| 1 |
/* { dg-additional-options "-fopenacc-dim=16:16" } */ |
| 2 |
|
| 3 |
#include <openacc.h> |
| 4 |
#include <string.h> |
| 5 |
#include <stdio.h> |
| 6 |
#include <gomp-constants.h> |
| 7 |
|
| 8 |
#pragma acc routine |
| 9 |
static int __attribute__ ((noinline)) coord () |
| 10 |
{ |
| 11 |
int res = 0; |
| 12 |
|
| 13 |
if (acc_on_device (acc_device_not_host)) |
| 14 |
{ |
| 15 |
int g, w, v; |
| 16 |
|
| 17 |
g = __builtin_goacc_parlevel_id (GOMP_DIM_GANG); |
| 18 |
w = __builtin_goacc_parlevel_id (GOMP_DIM_WORKER); |
| 19 |
v = __builtin_goacc_parlevel_id (GOMP_DIM_VECTOR); |
| 20 |
res = (1 << 24) | (g << 16) | (w << 8) | v; |
| 21 |
} |
| 22 |
return res; |
| 23 |
} |
| 24 |
|
| 25 |
|
| 26 |
int check (const int *ary, int size, int gp, int wp, int vp) |
| 27 |
{ |
| 28 |
int exit = 0; |
| 29 |
int ix; |
| 30 |
int *gangs = (int *)__builtin_alloca (gp * sizeof (int)); |
| 31 |
int *workers = (int *)__builtin_alloca (wp * sizeof (int)); |
| 32 |
int *vectors = (int *)__builtin_alloca (vp * sizeof (int)); |
| 33 |
int offloaded = 0; |
| 34 |
|
| 35 |
memset (gangs, 0, gp * sizeof (int)); |
| 36 |
memset (workers, 0, wp * sizeof (int)); |
| 37 |
memset (vectors, 0, vp * sizeof (int)); |
| 38 |
|
| 39 |
for (ix = 0; ix < size; ix++) |
| 40 |
{ |
| 41 |
int g = (ary[ix] >> 16) & 0xff; |
| 42 |
int w = (ary[ix] >> 8) & 0xff; |
| 43 |
int v = (ary[ix] >> 0) & 0xff; |
| 44 |
|
| 45 |
if (g >= gp || w >= wp || v >= vp) |
| 46 |
{ |
| 47 |
printf ("unexpected cpu %#x used\n", ary[ix]); |
| 48 |
exit = 1; |
| 49 |
} |
| 50 |
else |
| 51 |
{ |
| 52 |
vectors[v]++; |
| 53 |
workers[w]++; |
| 54 |
gangs[g]++; |
| 55 |
} |
| 56 |
offloaded += ary[ix] >> 24; |
| 57 |
} |
| 58 |
|
| 59 |
if (!offloaded) |
| 60 |
return 0; |
| 61 |
|
| 62 |
if (offloaded != size) |
| 63 |
{ |
| 64 |
printf ("offloaded %d times, expected %d\n", offloaded, size); |
| 65 |
return 1; |
| 66 |
} |
| 67 |
|
| 68 |
for (ix = 0; ix < gp; ix++) |
| 69 |
if (gangs[ix] != gangs[0]) |
| 70 |
{ |
| 71 |
printf ("gang %d not used %d times\n", ix, gangs[0]); |
| 72 |
exit = 1; |
| 73 |
} |
| 74 |
|
| 75 |
for (ix = 0; ix < wp; ix++) |
| 76 |
if (workers[ix] != workers[0]) |
| 77 |
{ |
| 78 |
printf ("worker %d not used %d times\n", ix, workers[0]); |
| 79 |
exit = 1; |
| 80 |
} |
| 81 |
|
| 82 |
for (ix = 0; ix < vp; ix++) |
| 83 |
if (vectors[ix] != vectors[0]) |
| 84 |
{ |
| 85 |
printf ("vector %d not used %d times\n", ix, vectors[0]); |
| 86 |
exit = 1; |
| 87 |
} |
| 88 |
|
| 89 |
return exit; |
| 90 |
} |
| 91 |
|
| 92 |
#define N (32 *32*32) |
| 93 |
|
| 94 |
int test_1 (int gp, int wp, int vp) |
| 95 |
{ |
| 96 |
int ary[N]; |
| 97 |
int exit = 0; |
| 98 |
|
| 99 |
#pragma acc parallel copyout (ary) |
| 100 |
{ |
| 101 |
#pragma acc loop gang (static:1) |
| 102 |
for (int ix = 0; ix < N; ix++) |
| 103 |
ary[ix] = coord (); |
| 104 |
} |
| 105 |
|
| 106 |
exit |= check (ary, N, gp, 1, 1); |
| 107 |
|
| 108 |
#pragma acc parallel copyout (ary) |
| 109 |
{ |
| 110 |
#pragma acc loop worker |
| 111 |
for (int ix = 0; ix < N; ix++) |
| 112 |
ary[ix] = coord (); |
| 113 |
} |
| 114 |
|
| 115 |
exit |= check (ary, N, 1, wp, 1); |
| 116 |
|
| 117 |
#pragma acc parallel copyout (ary) |
| 118 |
{ |
| 119 |
#pragma acc loop vector |
| 120 |
for (int ix = 0; ix < N; ix++) |
| 121 |
ary[ix] = coord (); |
| 122 |
} |
| 123 |
|
| 124 |
exit |= check (ary, N, 1, 1, vp); |
| 125 |
|
| 126 |
return exit; |
| 127 |
} |
| 128 |
|
| 129 |
int main () |
| 130 |
{ |
| 131 |
#ifdef ACC_DEVICE_TYPE_radeon |
| 132 |
/* AMD GCN uses the autovectorizer for the vector dimension: the use |
| 133 |
of a function call in vector-partitioned code in this test is not |
| 134 |
currently supported. */ |
| 135 |
return test_1 (16, 16, 1); |
| 136 |
#else |
| 137 |
return test_1 (16, 16, 32); |
| 138 |
#endif |
| 139 |
} |