| 1 |
/* { dg-do run { target aarch64_sve256_hw } } */ |
| 2 |
/* { dg-options "-march=armv8-a+sve -msve-vector-bits=256 -fopenmp -O2" } */ |
| 3 |
|
| 4 |
#include <arm_sve.h> |
| 5 |
|
| 6 |
#pragma omp declare reduction (+:svint32_t: omp_out = svadd_s32_z (svptrue_b32(), omp_in, omp_out)) \ |
| 7 |
initializer (omp_priv = svindex_s32 (0, 0)) |
| 8 |
|
| 9 |
void __attribute__ ((noipa)) |
| 10 |
parallel_reduction () |
| 11 |
{ |
| 12 |
int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; |
| 13 |
int b[8] = {0, 0, 0, 0, 0, 0, 0, 0}; |
| 14 |
svint32_t va = svld1_s32 (svptrue_b32 (), b); |
| 15 |
int i = 0; |
| 16 |
int64_t res; |
| 17 |
|
| 18 |
#pragma omp parallel reduction (+:va, i) |
| 19 |
{ |
| 20 |
va = svld1_s32 (svptrue_b32 (), a); |
| 21 |
i++; |
| 22 |
} |
| 23 |
|
| 24 |
res = svaddv_s32 (svptrue_b32 (), va); |
| 25 |
|
| 26 |
if (res != i * 8) |
| 27 |
__builtin_abort (); |
| 28 |
} |
| 29 |
|
| 30 |
void __attribute__ ((noipa)) |
| 31 |
for_reduction () |
| 32 |
{ |
| 33 |
int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; |
| 34 |
int b[8] = {0, 0, 0, 0, 0, 0, 0, 0}; |
| 35 |
svint32_t va = svld1_s32 (svptrue_b32 (), b); |
| 36 |
int j; |
| 37 |
int64_t res; |
| 38 |
|
| 39 |
#pragma omp parallel for reduction (+:va) |
| 40 |
for (j = 0; j < 8; j++) |
| 41 |
va += svld1_s32 (svptrue_b32 (), a); |
| 42 |
|
| 43 |
res = svaddv_s32 (svptrue_b32 (), va); |
| 44 |
|
| 45 |
if (res != 64) |
| 46 |
__builtin_abort (); |
| 47 |
} |
| 48 |
|
| 49 |
void __attribute__ ((noipa)) |
| 50 |
simd_reduction () |
| 51 |
{ |
| 52 |
int a[8]; |
| 53 |
svint32_t va = svindex_s32 (0, 0); |
| 54 |
int i = 0; |
| 55 |
int j; |
| 56 |
int64_t res = 0; |
| 57 |
|
| 58 |
for (j = 0; j < 8; j++) |
| 59 |
a[j] = 1; |
| 60 |
|
| 61 |
#pragma omp simd reduction (+:va) |
| 62 |
for (j = 0; j < 16; j++) |
| 63 |
va += svld1_s32 (svptrue_b32 (), a); |
| 64 |
|
| 65 |
res = svaddv_s32 (svptrue_b32 (), va); |
| 66 |
|
| 67 |
if (res != 128) |
| 68 |
__builtin_abort (); |
| 69 |
} |
| 70 |
|
| 71 |
void __attribute__ ((noipa)) |
| 72 |
inscan_reduction_incl () |
| 73 |
{ |
| 74 |
svint32_t va = svindex_s32 (0, 0); |
| 75 |
int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; |
| 76 |
int b[64] = { 0 }; |
| 77 |
int j; |
| 78 |
int64_t res = 0; |
| 79 |
|
| 80 |
#pragma omp parallel for reduction (inscan, +:va) |
| 81 |
for (j = 0; j < 8; j++) |
| 82 |
{ |
| 83 |
va += svld1_s32 (svptrue_b32 (), a); |
| 84 |
#pragma omp scan inclusive (va) |
| 85 |
svst1_s32 (svptrue_b32 (), b + j * 8, va); |
| 86 |
} |
| 87 |
|
| 88 |
res = svaddv_s32 (svptrue_b32 (), va); |
| 89 |
|
| 90 |
if (res != 64) |
| 91 |
__builtin_abort (); |
| 92 |
|
| 93 |
for (j = 0; j < 64; j+=8) |
| 94 |
if (b[j] != (j / 8 + 1)) |
| 95 |
__builtin_abort (); |
| 96 |
} |
| 97 |
|
| 98 |
void __attribute__ ((noipa)) |
| 99 |
inscan_reduction_excl () |
| 100 |
{ |
| 101 |
svint32_t va = svindex_s32 (0, 0); |
| 102 |
int a[8] = {1, 1, 1, 1, 1, 1, 1, 1}; |
| 103 |
int b[64] = { 0 }; |
| 104 |
int j; |
| 105 |
int64_t res = 0; |
| 106 |
|
| 107 |
#pragma omp parallel for reduction (inscan, +:va) |
| 108 |
for (j = 0; j < 8; j++) |
| 109 |
{ |
| 110 |
svst1_s32 (svptrue_b32 (), b + j * 8, va); |
| 111 |
#pragma omp scan exclusive (va) |
| 112 |
va += svld1_s32 (svptrue_b32 (), a); |
| 113 |
} |
| 114 |
|
| 115 |
res = svaddv_s32 (svptrue_b32 (), va); |
| 116 |
|
| 117 |
if (res != 64) |
| 118 |
__builtin_abort (); |
| 119 |
|
| 120 |
for (j = 0; j < 64; j+=8) |
| 121 |
if (b[j] != j / 8) |
| 122 |
__builtin_abort (); |
| 123 |
} |
| 124 |
|
| 125 |
|
| 126 |
int |
| 127 |
main () |
| 128 |
{ |
| 129 |
parallel_reduction (); |
| 130 |
for_reduction (); |
| 131 |
simd_reduction (); |
| 132 |
inscan_reduction_incl (); |
| 133 |
inscan_reduction_excl (); |
| 134 |
} |