| 1 |
/* { dg-do run { target aarch64_sve256_hw } } */ |
| 2 |
/* { dg-options "-msve-vector-bits=256 -fopenmp -O2" } */ |
| 3 |
|
| 4 |
#pragma GCC target "+sve" |
| 5 |
|
| 6 |
#include <arm_sve.h> |
| 7 |
#include <omp.h> |
| 8 |
|
| 9 |
static void __attribute__ ((noipa)) |
| 10 |
vec_compare (svint32_t *x, svint32_t y) |
| 11 |
{ |
| 12 |
svbool_t p = svnot_b_z (svptrue_b32 (), svcmpeq_s32 (svptrue_b32 (), *x, y)); |
| 13 |
|
| 14 |
if (svptest_any (svptrue_b32 (), p)) |
| 15 |
__builtin_abort (); |
| 16 |
} |
| 17 |
|
| 18 |
void __attribute__ ((noipa)) |
| 19 |
firstprivate_sections () |
| 20 |
{ |
| 21 |
int b[8], c[8]; |
| 22 |
svint32_t vb, vc; |
| 23 |
int i; |
| 24 |
|
| 25 |
#pragma omp parallel for |
| 26 |
for (i = 0; i < 8; i++) |
| 27 |
{ |
| 28 |
b[i] = i; |
| 29 |
c[i] = i + 1; |
| 30 |
} |
| 31 |
|
| 32 |
vb = svld1_s32 (svptrue_b32 (), b); |
| 33 |
vc = svld1_s32 (svptrue_b32 (), c); |
| 34 |
|
| 35 |
#pragma omp parallel sections firstprivate (vb, vc) |
| 36 |
{ |
| 37 |
#pragma omp section |
| 38 |
vec_compare (&vb, svindex_s32 (0, 1)); |
| 39 |
vec_compare (&vc, svindex_s32 (1, 1)); |
| 40 |
|
| 41 |
#pragma omp section |
| 42 |
vec_compare (&vb, svindex_s32 (0, 1)); |
| 43 |
vec_compare (&vc, svindex_s32 (1, 1)); |
| 44 |
} |
| 45 |
|
| 46 |
} |
| 47 |
|
| 48 |
void __attribute__ ((noipa)) |
| 49 |
firstprivate_for () |
| 50 |
{ |
| 51 |
|
| 52 |
int a[32], b[32], c[32]; |
| 53 |
svint32_t va, vb, vc; |
| 54 |
int i; |
| 55 |
|
| 56 |
#pragma omp parallel for |
| 57 |
for (i = 0; i < 32; i++) |
| 58 |
{ |
| 59 |
b[i] = i; |
| 60 |
c[i] = i + 1; |
| 61 |
} |
| 62 |
|
| 63 |
vb = svindex_s32 (1, 0); |
| 64 |
vc = svindex_s32 (0, 1); |
| 65 |
|
| 66 |
#pragma omp parallel for firstprivate (vb, vc) private (va) |
| 67 |
for (i = 0; i < 4; i++) |
| 68 |
{ |
| 69 |
svint32_t tb, tc; |
| 70 |
vec_compare (&vb, svindex_s32 (1, 0)); |
| 71 |
vec_compare (&vc, svindex_s32 (0, 1)); |
| 72 |
tb = svld1_s32 (svptrue_b32 (), b + i * 8); |
| 73 |
tc = svld1_s32 (svptrue_b32 (), c + i * 8); |
| 74 |
va = svadd_s32_z (svptrue_b32 (), vb, vc); |
| 75 |
va = svadd_s32_z (svptrue_b32 (), va, tb); |
| 76 |
va = svadd_s32_z (svptrue_b32 (), va, tc); |
| 77 |
svst1_s32 (svptrue_b32 (), a + i * 8, va); |
| 78 |
} |
| 79 |
|
| 80 |
for (i = 0; i < 32; i++) |
| 81 |
if (a[i] != b[i] + c[i] + vb[i % 8] + vc[i % 8]) |
| 82 |
__builtin_abort (); |
| 83 |
} |
| 84 |
|
| 85 |
void __attribute__ ((noipa)) |
| 86 |
firstprivate_distribute () |
| 87 |
{ |
| 88 |
|
| 89 |
int a[32], b[32], c[32]; |
| 90 |
svint32_t va, vb, vc; |
| 91 |
int i; |
| 92 |
|
| 93 |
#pragma omp parallel for |
| 94 |
for (i = 0; i < 32; i++) |
| 95 |
{ |
| 96 |
b[i] = i; |
| 97 |
c[i] = i + 1; |
| 98 |
} |
| 99 |
|
| 100 |
vb = svindex_s32 (1, 0); |
| 101 |
vc = svindex_s32 (0, 1); |
| 102 |
|
| 103 |
#pragma omp teams |
| 104 |
#pragma omp distribute firstprivate (vb, vc) private (va) |
| 105 |
for (i = 0; i < 4; i++) |
| 106 |
{ |
| 107 |
svint32_t tb, tc; |
| 108 |
vec_compare (&vb, svindex_s32 (1, 0)); |
| 109 |
vec_compare (&vc, svindex_s32 (0, 1)); |
| 110 |
tb = svld1_s32 (svptrue_b32 (), b + i * 8); |
| 111 |
tc = svld1_s32 (svptrue_b32 (), c + i * 8); |
| 112 |
va = svadd_s32_z (svptrue_b32 (), vb, vc); |
| 113 |
va = svadd_s32_z (svptrue_b32 (), va, tb); |
| 114 |
va = svadd_s32_z (svptrue_b32 (), va, tc); |
| 115 |
svst1_s32 (svptrue_b32 (), a + i * 8, va); |
| 116 |
} |
| 117 |
|
| 118 |
for (i = 0; i < 32; i++) |
| 119 |
if (a[i] != b[i] + c[i] + vb[i % 8] + vc[i % 8]) |
| 120 |
__builtin_abort (); |
| 121 |
} |
| 122 |
|
| 123 |
int |
| 124 |
main () |
| 125 |
{ |
| 126 |
firstprivate_for (); |
| 127 |
firstprivate_sections (); |
| 128 |
firstprivate_distribute (); |
| 129 |
} |