| 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 <stdlib.h> |
| 8 |
#include <omp.h> |
| 9 |
|
| 10 |
static void __attribute__ ((noipa)) |
| 11 |
compare_vec (svint32_t x, svint32_t y) |
| 12 |
{ |
| 13 |
svbool_t p = svnot_b_z (svptrue_b32 (), svcmpeq_s32 (svptrue_b32 (), x, y)); |
| 14 |
|
| 15 |
if (svptest_any (svptrue_b32 (), p)) |
| 16 |
__builtin_abort (); |
| 17 |
} |
| 18 |
|
| 19 |
static void __attribute__ ((noipa)) |
| 20 |
compare_vecb (svbool_t x, svbool_t y) |
| 21 |
{ |
| 22 |
svbool_t p = sveor_b_z (svptrue_b32 (), x, y); |
| 23 |
|
| 24 |
if (svptest_any (svptrue_b32 (), p)) |
| 25 |
__builtin_abort (); |
| 26 |
} |
| 27 |
|
| 28 |
void __attribute__ ((noipa)) |
| 29 |
implicit_shared_default (svint32_t a, svint32_t b, svbool_t p) |
| 30 |
{ |
| 31 |
|
| 32 |
#pragma omp parallel default (shared) num_threads (10) |
| 33 |
{ |
| 34 |
/* 'a', 'b' and 'p' are implicitly shared. */ |
| 35 |
compare_vec (a, svindex_s32 (0, 1)); |
| 36 |
compare_vec (b, svindex_s32 (8, 1)); |
| 37 |
compare_vecb (p, svptrue_b32 ()); |
| 38 |
|
| 39 |
#pragma omp barrier |
| 40 |
if (omp_get_thread_num () == 2) |
| 41 |
a = svadd_s32_z (p, a, b); |
| 42 |
|
| 43 |
#pragma omp barrier |
| 44 |
if (omp_get_thread_num () == 0) |
| 45 |
{ |
| 46 |
compare_vec (a, svindex_s32 (8, 2)); |
| 47 |
compare_vec (b, svindex_s32 (8, 1)); |
| 48 |
compare_vecb (p, svptrue_b32 ()); |
| 49 |
b = svadd_s32_z (p, a, b); |
| 50 |
} |
| 51 |
|
| 52 |
#pragma omp barrier |
| 53 |
compare_vec (a, svindex_s32 (8, 2)); |
| 54 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 55 |
|
| 56 |
#pragma omp barrier |
| 57 |
if (omp_get_thread_num () == 0 || omp_get_thread_num () == 2) |
| 58 |
{ |
| 59 |
compare_vec (a, svindex_s32 (8, 2)); |
| 60 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 61 |
} |
| 62 |
} |
| 63 |
} |
| 64 |
|
| 65 |
void __attribute__ ((noipa)) |
| 66 |
explicit_shared (svint32_t a, svint32_t b, svbool_t p) |
| 67 |
{ |
| 68 |
|
| 69 |
#pragma omp parallel shared (a, b, p) num_threads (12) |
| 70 |
{ |
| 71 |
/* 'a', 'b' and 'p' are explicitly shared. */ |
| 72 |
compare_vec (a, svindex_s32 (0, 1)); |
| 73 |
compare_vec (b, svindex_s32 (8, 1)); |
| 74 |
compare_vecb (p, svptrue_b32 ()); |
| 75 |
|
| 76 |
#pragma omp barrier |
| 77 |
if (omp_get_thread_num () == 2) |
| 78 |
a = svadd_s32_z (p, a, b); |
| 79 |
|
| 80 |
#pragma omp barrier |
| 81 |
if (omp_get_thread_num () == 0) |
| 82 |
{ |
| 83 |
compare_vec (a, svindex_s32 (8, 2)); |
| 84 |
compare_vec (b, svindex_s32 (8, 1)); |
| 85 |
compare_vecb (p, svptrue_b32 ()); |
| 86 |
b = svadd_s32_z (p, a, b); |
| 87 |
} |
| 88 |
|
| 89 |
#pragma omp barrier |
| 90 |
compare_vec (a, svindex_s32 (8, 2)); |
| 91 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 92 |
|
| 93 |
#pragma omp barrier |
| 94 |
if (omp_get_thread_num () == 0 || omp_get_thread_num () == 2) |
| 95 |
{ |
| 96 |
compare_vec (a, svindex_s32 (8, 2)); |
| 97 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 98 |
} |
| 99 |
} |
| 100 |
} |
| 101 |
|
| 102 |
void __attribute__ ((noipa)) |
| 103 |
implicit_shared_no_default (svint32_t a, svint32_t b, svbool_t p) |
| 104 |
{ |
| 105 |
|
| 106 |
#pragma omp parallel num_threads (16) |
| 107 |
{ |
| 108 |
/* 'a', 'b' and 'p' are implicitly shared without default clause. */ |
| 109 |
compare_vec (a, svindex_s32 (0, 1)); |
| 110 |
compare_vec (b, svindex_s32 (8, 1)); |
| 111 |
compare_vecb (p, svptrue_b32 ()); |
| 112 |
|
| 113 |
#pragma omp barrier |
| 114 |
if (omp_get_thread_num () == 12) |
| 115 |
a = svadd_s32_z (p, a, b); |
| 116 |
|
| 117 |
#pragma omp barrier |
| 118 |
if (omp_get_thread_num () == 15) |
| 119 |
{ |
| 120 |
compare_vec (a, svindex_s32 (8, 2)); |
| 121 |
compare_vec (b, svindex_s32 (8, 1)); |
| 122 |
compare_vecb (p, svptrue_b32 ()); |
| 123 |
b = svadd_s32_z (p, a, b); |
| 124 |
} |
| 125 |
|
| 126 |
#pragma omp barrier |
| 127 |
compare_vec (a, svindex_s32 (8, 2)); |
| 128 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 129 |
|
| 130 |
#pragma omp barrier |
| 131 |
if (omp_get_thread_num () == 12 || omp_get_thread_num () == 15) |
| 132 |
{ |
| 133 |
compare_vec (a, svindex_s32 (8, 2)); |
| 134 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 135 |
} |
| 136 |
} |
| 137 |
|
| 138 |
} |
| 139 |
|
| 140 |
void __attribute__ ((noipa)) |
| 141 |
mix_shared (svint32_t b, svbool_t p) |
| 142 |
{ |
| 143 |
|
| 144 |
svint32_t a = svindex_s32 (0, 0); |
| 145 |
int *m = (int *) malloc (8 * sizeof (int)); |
| 146 |
int i; |
| 147 |
|
| 148 |
#pragma omp parallel for |
| 149 |
for (i = 0; i < 8; i++) |
| 150 |
m[i] = i; |
| 151 |
|
| 152 |
#pragma omp parallel num_threads (16) |
| 153 |
{ |
| 154 |
compare_vec (a, svindex_s32 (0, 0)); |
| 155 |
compare_vec (b, svindex_s32 (8, 1)); |
| 156 |
|
| 157 |
#pragma omp barrier |
| 158 |
/* 'm' is predetermined shared here. 'a' is implicitly shared here. */ |
| 159 |
if (omp_get_thread_num () == 10) |
| 160 |
a = svld1_s32 (svptrue_b32 (), m); |
| 161 |
|
| 162 |
#pragma omp barrier |
| 163 |
/* 'a', 'b' and 'p' are implicitly shared without default clause. */ |
| 164 |
compare_vec (a, svindex_s32 (0, 1)); |
| 165 |
compare_vec (b, svindex_s32 (8, 1)); |
| 166 |
compare_vecb (p, svptrue_b32 ()); |
| 167 |
|
| 168 |
#pragma omp barrier |
| 169 |
if (omp_get_thread_num () == 12) |
| 170 |
a = svadd_s32_z (p, a, b); |
| 171 |
|
| 172 |
#pragma omp barrier |
| 173 |
if (omp_get_thread_num () == 15) |
| 174 |
{ |
| 175 |
compare_vec (a, svindex_s32 (8, 2)); |
| 176 |
compare_vec (b, svindex_s32 (8, 1)); |
| 177 |
compare_vecb (p, svptrue_b32 ()); |
| 178 |
b = svadd_s32_z (p, a, b); |
| 179 |
} |
| 180 |
|
| 181 |
#pragma omp barrier |
| 182 |
if (omp_get_thread_num () == 12 || omp_get_thread_num () == 15) |
| 183 |
{ |
| 184 |
compare_vec (a, svindex_s32 (8, 2)); |
| 185 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 186 |
} |
| 187 |
|
| 188 |
#pragma omp barrier |
| 189 |
compare_vec (a, svindex_s32 (8, 2)); |
| 190 |
compare_vec (b, svadd_s32_z (p, svindex_s32 (8, 2), svindex_s32 (8, 1))); |
| 191 |
} |
| 192 |
} |
| 193 |
|
| 194 |
#define N __ARM_FEATURE_SVE_BITS |
| 195 |
#define FIXED_ATTR __attribute__((arm_sve_vector_bits (N))) |
| 196 |
|
| 197 |
typedef svint32_t v8si FIXED_ATTR; |
| 198 |
|
| 199 |
void __attribute__ ((noipa)) |
| 200 |
predetermined_shared_static (int n) |
| 201 |
{ |
| 202 |
|
| 203 |
int *m = (int *) malloc (8 * sizeof (int)); |
| 204 |
int i; |
| 205 |
|
| 206 |
#pragma omp parallel for |
| 207 |
/* 'm' is predetermined shared here. */ |
| 208 |
for (i = 0; i < 8; i++) |
| 209 |
m[i] = i; |
| 210 |
|
| 211 |
static v8si a = { 0, 1, 2, 3, 4, 5, 6, 7 }; |
| 212 |
|
| 213 |
#pragma omp parallel num_threads (16) |
| 214 |
{ |
| 215 |
/* 'a' is implicit shared here. */ |
| 216 |
if (n == 0) |
| 217 |
compare_vec (a, svindex_s32 (0, 1)); |
| 218 |
|
| 219 |
if (n == 1) |
| 220 |
compare_vec (a, svindex_s32 (1, 1)); |
| 221 |
|
| 222 |
#pragma omp barrier |
| 223 |
if (omp_get_thread_num () == 12) |
| 224 |
{ |
| 225 |
if (n == 0) |
| 226 |
compare_vec (a, svindex_s32 (0, 1)); |
| 227 |
|
| 228 |
if (n == 1) |
| 229 |
compare_vec (a, svindex_s32 (1, 1)); |
| 230 |
|
| 231 |
a = svadd_s32_z (svptrue_b32 (), a, svindex_s32 (1, 0)); |
| 232 |
} |
| 233 |
|
| 234 |
#pragma omp barrier |
| 235 |
if (n == 0) |
| 236 |
compare_vec (a, svindex_s32 (1, 1)); |
| 237 |
|
| 238 |
if (n == 1) |
| 239 |
compare_vec (a, svindex_s32 (2, 1)); |
| 240 |
} |
| 241 |
} |
| 242 |
|
| 243 |
|
| 244 |
int |
| 245 |
main () |
| 246 |
{ |
| 247 |
svint32_t x = svindex_s32 (0, 1); |
| 248 |
svint32_t y = svindex_s32 (8, 1); |
| 249 |
svbool_t p = svptrue_b32 (); |
| 250 |
|
| 251 |
/* Implicit shared. */ |
| 252 |
implicit_shared_default (x, y, p); |
| 253 |
|
| 254 |
/* Explicit shared. */ |
| 255 |
explicit_shared (x, y, p); |
| 256 |
|
| 257 |
/* Implicit shared with no default clause. */ |
| 258 |
implicit_shared_no_default (x, y, p); |
| 259 |
|
| 260 |
/* Mix shared. */ |
| 261 |
mix_shared (y, p); |
| 262 |
|
| 263 |
/* Predetermined and static shared. */ |
| 264 |
predetermined_shared_static (0); |
| 265 |
predetermined_shared_static (1); |
| 266 |
} |