3 #include <gomp-constants.h>
5 #define N (32*32*32+17)
7 #pragma acc routine worker
8 void __attribute__ ((noinline
)) worker (int ary
[N
])
10 #pragma acc loop worker vector
11 for (unsigned ix
= 0; ix
< N
; ix
++)
13 if (acc_on_device (acc_device_not_host
))
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 ary
[ix
] = (g
<< 16) | (w
<< 8) | v
;
33 int workersize
, vectorsize
;
35 for (ix
= 0; ix
< N
;ix
++)
40 #pragma acc parallel num_workers(NW) vector_length(VL) \
43 ondev
= acc_on_device (acc_device_not_host
);
48 #ifdef ACC_DEVICE_TYPE_radeon
49 /* AMD GCN has an upper limit of 'num_workers(16)'. */
52 /* AMD GCN uses the autovectorizer for the vector dimension: the use
53 of a function call in vector-partitioned code in this test is not
54 currently supported. */
58 for (ix
= 0; ix
< N
; ix
++)
64 int w
= (ix
/ vectorsize
) % workersize
;
65 int v
= ix
% vectorsize
;
67 expected
= (g
<< 16) | (w
<< 8) | v
;
70 if (ary
[ix
] != expected
)
73 printf ("ary[%d]=%x expected %x\n", ix
, ary
[ix
], expected
);