| 1 |
/* OpenACC parallelism dimensions clauses: num_gangs, num_workers, |
| 2 |
vector_length. */ |
| 3 |
|
| 4 |
/* { dg-additional-options "-DEXPENSIVE" { target run_expensive_tests } } */ |
| 5 |
|
| 6 |
/* { dg-additional-options "--param=openacc-kernels=decompose" } */ |
| 7 |
|
| 8 |
/* { dg-additional-options "-fopt-info-all-omp" } |
| 9 |
{ dg-additional-options "-foffload=-fopt-info-all-omp" } */ |
| 10 |
|
| 11 |
/* { dg-additional-options "--param=openacc-privatization=noisy" } |
| 12 |
{ dg-additional-options "-foffload=--param=openacc-privatization=noisy" } |
| 13 |
Prune a few: uninteresting, and potentially varying depending on GCC configuration (data types): |
| 14 |
{ dg-prune-output {note: variable 'D\.[0-9]+' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} } */ |
| 15 |
|
| 16 |
/* It's only with Tcl 8.5 (released in 2007) that "the variable 'varName' |
| 17 |
passed to 'incr' may be unset, and in that case, it will be set to [...]", |
| 18 |
so to maintain compatibility with earlier Tcl releases, we manually |
| 19 |
initialize counter variables: |
| 20 |
{ dg-line l_dummy[variable c_compute 0 c_loop_i 0 c_loop_j 0 c_loop_k 0] } |
| 21 |
{ dg-message dummy {} { target iN-VAl-Id } l_dummy } to avoid |
| 22 |
"WARNING: dg-line var l_dummy defined, but not used". */ |
| 23 |
|
| 24 |
/* { dg-additional-options "-Wopenacc-parallelism" } for testing/documenting |
| 25 |
aspects of that functionality. */ |
| 26 |
|
| 27 |
/* See also '../libgomp.oacc-fortran/parallel-dims.f90'. */ |
| 28 |
|
| 29 |
#include <limits.h> |
| 30 |
#include <openacc.h> |
| 31 |
#include <gomp-constants.h> |
| 32 |
|
| 33 |
#pragma acc routine seq |
| 34 |
inline __attribute__ ((always_inline)) |
| 35 |
static int acc_gang () |
| 36 |
{ |
| 37 |
return __builtin_goacc_parlevel_id (GOMP_DIM_GANG); |
| 38 |
} |
| 39 |
|
| 40 |
#pragma acc routine seq |
| 41 |
inline __attribute__ ((always_inline)) |
| 42 |
static int acc_worker () |
| 43 |
{ |
| 44 |
return __builtin_goacc_parlevel_id (GOMP_DIM_WORKER); |
| 45 |
} |
| 46 |
|
| 47 |
#pragma acc routine seq |
| 48 |
inline __attribute__ ((always_inline)) |
| 49 |
static int acc_vector () |
| 50 |
{ |
| 51 |
return __builtin_goacc_parlevel_id (GOMP_DIM_VECTOR); |
| 52 |
} |
| 53 |
|
| 54 |
#ifdef EXPENSIVE |
| 55 |
#define N 100 |
| 56 |
#else |
| 57 |
#define N 50 |
| 58 |
#endif |
| 59 |
|
| 60 |
int main () |
| 61 |
{ |
| 62 |
acc_init (acc_device_default); |
| 63 |
|
| 64 |
/* OpenACC parallel construct. */ |
| 65 |
|
| 66 |
/* Non-positive value. */ |
| 67 |
|
| 68 |
/* GR, WS, VS. */ |
| 69 |
{ |
| 70 |
#define GANGS 0 |
| 71 |
/* { dg-warning {'num_gangs' value must be positive} {} { target c } .-1 } */ |
| 72 |
int gangs_actual = GANGS; |
| 73 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 74 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 75 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 76 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 77 |
copy (gangs_actual) \ |
| 78 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) \ |
| 79 |
num_gangs (GANGS) |
| 80 |
/* { dg-note {in expansion of macro 'GANGS'} {} { target c } .-1 } */ |
| 81 |
/* { dg-warning {'num_gangs' value must be positive} {} { target c++ } .-2 } */ |
| 82 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 83 |
{ |
| 84 |
/* We're actually executing with num_gangs (1). */ |
| 85 |
gangs_actual = 1; |
| 86 |
for (int i = N * gangs_actual; i > -N * gangs_actual; --i) |
| 87 |
{ |
| 88 |
gangs_min = gangs_max = acc_gang (); |
| 89 |
workers_min = workers_max = acc_worker (); |
| 90 |
vectors_min = vectors_max = acc_vector (); |
| 91 |
} |
| 92 |
} |
| 93 |
if (gangs_actual != 1) |
| 94 |
__builtin_abort (); |
| 95 |
if (gangs_min != 0 || gangs_max != gangs_actual - 1 |
| 96 |
|| workers_min != 0 || workers_max != 0 |
| 97 |
|| vectors_min != 0 || vectors_max != 0) |
| 98 |
__builtin_abort (); |
| 99 |
#undef GANGS |
| 100 |
} |
| 101 |
|
| 102 |
/* GP, WS, VS. */ |
| 103 |
{ |
| 104 |
#define GANGS 0 |
| 105 |
/* { dg-warning {'num_gangs' value must be positive} {} { target c } .-1 } */ |
| 106 |
int gangs_actual = GANGS; |
| 107 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 108 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 109 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 110 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 111 |
copy (gangs_actual) \ |
| 112 |
num_gangs (GANGS) |
| 113 |
/* { dg-note {in expansion of macro 'GANGS'} {} { target c } .-1 } */ |
| 114 |
/* { dg-warning {'num_gangs' value must be positive} {} { target c++ } .-2 } */ |
| 115 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 116 |
/* { dg-warning {region contains gang partitioned code but is not gang partitioned} {} { target *-*-* } l_compute$c_compute } */ |
| 117 |
{ |
| 118 |
/* We're actually executing with num_gangs (1). */ |
| 119 |
gangs_actual = 1; |
| 120 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 121 |
gang \ |
| 122 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 123 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 124 |
/* { dg-optimized {assigned OpenACC gang loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 125 |
for (int i = N * gangs_actual; i > -N * gangs_actual; --i) |
| 126 |
{ |
| 127 |
gangs_min = gangs_max = acc_gang (); |
| 128 |
workers_min = workers_max = acc_worker (); |
| 129 |
vectors_min = vectors_max = acc_vector (); |
| 130 |
} |
| 131 |
} |
| 132 |
if (gangs_actual != 1) |
| 133 |
__builtin_abort (); |
| 134 |
if (gangs_min != 0 || gangs_max != gangs_actual - 1 |
| 135 |
|| workers_min != 0 || workers_max != 0 |
| 136 |
|| vectors_min != 0 || vectors_max != 0) |
| 137 |
__builtin_abort (); |
| 138 |
#undef GANGS |
| 139 |
} |
| 140 |
|
| 141 |
/* GR, WP, VS. */ |
| 142 |
{ |
| 143 |
#define WORKERS 0 |
| 144 |
/* { dg-warning {'num_workers' value must be positive} {} { target c } .-1 } */ |
| 145 |
int workers_actual = WORKERS; |
| 146 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 147 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 148 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 149 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 150 |
copy (workers_actual) \ |
| 151 |
num_workers (WORKERS) |
| 152 |
/* { dg-note {in expansion of macro 'WORKERS'} {} { target c } .-1 } */ |
| 153 |
/* { dg-warning {'num_workers' value must be positive} {} { target c++ } .-2 } */ |
| 154 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 155 |
/* { dg-warning {region contains worker partitioned code but is not worker partitioned} {} { target *-*-* } l_compute$c_compute } */ |
| 156 |
{ |
| 157 |
/* We're actually executing with num_workers (1). */ |
| 158 |
workers_actual = 1; |
| 159 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 160 |
worker \ |
| 161 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 162 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 163 |
/* { dg-optimized {assigned OpenACC worker loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 164 |
for (int i = N * workers_actual; i > -N * workers_actual; --i) |
| 165 |
{ |
| 166 |
gangs_min = gangs_max = acc_gang (); |
| 167 |
workers_min = workers_max = acc_worker (); |
| 168 |
vectors_min = vectors_max = acc_vector (); |
| 169 |
} |
| 170 |
} |
| 171 |
if (workers_actual != 1) |
| 172 |
__builtin_abort (); |
| 173 |
if (gangs_min != 0 || gangs_max != 0 |
| 174 |
|| workers_min != 0 || workers_max != workers_actual - 1 |
| 175 |
|| vectors_min != 0 || vectors_max != 0) |
| 176 |
__builtin_abort (); |
| 177 |
#undef WORKERS |
| 178 |
} |
| 179 |
|
| 180 |
/* GR, WS, VP. */ |
| 181 |
{ |
| 182 |
#define VECTORS 0 |
| 183 |
/* { dg-warning {'vector_length' value must be positive} {} { target c } .-1 } */ |
| 184 |
int vectors_actual = VECTORS; |
| 185 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 186 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 187 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 188 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 189 |
copy (vectors_actual) \ |
| 190 |
vector_length (VECTORS) |
| 191 |
/* { dg-note {in expansion of macro 'VECTORS'} {} { target c } .-1 } */ |
| 192 |
/* { dg-warning {'vector_length' value must be positive} {} { target c++ } .-2 } */ |
| 193 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 194 |
/* { dg-warning {region contains vector partitioned code but is not vector partitioned} {} { target *-*-* } l_compute$c_compute } */ |
| 195 |
/* { dg-warning {using 'vector_length \(32\)', ignoring 1} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 196 |
{ |
| 197 |
/* We're actually executing with vector_length (1), just the GCC nvptx |
| 198 |
back end enforces vector_length (32). */ |
| 199 |
if (acc_on_device (acc_device_nvidia)) |
| 200 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 201 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 202 |
vectors_actual = 32; |
| 203 |
else |
| 204 |
vectors_actual = 1; |
| 205 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 206 |
vector \ |
| 207 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 208 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 209 |
/* { dg-optimized {assigned OpenACC vector loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 210 |
for (int i = N * vectors_actual; i > -N * vectors_actual; --i) |
| 211 |
{ |
| 212 |
gangs_min = gangs_max = acc_gang (); |
| 213 |
workers_min = workers_max = acc_worker (); |
| 214 |
vectors_min = vectors_max = acc_vector (); |
| 215 |
} |
| 216 |
} |
| 217 |
if (acc_get_device_type () == acc_device_nvidia) |
| 218 |
{ |
| 219 |
if (vectors_actual != 32) |
| 220 |
__builtin_abort (); |
| 221 |
} |
| 222 |
else |
| 223 |
if (vectors_actual != 1) |
| 224 |
__builtin_abort (); |
| 225 |
if (gangs_min != 0 || gangs_max != 0 |
| 226 |
|| workers_min != 0 || workers_max != 0 |
| 227 |
|| vectors_min != 0 || vectors_max != vectors_actual - 1) |
| 228 |
__builtin_abort (); |
| 229 |
#undef VECTORS |
| 230 |
} |
| 231 |
|
| 232 |
|
| 233 |
/* High value. */ |
| 234 |
|
| 235 |
/* GR, WS, VS. */ |
| 236 |
{ |
| 237 |
/* There is no actual limit for the number of gangs, so we try with a |
| 238 |
rather high value. */ |
| 239 |
int gangs = 12345; |
| 240 |
int gangs_actual = gangs; |
| 241 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 242 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 243 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 244 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 245 |
copy (gangs_actual) \ |
| 246 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) \ |
| 247 |
num_gangs (gangs) |
| 248 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 249 |
/* { dg-bogus {warning: region is gang partitioned but does not contain gang partitioned code} {TODO 'reduction'} { xfail *-*-* } l_compute$c_compute } */ |
| 250 |
{ |
| 251 |
if (acc_on_device (acc_device_host)) |
| 252 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 253 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 254 |
{ |
| 255 |
/* We're actually executing with num_gangs (1). */ |
| 256 |
gangs_actual = 1; |
| 257 |
} |
| 258 |
/* As we're executing GR not GP, don't multiply with a "gangs_actual" |
| 259 |
factor. */ |
| 260 |
for (int i = N /* * gangs_actual */; i > -N /* * gangs_actual */; --i) |
| 261 |
{ |
| 262 |
gangs_min = gangs_max = acc_gang (); |
| 263 |
workers_min = workers_max = acc_worker (); |
| 264 |
vectors_min = vectors_max = acc_vector (); |
| 265 |
} |
| 266 |
} |
| 267 |
if (gangs_actual < 1) |
| 268 |
__builtin_abort (); |
| 269 |
if (gangs_min != 0 || gangs_max != gangs_actual - 1 |
| 270 |
|| workers_min != 0 || workers_max != 0 |
| 271 |
|| vectors_min != 0 || vectors_max != 0) |
| 272 |
__builtin_abort (); |
| 273 |
} |
| 274 |
|
| 275 |
/* GP, WS, VS. */ |
| 276 |
{ |
| 277 |
/* There is no actual limit for the number of gangs, so we try with a |
| 278 |
rather high value. */ |
| 279 |
int gangs = 12345; |
| 280 |
int gangs_actual = gangs; |
| 281 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 282 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 283 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 284 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 285 |
copy (gangs_actual) \ |
| 286 |
num_gangs (gangs) |
| 287 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 288 |
{ |
| 289 |
if (acc_on_device (acc_device_host)) |
| 290 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 291 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 292 |
{ |
| 293 |
/* We're actually executing with num_gangs (1). */ |
| 294 |
gangs_actual = 1; |
| 295 |
} |
| 296 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 297 |
gang \ |
| 298 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 299 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 300 |
/* { dg-optimized {assigned OpenACC gang loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 301 |
for (int i = N * gangs_actual; i > -N * gangs_actual; --i) |
| 302 |
{ |
| 303 |
gangs_min = gangs_max = acc_gang (); |
| 304 |
workers_min = workers_max = acc_worker (); |
| 305 |
vectors_min = vectors_max = acc_vector (); |
| 306 |
} |
| 307 |
} |
| 308 |
if (gangs_actual < 1) |
| 309 |
__builtin_abort (); |
| 310 |
if (gangs_min != 0 || gangs_max != gangs_actual - 1 |
| 311 |
|| workers_min != 0 || workers_max != 0 |
| 312 |
|| vectors_min != 0 || vectors_max != 0) |
| 313 |
__builtin_abort (); |
| 314 |
} |
| 315 |
|
| 316 |
/* GR, WP, VS. */ |
| 317 |
{ |
| 318 |
/* We try with an outrageously large value. */ |
| 319 |
#define WORKERS 2 << 20 |
| 320 |
int workers_actual = WORKERS; |
| 321 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 322 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 323 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 324 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 325 |
copy (workers_actual) \ |
| 326 |
num_workers (WORKERS) |
| 327 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 328 |
/* { dg-warning {using 'num_workers \(32\)', ignoring 2097152} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 329 |
{ |
| 330 |
if (acc_on_device (acc_device_host)) |
| 331 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 332 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 333 |
{ |
| 334 |
/* We're actually executing with num_workers (1). */ |
| 335 |
workers_actual = 1; |
| 336 |
} |
| 337 |
else if (acc_on_device (acc_device_nvidia)) |
| 338 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 339 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 340 |
{ |
| 341 |
/* The GCC nvptx back end enforces num_workers (32). */ |
| 342 |
workers_actual = 32; |
| 343 |
} |
| 344 |
else if (acc_on_device (acc_device_radeon)) |
| 345 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 346 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 347 |
{ |
| 348 |
/* The GCC GCN back end is limited to num_workers (16). */ |
| 349 |
workers_actual = 16; |
| 350 |
} |
| 351 |
else |
| 352 |
__builtin_abort (); |
| 353 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 354 |
worker \ |
| 355 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 356 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 357 |
/* { dg-optimized {assigned OpenACC worker loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 358 |
for (int i = N * workers_actual; i > -N * workers_actual; --i) |
| 359 |
{ |
| 360 |
gangs_min = gangs_max = acc_gang (); |
| 361 |
workers_min = workers_max = acc_worker (); |
| 362 |
vectors_min = vectors_max = acc_vector (); |
| 363 |
} |
| 364 |
} |
| 365 |
if (workers_actual < 1) |
| 366 |
__builtin_abort (); |
| 367 |
if (gangs_min != 0 || gangs_max != 0 |
| 368 |
|| workers_min != 0 || workers_max != workers_actual - 1 |
| 369 |
|| vectors_min != 0 || vectors_max != 0) |
| 370 |
__builtin_abort (); |
| 371 |
#undef WORKERS |
| 372 |
} |
| 373 |
|
| 374 |
/* GR, WP, VS. */ |
| 375 |
{ |
| 376 |
/* We try with an outrageously large value. */ |
| 377 |
int workers = 2 << 20; |
| 378 |
/* For nvptx offloading, this one will not result in "using num_workers |
| 379 |
(32), ignoring runtime setting", and will in fact try to launch with |
| 380 |
"num_workers (workers)", which will run into "libgomp: cuLaunchKernel |
| 381 |
error: invalid argument". So, limit ourselves here. */ |
| 382 |
if (acc_get_device_type () == acc_device_nvidia) |
| 383 |
workers = 32; |
| 384 |
int workers_actual = workers; |
| 385 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 386 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 387 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 388 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 389 |
copy (workers_actual) \ |
| 390 |
num_workers (workers) |
| 391 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 392 |
{ |
| 393 |
if (acc_on_device (acc_device_host)) |
| 394 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 395 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 396 |
{ |
| 397 |
/* We're actually executing with num_workers (1). */ |
| 398 |
workers_actual = 1; |
| 399 |
} |
| 400 |
else if (acc_on_device (acc_device_nvidia)) |
| 401 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 402 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 403 |
{ |
| 404 |
/* We're actually executing with num_workers (32). */ |
| 405 |
/* workers_actual = 32; */ |
| 406 |
} |
| 407 |
else if (acc_on_device (acc_device_radeon)) |
| 408 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 409 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 410 |
{ |
| 411 |
/* The GCC GCN back end is limited to num_workers (16). */ |
| 412 |
workers_actual = 16; |
| 413 |
} |
| 414 |
else |
| 415 |
__builtin_abort (); |
| 416 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 417 |
worker \ |
| 418 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 419 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 420 |
/* { dg-optimized {assigned OpenACC worker loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 421 |
for (int i = N * workers_actual; i > -N * workers_actual; --i) |
| 422 |
{ |
| 423 |
gangs_min = gangs_max = acc_gang (); |
| 424 |
workers_min = workers_max = acc_worker (); |
| 425 |
vectors_min = vectors_max = acc_vector (); |
| 426 |
} |
| 427 |
} |
| 428 |
if (workers_actual < 1) |
| 429 |
__builtin_abort (); |
| 430 |
if (gangs_min != 0 || gangs_max != 0 |
| 431 |
|| workers_min != 0 || workers_max != workers_actual - 1 |
| 432 |
|| vectors_min != 0 || vectors_max != 0) |
| 433 |
__builtin_abort (); |
| 434 |
} |
| 435 |
|
| 436 |
/* GR, WS, VP. */ |
| 437 |
{ |
| 438 |
/* We try with an outrageously large value. */ |
| 439 |
#define VECTORS 2 << 20 |
| 440 |
int vectors_actual = VECTORS; |
| 441 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 442 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 443 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 444 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 445 |
copy (vectors_actual) \ |
| 446 |
vector_length (VECTORS) |
| 447 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 448 |
/* { dg-warning {using 'vector_length \(1024\)', ignoring 2097152} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 449 |
{ |
| 450 |
if (acc_on_device (acc_device_host)) |
| 451 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 452 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 453 |
{ |
| 454 |
/* We're actually executing with vector_length (1). */ |
| 455 |
vectors_actual = 1; |
| 456 |
} |
| 457 |
else if (acc_on_device (acc_device_nvidia)) |
| 458 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 459 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 460 |
{ |
| 461 |
/* The GCC nvptx back end reduces to vector_length (1024). */ |
| 462 |
vectors_actual = 1024; |
| 463 |
} |
| 464 |
else if (acc_on_device (acc_device_radeon)) |
| 465 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 466 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 467 |
{ |
| 468 |
/* The GCC GCN back end enforces vector_length (1): autovectorize. */ |
| 469 |
vectors_actual = 1; |
| 470 |
} |
| 471 |
else |
| 472 |
__builtin_abort (); |
| 473 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 474 |
vector \ |
| 475 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 476 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 477 |
/* { dg-optimized {assigned OpenACC vector loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 478 |
for (int i = N * vectors_actual; i > -N * vectors_actual; --i) |
| 479 |
{ |
| 480 |
gangs_min = gangs_max = acc_gang (); |
| 481 |
workers_min = workers_max = acc_worker (); |
| 482 |
vectors_min = vectors_max = acc_vector (); |
| 483 |
} |
| 484 |
} |
| 485 |
if (vectors_actual < 1) |
| 486 |
__builtin_abort (); |
| 487 |
if (gangs_min != 0 || gangs_max != 0 |
| 488 |
|| workers_min != 0 || workers_max != 0 |
| 489 |
|| vectors_min != 0 || vectors_max != vectors_actual - 1) |
| 490 |
__builtin_abort (); |
| 491 |
#undef VECTORS |
| 492 |
} |
| 493 |
|
| 494 |
/* GR, WS, VP. */ |
| 495 |
{ |
| 496 |
/* We try with an outrageously large value. */ |
| 497 |
int vectors = 2 << 20; |
| 498 |
int vectors_actual = vectors; |
| 499 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 500 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 501 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 502 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 503 |
copy (vectors_actual) \ |
| 504 |
vector_length (vectors) |
| 505 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 506 |
/* { dg-warning {using 'vector_length \(32\)', ignoring runtime setting} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 507 |
{ |
| 508 |
if (acc_on_device (acc_device_host)) |
| 509 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 510 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 511 |
{ |
| 512 |
/* We're actually executing with vector_length (1). */ |
| 513 |
vectors_actual = 1; |
| 514 |
} |
| 515 |
else if (acc_on_device (acc_device_nvidia)) |
| 516 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 517 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 518 |
{ |
| 519 |
/* The GCC nvptx back end enforces vector_length (32). */ |
| 520 |
vectors_actual = 32; |
| 521 |
} |
| 522 |
else if (acc_on_device (acc_device_radeon)) |
| 523 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 524 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 525 |
{ |
| 526 |
/* Because of the way vectors are implemented for GCN, a vector loop |
| 527 |
containing a seq routine call will not vectorize calls to that |
| 528 |
routine. Hence, we'll only get one "vector". */ |
| 529 |
vectors_actual = 1; |
| 530 |
} |
| 531 |
else |
| 532 |
__builtin_abort (); |
| 533 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 534 |
vector \ |
| 535 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 536 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 537 |
/* { dg-optimized {assigned OpenACC vector loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 538 |
for (int i = N * vectors_actual; i > -N * vectors_actual; --i) |
| 539 |
{ |
| 540 |
gangs_min = gangs_max = acc_gang (); |
| 541 |
workers_min = workers_max = acc_worker (); |
| 542 |
vectors_min = vectors_max = acc_vector (); |
| 543 |
} |
| 544 |
} |
| 545 |
if (vectors_actual < 1) |
| 546 |
__builtin_abort (); |
| 547 |
if (gangs_min != 0 || gangs_max != 0 |
| 548 |
|| workers_min != 0 || workers_max != 0 |
| 549 |
|| vectors_min != 0 || vectors_max != vectors_actual - 1) |
| 550 |
__builtin_abort (); |
| 551 |
} |
| 552 |
|
| 553 |
|
| 554 |
/* Composition of GP, WP, VP. */ |
| 555 |
{ |
| 556 |
int gangs = 12345; |
| 557 |
/* With nvptx offloading, multi-level reductions apparently are very slow |
| 558 |
in the following case. So, limit ourselves here. */ |
| 559 |
if (acc_get_device_type () == acc_device_nvidia) |
| 560 |
gangs = 3; |
| 561 |
/* Similar appears to be true for GCN. */ |
| 562 |
if (acc_get_device_type () == acc_device_radeon) |
| 563 |
gangs = 3; |
| 564 |
int gangs_actual = gangs; |
| 565 |
#define WORKERS 3 |
| 566 |
int workers_actual = WORKERS; |
| 567 |
#define VECTORS 11 |
| 568 |
int vectors_actual = VECTORS; |
| 569 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 570 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 571 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 572 |
#pragma acc parallel /* { dg-line l_compute[incr c_compute] } */ \ |
| 573 |
copy (gangs_actual, workers_actual, vectors_actual) \ |
| 574 |
num_gangs (gangs) \ |
| 575 |
num_workers (WORKERS) \ |
| 576 |
vector_length (VECTORS) |
| 577 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 578 |
/* { dg-warning {using 'vector_length \(32\)', ignoring 11} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 579 |
{ |
| 580 |
if (acc_on_device (acc_device_host)) |
| 581 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 582 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 583 |
{ |
| 584 |
/* We're actually executing with num_gangs (1), num_workers (1), |
| 585 |
vector_length (1). */ |
| 586 |
gangs_actual = 1; |
| 587 |
workers_actual = 1; |
| 588 |
vectors_actual = 1; |
| 589 |
} |
| 590 |
else if (acc_on_device (acc_device_nvidia)) |
| 591 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 592 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 593 |
{ |
| 594 |
/* The GCC nvptx back end enforces vector_length (32). */ |
| 595 |
vectors_actual = 32; |
| 596 |
} |
| 597 |
else if (acc_on_device (acc_device_radeon)) |
| 598 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 599 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 600 |
{ |
| 601 |
/* See above comments about GCN vectors_actual. */ |
| 602 |
vectors_actual = 1; |
| 603 |
} |
| 604 |
else |
| 605 |
__builtin_abort (); |
| 606 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 607 |
gang \ |
| 608 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 609 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 610 |
/* { dg-note {variable 'j' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 611 |
/* { dg-optimized {assigned OpenACC gang loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 612 |
for (int i = N * gangs_actual; i > -N * gangs_actual; --i) |
| 613 |
#pragma acc loop /* { dg-line l_loop_j[incr c_loop_j] } */ \ |
| 614 |
worker \ |
| 615 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 616 |
/* { dg-note {variable 'j' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_j$c_loop_j } */ |
| 617 |
/* { dg-note {variable 'k' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_j$c_loop_j } */ |
| 618 |
/* { dg-optimized {assigned OpenACC worker loop parallelism} {} { target *-*-* } l_loop_j$c_loop_j } */ |
| 619 |
for (int j = N * workers_actual; j > -N * workers_actual; --j) |
| 620 |
#pragma acc loop /* { dg-line l_loop_k[incr c_loop_k] } */ \ |
| 621 |
vector \ |
| 622 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 623 |
/* { dg-note {variable 'k' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_k$c_loop_k } */ |
| 624 |
/* { dg-optimized {assigned OpenACC vector loop parallelism} {} { target *-*-* } l_loop_k$c_loop_k } */ |
| 625 |
for (int k = N * vectors_actual; k > -N * vectors_actual; --k) |
| 626 |
{ |
| 627 |
gangs_min = gangs_max = acc_gang (); |
| 628 |
workers_min = workers_max = acc_worker (); |
| 629 |
vectors_min = vectors_max = acc_vector (); |
| 630 |
} |
| 631 |
} |
| 632 |
if (gangs_min != 0 || gangs_max != gangs_actual - 1 |
| 633 |
|| workers_min != 0 || workers_max != workers_actual - 1 |
| 634 |
|| vectors_min != 0 || vectors_max != vectors_actual - 1) |
| 635 |
__builtin_abort (); |
| 636 |
#undef VECTORS |
| 637 |
#undef WORKERS |
| 638 |
} |
| 639 |
|
| 640 |
|
| 641 |
/* OpenACC kernels construct. */ |
| 642 |
|
| 643 |
/* We can't test parallelized OpenACC kernels constructs in this way: use of |
| 644 |
the acc_gang, acc_worker, acc_vector functions will make the construct |
| 645 |
unparallelizable. */ |
| 646 |
|
| 647 |
|
| 648 |
/* Unparallelized OpenACC kernels constructs must get launched as 1 x 1 x 1 |
| 649 |
kernels. */ |
| 650 |
{ |
| 651 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 652 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 653 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 654 |
#pragma acc kernels /* { dg-line l_compute[incr c_compute] } */ |
| 655 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'vectors_max' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 656 |
{ dg-note {variable 'vectors_max' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 657 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'vectors_min' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 658 |
{ dg-note {variable 'vectors_min' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 659 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'workers_max' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 660 |
{ dg-note {variable 'workers_max' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 661 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'workers_min' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 662 |
{ dg-note {variable 'workers_min' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 663 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'gangs_max' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 664 |
{ dg-note {variable 'gangs_max' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 665 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'gangs_min' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 666 |
{ dg-note {variable 'gangs_min' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 667 |
{ |
| 668 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 669 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 670 |
/* { dg-note {forwarded loop nest in OpenACC 'kernels' region to 'parloops' for analysis} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 671 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 672 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 673 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 674 |
for (int i = N; i > -N; --i) |
| 675 |
{ |
| 676 |
/* This is to make the loop unparallelizable. */ |
| 677 |
asm volatile ("" : : : "memory"); |
| 678 |
|
| 679 |
gangs_min = gangs_max = acc_gang (); |
| 680 |
workers_min = workers_max = acc_worker (); |
| 681 |
vectors_min = vectors_max = acc_vector (); |
| 682 |
} |
| 683 |
} |
| 684 |
if (gangs_min != 0 || gangs_max != 1 - 1 |
| 685 |
|| workers_min != 0 || workers_max != 1 - 1 |
| 686 |
|| vectors_min != 0 || vectors_max != 1 - 1) |
| 687 |
__builtin_abort (); |
| 688 |
} |
| 689 |
|
| 690 |
|
| 691 |
/* Unparallelized OpenACC kernels constructs must get launched as 1 x 1 x 1 |
| 692 |
kernels even when there are explicit num_gangs, num_workers, or |
| 693 |
vector_length clauses. */ |
| 694 |
{ |
| 695 |
int gangs = 5; |
| 696 |
#define WORKERS 5 |
| 697 |
#define VECTORS 13 |
| 698 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 699 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 700 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 701 |
#pragma acc kernels /* { dg-line l_compute[incr c_compute] } */ \ |
| 702 |
num_gangs (gangs) \ |
| 703 |
num_workers (WORKERS) \ |
| 704 |
vector_length (VECTORS) |
| 705 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'vectors_max' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 706 |
{ dg-note {variable 'vectors_max' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 707 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'vectors_min' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 708 |
{ dg-note {variable 'vectors_min' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 709 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'workers_max' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 710 |
{ dg-note {variable 'workers_max' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 711 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'workers_min' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 712 |
{ dg-note {variable 'workers_min' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 713 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'gangs_max' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 714 |
{ dg-note {variable 'gangs_max' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 715 |
/* { dg-note {OpenACC 'kernels' decomposition: variable 'gangs_min' in 'copy' clause requested to be made addressable} {} { target *-*-* } l_compute$c_compute } |
| 716 |
{ dg-note {variable 'gangs_min' made addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 717 |
{ |
| 718 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 719 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 720 |
/* { dg-note {forwarded loop nest in OpenACC 'kernels' region to 'parloops' for analysis} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 721 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 722 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 723 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 724 |
for (int i = N; i > -N; --i) |
| 725 |
{ |
| 726 |
/* This is to make the loop unparallelizable. */ |
| 727 |
asm volatile ("" : : : "memory"); |
| 728 |
|
| 729 |
gangs_min = gangs_max = acc_gang (); |
| 730 |
workers_min = workers_max = acc_worker (); |
| 731 |
vectors_min = vectors_max = acc_vector (); |
| 732 |
} |
| 733 |
} |
| 734 |
if (gangs_min != 0 || gangs_max != 1 - 1 |
| 735 |
|| workers_min != 0 || workers_max != 1 - 1 |
| 736 |
|| vectors_min != 0 || vectors_max != 1 - 1) |
| 737 |
__builtin_abort (); |
| 738 |
#undef VECTORS |
| 739 |
#undef WORKERS |
| 740 |
} |
| 741 |
|
| 742 |
|
| 743 |
/* OpenACC serial construct. */ |
| 744 |
|
| 745 |
/* GR, WS, VS. */ |
| 746 |
{ |
| 747 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 748 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 749 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 750 |
#pragma acc serial /* { dg-line l_compute[incr c_compute] } */ \ |
| 751 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 752 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 753 |
/* { dg-warning {using 'vector_length \(32\)', ignoring 1} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 754 |
{ |
| 755 |
for (int i = N; i > -N; i--) |
| 756 |
{ |
| 757 |
gangs_min = gangs_max = acc_gang (); |
| 758 |
workers_min = workers_max = acc_worker (); |
| 759 |
vectors_min = vectors_max = acc_vector (); |
| 760 |
} |
| 761 |
} |
| 762 |
if (gangs_min != 0 || gangs_max != 1 - 1 |
| 763 |
|| workers_min != 0 || workers_max != 1 - 1 |
| 764 |
|| vectors_min != 0 || vectors_max != 1 - 1) |
| 765 |
__builtin_abort (); |
| 766 |
} |
| 767 |
|
| 768 |
/* Composition of GP, WP, VP. */ |
| 769 |
{ |
| 770 |
int vectors_actual = 1; /* Implicit 'vector_length (1)' clause. */ |
| 771 |
int gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max; |
| 772 |
gangs_min = workers_min = vectors_min = INT_MAX; |
| 773 |
gangs_max = workers_max = vectors_max = INT_MIN; |
| 774 |
#pragma acc serial /* { dg-line l_compute[incr c_compute] } */ \ |
| 775 |
copy (vectors_actual) \ |
| 776 |
copy (gangs_min, gangs_max, workers_min, workers_max, vectors_min, vectors_max) |
| 777 |
/* { dg-note {variable 'i' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_compute$c_compute } */ |
| 778 |
/* { dg-bogus {warning: region contains gang partitioned code but is not gang partitioned} {TODO 'serial'} { xfail *-*-* } l_compute$c_compute } |
| 779 |
{ dg-bogus {warning: region contains worker partitioned code but is not worker partitioned} {TODO 'serial'} { xfail *-*-* } l_compute$c_compute } |
| 780 |
{ dg-bogus {warning: region contains vector partitioned code but is not vector partitioned} {TODO 'serial'} { xfail *-*-* } l_compute$c_compute } */ |
| 781 |
/* { dg-warning {using 'vector_length \(32\)', ignoring 1} {} { target openacc_nvidia_accel_selected } l_compute$c_compute } */ |
| 782 |
{ |
| 783 |
if (acc_on_device (acc_device_nvidia)) |
| 784 |
/* { dg-optimized {assigned OpenACC seq loop parallelism} {} { target { c++ && { ! __OPTIMIZE__ } } } .-1 } |
| 785 |
..., as without optimizations, we're not inlining the C++ 'acc_on_device' wrapper. */ |
| 786 |
{ |
| 787 |
/* The GCC nvptx back end enforces vector_length (32). */ |
| 788 |
/* It's unclear if that's actually permissible here; |
| 789 |
<https://github.com/OpenACC/openacc-spec/issues/238> "OpenACC |
| 790 |
'serial' construct might not actually be serial". */ |
| 791 |
vectors_actual = 32; |
| 792 |
} |
| 793 |
#pragma acc loop /* { dg-line l_loop_i[incr c_loop_i] } */ \ |
| 794 |
gang \ |
| 795 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 796 |
/* { dg-note {variable 'i' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 797 |
/* { dg-note {variable 'j' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 798 |
/* { dg-optimized {assigned OpenACC gang loop parallelism} {} { target *-*-* } l_loop_i$c_loop_i } */ |
| 799 |
for (int i = N; i > -N; i--) |
| 800 |
#pragma acc loop /* { dg-line l_loop_j[incr c_loop_j] } */ \ |
| 801 |
worker \ |
| 802 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 803 |
/* { dg-note {variable 'j' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_j$c_loop_j } */ |
| 804 |
/* { dg-note {variable 'k' declared in block isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_j$c_loop_j } */ |
| 805 |
/* { dg-optimized {assigned OpenACC worker loop parallelism} {} { target *-*-* } l_loop_j$c_loop_j } */ |
| 806 |
for (int j = N; j > -N; j--) |
| 807 |
#pragma acc loop /* { dg-line l_loop_k[incr c_loop_k] } */ \ |
| 808 |
vector \ |
| 809 |
reduction (min: gangs_min, workers_min, vectors_min) reduction (max: gangs_max, workers_max, vectors_max) |
| 810 |
/* { dg-note {variable 'k' in 'private' clause isn't candidate for adjusting OpenACC privatization level: not addressable} {} { target *-*-* } l_loop_k$c_loop_k } */ |
| 811 |
/* { dg-optimized {assigned OpenACC vector loop parallelism} {} { target *-*-* } l_loop_k$c_loop_k } */ |
| 812 |
for (int k = N * vectors_actual; k > -N * vectors_actual; k--) |
| 813 |
{ |
| 814 |
gangs_min = gangs_max = acc_gang (); |
| 815 |
workers_min = workers_max = acc_worker (); |
| 816 |
vectors_min = vectors_max = acc_vector (); |
| 817 |
} |
| 818 |
} |
| 819 |
if (acc_get_device_type () == acc_device_nvidia) |
| 820 |
{ |
| 821 |
if (vectors_actual != 32) |
| 822 |
__builtin_abort (); |
| 823 |
} |
| 824 |
else |
| 825 |
if (vectors_actual != 1) |
| 826 |
__builtin_abort (); |
| 827 |
if (gangs_min != 0 || gangs_max != 1 - 1 |
| 828 |
|| workers_min != 0 || workers_max != 1 - 1 |
| 829 |
|| vectors_min != 0 || vectors_max != vectors_actual - 1) |
| 830 |
__builtin_abort (); |
| 831 |
} |
| 832 |
|
| 833 |
|
| 834 |
return 0; |
| 835 |
} |