1 | // RUN: %libomp-compile-and-run |
2 | // RUN: %libomp-compile && env KMP_TASKLOOP_MIN_TASKS=1 %libomp-run |
3 | |
4 | // These compilers don't support the taskloop construct |
5 | // UNSUPPORTED: gcc-4, gcc-5, icc-16 |
6 | // GCC 6 has support for taskloops, but at least 6.3.0 is crashing on this test |
7 | // UNSUPPORTED: gcc-6 |
8 | |
9 | /* |
10 | * Test for taskloop |
11 | * Method: calculate how many times the iteration space is dispatched |
12 | * and judge if each dispatch has the requested grainsize |
13 | * It is possible for two adjacent chunks are executed by the same thread |
14 | */ |
15 | #include <stdio.h> |
16 | #include <omp.h> |
17 | #include <stdlib.h> |
18 | #include "omp_testsuite.h" |
19 | |
20 | #define CFDMAX_SIZE 1120 |
21 | |
22 | int test_omp_taskloop_grainsize() |
23 | { |
24 | int result = 0; |
25 | int i, grainsize, count, tmp_count, num_off; |
26 | int *tmp, *tids, *tidsArray; |
27 | |
28 | tidsArray = (int *)malloc(size: sizeof(int) * CFDMAX_SIZE); |
29 | tids = tidsArray; |
30 | |
31 | for (grainsize = 1; grainsize < 48; ++grainsize) { |
32 | fprintf(stderr, format: "Grainsize %d\n" , grainsize); |
33 | count = tmp_count = num_off = 0; |
34 | |
35 | for (i = 0; i < CFDMAX_SIZE; ++i) { |
36 | tids[i] = -1; |
37 | } |
38 | |
39 | #pragma omp parallel shared(tids) |
40 | { |
41 | #pragma omp master |
42 | #pragma omp taskloop grainsize(grainsize) |
43 | for (i = 0; i < CFDMAX_SIZE; i++) { |
44 | tids[i] = omp_get_thread_num(); |
45 | } |
46 | } |
47 | |
48 | for (i = 0; i < CFDMAX_SIZE; ++i) { |
49 | if (tids[i] == -1) { |
50 | fprintf(stderr, format: " Iteration %d not touched!\n" , i); |
51 | result++; |
52 | } |
53 | } |
54 | |
55 | for (i = 0; i < CFDMAX_SIZE - 1; ++i) { |
56 | if (tids[i] != tids[i + 1]) { |
57 | count++; |
58 | } |
59 | } |
60 | |
61 | tmp = (int *)malloc(size: sizeof(int) * (count + 1)); |
62 | tmp[0] = 1; |
63 | |
64 | for (i = 0; i < CFDMAX_SIZE - 1; ++i) { |
65 | if (tmp_count > count) { |
66 | printf(format: "--------------------\nTestinternal Error: List too " |
67 | "small!!!\n--------------------\n" ); |
68 | break; |
69 | } |
70 | if (tids[i] != tids[i + 1]) { |
71 | tmp_count++; |
72 | tmp[tmp_count] = 1; |
73 | } else { |
74 | tmp[tmp_count]++; |
75 | } |
76 | } |
77 | |
78 | // is grainsize statement working? |
79 | int num_tasks = CFDMAX_SIZE / grainsize; |
80 | int multiple1 = CFDMAX_SIZE / num_tasks; |
81 | int multiple2 = CFDMAX_SIZE / num_tasks + 1; |
82 | for (i = 0; i < count; i++) { |
83 | // it is possible for 2 adjacent chunks assigned to a same thread |
84 | if (tmp[i] % multiple1 != 0 && tmp[i] % multiple2 != 0) { |
85 | num_off++; |
86 | } |
87 | } |
88 | |
89 | if (num_off > 1) { |
90 | fprintf(stderr, format: " The number of bad chunks is %d\n" , num_off); |
91 | result++; |
92 | } else { |
93 | fprintf(stderr, format: " Everything ok\n" ); |
94 | } |
95 | |
96 | free(ptr: tmp); |
97 | } |
98 | free(ptr: tidsArray); |
99 | return (result==0); |
100 | } |
101 | |
102 | int main() |
103 | { |
104 | int i; |
105 | int num_failed=0; |
106 | |
107 | for (i = 0; i < REPETITIONS; i++) { |
108 | if (!test_omp_taskloop_grainsize()) { |
109 | num_failed++; |
110 | } |
111 | } |
112 | return num_failed; |
113 | } |
114 | |