1 | // RUN: %libomp-compile-and-run |
2 | |
3 | /* Test for guided scheduling |
4 | * Ensure threads get chunks interleavely first |
5 | * Then judge the chunk sizes are decreasing to a stable value |
6 | * Modified by Chunhua Liao |
7 | * For example, 100 iteration on 2 threads, chunksize 7 |
8 | * one line for each dispatch, 0/1 means thread id |
9 | * 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 24 |
10 | * 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 18 |
11 | * 0 0 0 0 0 0 0 0 0 0 0 0 0 0 14 |
12 | * 1 1 1 1 1 1 1 1 1 1 10 |
13 | * 0 0 0 0 0 0 0 0 8 |
14 | * 1 1 1 1 1 1 1 7 |
15 | * 0 0 0 0 0 0 0 7 |
16 | * 1 1 1 1 1 1 1 7 |
17 | * 0 0 0 0 0 5 |
18 | */ |
19 | #include <stdio.h> |
20 | #include <stdlib.h> |
21 | #include "omp_testsuite.h" |
22 | #include "omp_my_sleep.h" |
23 | |
24 | #define CFSMAX_SIZE 1000 |
25 | #define MAX_TIME 0.005 |
26 | |
27 | #ifdef SLEEPTIME |
28 | #undef SLEEPTIME |
29 | #define SLEEPTIME 0.0001 |
30 | #endif |
31 | |
32 | int test_omp_for_schedule_guided() |
33 | { |
34 | int * tids; |
35 | int * chunksizes; |
36 | int notout; |
37 | int maxiter; |
38 | int threads; |
39 | int i; |
40 | int result; |
41 | |
42 | tids = (int *) malloc (size: sizeof (int) * (CFSMAX_SIZE + 1)); |
43 | maxiter = 0; |
44 | result = 1; |
45 | notout = 1; |
46 | |
47 | /* Testing if enough threads are available for this check. */ |
48 | #pragma omp parallel |
49 | { |
50 | #pragma omp single |
51 | { |
52 | threads = omp_get_num_threads(); |
53 | } |
54 | } |
55 | |
56 | /* ensure there are at least two threads */ |
57 | if (threads < 2) { |
58 | omp_set_num_threads(2); |
59 | threads = 2; |
60 | } |
61 | |
62 | /* Now the real parallel work: |
63 | * Each thread will start immediately with the first chunk. |
64 | */ |
65 | #pragma omp parallel shared(tids,maxiter) |
66 | { /* begin of parallel */ |
67 | double count; |
68 | int tid; |
69 | int j; |
70 | |
71 | tid = omp_get_thread_num (); |
72 | |
73 | #pragma omp for nowait schedule(guided) |
74 | for(j = 0; j < CFSMAX_SIZE; ++j) { |
75 | count = 0.; |
76 | #pragma omp flush(maxiter) |
77 | if (j > maxiter) { |
78 | #pragma omp critical |
79 | { |
80 | maxiter = j; |
81 | } |
82 | } |
83 | /*printf ("thread %d sleeping\n", tid);*/ |
84 | #pragma omp flush(maxiter,notout) |
85 | while (notout && (count < MAX_TIME) && (maxiter == j)) { |
86 | #pragma omp flush(maxiter,notout) |
87 | my_sleep (SLEEPTIME); |
88 | count += SLEEPTIME; |
89 | #ifdef VERBOSE |
90 | printf("." ); |
91 | #endif |
92 | } |
93 | #ifdef VERBOSE |
94 | if (count > 0.) printf(" waited %lf s\n" , count); |
95 | #endif |
96 | /*printf ("thread %d awake\n", tid);*/ |
97 | tids[j] = tid; |
98 | #ifdef VERBOSE |
99 | printf("%d finished by %d\n" ,j,tid); |
100 | #endif |
101 | } /* end of for */ |
102 | notout = 0; |
103 | #pragma omp flush(maxiter,notout) |
104 | } /* end of parallel */ |
105 | |
106 | /******************************************************* |
107 | * evaluation of the values * |
108 | *******************************************************/ |
109 | { |
110 | int determined_chunksize = 1; |
111 | int last_threadnr = tids[0]; |
112 | int global_chunknr = 0; |
113 | int openwork = CFSMAX_SIZE; |
114 | int expected_chunk_size; |
115 | int* local_chunknr = (int*)malloc(size: threads * sizeof(int)); |
116 | double c = 1; |
117 | |
118 | for (i = 0; i < threads; i++) |
119 | local_chunknr[i] = 0; |
120 | |
121 | tids[CFSMAX_SIZE] = -1; |
122 | |
123 | /* |
124 | * determine the number of global chunks |
125 | */ |
126 | // fprintf(stderr,"# global_chunknr thread local_chunknr chunksize\n"); |
127 | for(i = 1; i <= CFSMAX_SIZE; ++i) { |
128 | if (last_threadnr==tids[i]) { |
129 | determined_chunksize++; |
130 | } else { |
131 | /* fprintf(stderr, "%d\t%d\t%d\t%d\n", global_chunknr, |
132 | last_threadnr, local_chunknr[last_threadnr], m); */ |
133 | global_chunknr++; |
134 | local_chunknr[last_threadnr]++; |
135 | last_threadnr = tids[i]; |
136 | determined_chunksize = 1; |
137 | } |
138 | } |
139 | /* now allocate the memory for saving the sizes of the global chunks */ |
140 | chunksizes = (int*)malloc(size: global_chunknr * sizeof(int)); |
141 | |
142 | /* |
143 | * Evaluate the sizes of the global chunks |
144 | */ |
145 | global_chunknr = 0; |
146 | determined_chunksize = 1; |
147 | last_threadnr = tids[0]; |
148 | for (i = 1; i <= CFSMAX_SIZE; ++i) { |
149 | /* If the threadnumber was the same as before increase the |
150 | * detected chunksize for this chunk otherwise set the detected |
151 | * chunksize again to one and save the number of the next |
152 | * thread in last_threadnr. |
153 | */ |
154 | if (last_threadnr == tids[i]) { |
155 | determined_chunksize++; |
156 | } else { |
157 | chunksizes[global_chunknr] = determined_chunksize; |
158 | global_chunknr++; |
159 | local_chunknr[last_threadnr]++; |
160 | last_threadnr = tids[i]; |
161 | determined_chunksize = 1; |
162 | } |
163 | } |
164 | |
165 | #ifdef VERBOSE |
166 | fprintf(stderr, "found\texpected\tconstant\n" ); |
167 | #endif |
168 | |
169 | /* identify the constant c for the exponential |
170 | decrease of the chunksize */ |
171 | expected_chunk_size = openwork / threads; |
172 | c = (double) chunksizes[0] / expected_chunk_size; |
173 | |
174 | for (i = 0; i < global_chunknr; i++) { |
175 | /* calculate the new expected chunksize */ |
176 | if (expected_chunk_size > 1) |
177 | expected_chunk_size = c * openwork / threads; |
178 | #ifdef VERBOSE |
179 | fprintf(stderr, "%8d\t%8d\t%lf\n" , chunksizes[i], |
180 | expected_chunk_size, c * chunksizes[i]/expected_chunk_size); |
181 | #endif |
182 | /* check if chunksize is inside the rounding errors */ |
183 | if (abs (x: chunksizes[i] - expected_chunk_size) >= 2) { |
184 | result = 0; |
185 | #ifndef VERBOSE |
186 | fprintf(stderr, format: "Chunksize differed from expected " |
187 | "value: %d instead of %d\n" , chunksizes[i], |
188 | expected_chunk_size); |
189 | return 0; |
190 | #endif |
191 | } /* end if */ |
192 | |
193 | #ifndef VERBOSE |
194 | if (expected_chunk_size - chunksizes[i] < 0) |
195 | fprintf(stderr, format: "Chunksize did not decrease: %d" |
196 | " instead of %d\n" , chunksizes[i],expected_chunk_size); |
197 | #endif |
198 | |
199 | /* calculating the remaining amount of work */ |
200 | openwork -= chunksizes[i]; |
201 | } |
202 | } |
203 | return result; |
204 | } |
205 | |
206 | int main() |
207 | { |
208 | int i; |
209 | int num_failed=0; |
210 | |
211 | for(i = 0; i < REPETITIONS; i++) { |
212 | if(!test_omp_for_schedule_guided()) { |
213 | num_failed++; |
214 | } |
215 | } |
216 | return num_failed; |
217 | } |
218 | |