
219 | int gtid = tid; |
220 | if (gtid) { |
221 | gtid += __kmp_hidden_helper_threads_num; |
222 | } |
223 | int last; |
224 | #if DEBUG |
225 | printf("run_loop_<%d>(lb=%d, ub=%d, st=%d, ch=%d)\n" , |
226 | (int)sizeof(int), gtid, tid, |
227 | (int)loop_lb, (int)loop_ub, (int)loop_st, loop_chunk); |
228 | #endif |
229 | // Don't test degenerate cases that should have been discovered by codegen |
230 | if (loop_st == 0) |
231 | return 0; |
232 | if (loop_st > 0 ? loop_lb > loop_ub : loop_lb < loop_ub) |
233 | return 0; |
234 | |
235 | __kmpc_dispatch_init_4(&loc, gtid, kmp_sch_guided_simd, |
236 | loop_lb, loop_ub, loop_st, loop_chunk); |
237 | if (tid == 0) { |
238 | // Let the master thread handle the chunks alone |
239 | int chunk; // No of current chunk |
240 | int next_lb; // Lower bound of the next chunk |
241 | int last_ub; // Upper bound of the last processed chunk |
242 | u64 cur; // Number of interations in current chunk |
243 | u64 max; // Max allowed iterations for current chunk |
244 | int undersized = 0; |
245 | |
246 | chunk = 0; |
247 | next_lb = loop_lb; |
248 | max = (loop_ub - loop_lb) / loop_st + 1; |
249 | // The first chunk can consume all iterations |
250 | while (__kmpc_dispatch_next_4(&loc, gtid, &last, &lb, &ub, &st)) { |
251 | ++ chunk; |
252 | #if DEBUG |
253 | printf("chunk=%d, lb=%d, ub=%d\n" , chunk, (int)lb, (int)ub); |
254 | #endif |
255 | // Check if previous chunk (it is not the final chunk) is undersized |
256 | if (undersized) { |
257 | printf(format: "Error with chunk %d\n" , chunk); |
258 | err++; |
259 | } |
260 | // Check lower and upper bounds |
261 | if (lb != next_lb) { |
262 | printf(format: "Error with lb %d, %d, ch %d\n" , (int)lb, (int)next_lb, chunk); |
263 | err++; |
264 | } |
265 | if (loop_st > 0) { |
266 | if (!(ub <= loop_ub)) { |
267 | printf(format: "Error with ub %d, %d, ch %d\n" , (int)ub, (int)loop_ub, chunk); |
268 | err++; |
269 | } |
270 | if (!(lb <= ub)) { |
271 | printf(format: "Error with bounds %d, %d, %d\n" , (int)lb, (int)ub, chunk); |
272 | err++; |
273 | } |
274 | } else { |
275 | if (!(ub >= loop_ub)) { |
276 | printf(format: "Error with ub %d, %d, %d\n" , (int)ub, (int)loop_ub, chunk); |
277 | err++; |
278 | } |
279 | if (!(lb >= ub)) { |
280 | printf(format: "Error with bounds %d, %d, %d\n" , (int)lb, (int)ub, chunk); |
281 | err++; |
282 | } |
283 | }; // if |
284 | // Stride should not change |
285 | if (!(st == loop_st)) { |
286 | printf(format: "Error with st %d, %d, ch %d\n" , (int)st, (int)loop_st, chunk); |
287 | err++; |
288 | } |
289 | cur = (ub - lb) / loop_st + 1; |
290 | // Guided scheduling uses FP computations, so current chunk may |
291 | // be a bit bigger (+1) than allowed maximum |
292 | if (!(cur <= max + 1)) { |
293 | printf(format: "Error with iter %llu, %llu\n" , cur, max); |
294 | err++; |
295 | } |
296 | // Update maximum for the next chunk |
297 | if (cur < max) |
298 | max = cur; |
299 | next_lb = ub + loop_st; |
300 | last_ub = ub; |
301 | undersized = (cur < loop_chunk); |
302 | }; // while |
303 | // Must have at least one chunk |
304 | if (!(chunk > 0)) { |
305 | printf(format: "Error with chunk %d\n" , chunk); |
306 | err++; |
307 | } |
308 | // Must have the right last iteration index |
309 | if (loop_st > 0) { |
310 | if (!(last_ub <= loop_ub)) { |
311 | printf(format: "Error with last1 %d, %d, ch %d\n" , |
312 | (int)last_ub, (int)loop_ub, chunk); |
313 | err++; |
314 | } |
315 | if (!(last_ub + loop_st > loop_ub)) { |
316 | printf(format: "Error with last2 %d, %d, %d, ch %d\n" , |
317 | (int)last_ub, (int)loop_st, (int)loop_ub, chunk); |
318 | err++; |
319 | } |
320 | } else { |
321 | if (!(last_ub >= loop_ub)) { |
322 | printf(format: "Error with last1 %d, %d, ch %d\n" , |
323 | (int)last_ub, (int)loop_ub, chunk); |
324 | err++; |
325 | } |
326 | if (!(last_ub + loop_st < loop_ub)) { |
327 | printf(format: "Error with last2 %d, %d, %d, ch %d\n" , |
328 | (int)last_ub, (int)loop_st, (int)loop_ub, chunk); |
329 | err++; |
330 | } |
331 | }; // if |
332 | // Let non-master threads go |
333 | loop_sync = 1; |
334 | } else { |
335 | int i; |
336 | // Workers wait for master thread to finish, then call __kmpc_dispatch_next |
337 | for (i = 0; i < 1000000; ++ i) { |
338 | if (loop_sync != 0) { |
339 | break; |
340 | }; // if |
341 | }; // for i |
342 | while (loop_sync == 0) { |
343 | delay(); |
344 | }; // while |
345 | // At this moment we do not have any more chunks -- all the chunks already |
346 | // processed by the master thread |
347 | rc = __kmpc_dispatch_next_4(&loc, gtid, &last, &lb, &ub, &st); |
348 | if (rc) { |
349 | printf(format: "Error return value\n" ); |
350 | err++; |
351 | } |
352 | }; // if |
353 | |
354 | __kmpc_barrier(&loc, gtid); |
355 | if (tid == 0) { |
356 | loop_sync = 0; // Restore original state |
357 | #if DEBUG |
358 | printf("run_loop<>(): at the end\n" ); |
359 | #endif |
360 | }; // if |
361 | __kmpc_barrier(&loc, gtid); |
362 | return err; |
363 | } // run_loop |
364 | |
365 | // --------------------------------------------------------------------------- |
366 | int run_64(int num_th) |
367 | { |
368 | int err = 0; |
369 | #pragma omp parallel num_threads(num_th) |
370 | { |
371 | int chunk; |
372 | i64 st, lb, ub; |
373 | for (chunk = SIMD_LEN; chunk <= 3*SIMD_LEN; chunk += SIMD_LEN) { |
374 | for (st = 1; st <= 3; ++ st) { |
375 | for (lb = -3 * num_th * st; lb <= 3 * num_th * st; ++ lb) { |
376 | for (ub = lb; ub < lb + num_th * (chunk+1) * st; ++ ub) { |
377 | err += run_loop_64(loop_lb: lb, loop_ub: ub, loop_st: st, loop_chunk: chunk); |
378 | err += run_loop_64(loop_lb: ub, loop_ub: lb, loop_st: -st, loop_chunk: chunk); |
379 | }; // for ub |
380 | }; // for lb |
381 | }; // for st |
382 | }; // for chunk |
383 | } |
384 | return err; |
385 | } // run_all |
386 | |
387 | int run_32(int num_th) |
388 | { |
389 | int err = 0; |
390 | #pragma omp parallel num_threads(num_th) |
391 | { |
392 | int chunk, st, lb, ub; |
393 | for (chunk = SIMD_LEN; chunk <= 3*SIMD_LEN; chunk += SIMD_LEN) { |
394 | for (st = 1; st <= 3; ++ st) { |
395 | for (lb = -3 * num_th * st; lb <= 3 * num_th * st; ++ lb) { |
396 | for (ub = lb; ub < lb + num_th * (chunk+1) * st; ++ ub) { |
397 | err += run_loop_32(loop_lb: lb, loop_ub: ub, loop_st: st, loop_chunk: chunk); |
398 | err += run_loop_32(loop_lb: ub, loop_ub: lb, loop_st: -st, loop_chunk: chunk); |
399 | }; // for ub |
400 | }; // for lb |
401 | }; // for st |
402 | }; // for chunk |
403 | } |
404 | return err; |
405 | } // run_all |
406 | |
407 | // --------------------------------------------------------------------------- |
408 | int main() |
409 | { |
410 | { |
411 | const char *env = getenv(name: "LIBOMP_NUM_HIDDEN_HELPER_THREADS" ); |
412 | if (env) { |
413 | __kmp_hidden_helper_threads_num = atoi(nptr: env); |
414 | } |
415 | } |
416 | |
417 | int n, err = 0; |
418 | for (n = 1; n <= 4; ++ n) { |
419 | err += run_32(num_th: n); |
420 | err += run_64(num_th: n); |
421 | }; // for n |
422 | if (err) |
423 | printf(format: "failed with %d errors\n" , err); |
424 | else |
425 | printf(format: "passed\n" ); |
426 | return err; |
427 | } |
428 | |