| 1 |
/* { dg-do run { target openacc_nvidia_accel_selected } } */ |
| 2 |
/* { dg-additional-options "-lcuda -lcublas -lcudart" } */ |
| 3 |
/* { dg-require-effective-target openacc_cublas } */ |
| 4 |
/* { dg-require-effective-target openacc_cudart } */ |
| 5 |
|
| 6 |
#include <stdio.h> |
| 7 |
#include <stdlib.h> |
| 8 |
#include <cuda.h> |
| 9 |
#include <cuda_runtime_api.h> |
| 10 |
#include <cublas_v2.h> |
| 11 |
#include <openacc.h> |
| 12 |
|
| 13 |
void |
| 14 |
saxpy (int n, float a, float *x, float *y) |
| 15 |
{ |
| 16 |
int i; |
| 17 |
|
| 18 |
for (i = 0; i < n; i++) |
| 19 |
{ |
| 20 |
y[i] = a * x[i] + y[i]; |
| 21 |
} |
| 22 |
} |
| 23 |
|
| 24 |
void |
| 25 |
context_check (CUcontext ctx1) |
| 26 |
{ |
| 27 |
CUcontext ctx2, ctx3; |
| 28 |
CUresult r; |
| 29 |
|
| 30 |
r = cuCtxGetCurrent (&ctx2); |
| 31 |
if (r != CUDA_SUCCESS) |
| 32 |
{ |
| 33 |
fprintf (stderr, "cuCtxGetCurrent failed: %d\n", r); |
| 34 |
exit (EXIT_FAILURE); |
| 35 |
} |
| 36 |
|
| 37 |
if (ctx1 != ctx2) |
| 38 |
{ |
| 39 |
fprintf (stderr, "new context established\n"); |
| 40 |
exit (EXIT_FAILURE); |
| 41 |
} |
| 42 |
|
| 43 |
ctx3 = (CUcontext) acc_get_current_cuda_context (); |
| 44 |
|
| 45 |
if (ctx1 != ctx3) |
| 46 |
{ |
| 47 |
fprintf (stderr, "acc_get_current_cuda_context returned wrong value\n"); |
| 48 |
exit (EXIT_FAILURE); |
| 49 |
} |
| 50 |
|
| 51 |
return; |
| 52 |
} |
| 53 |
|
| 54 |
int |
| 55 |
main (int argc, char **argv) |
| 56 |
{ |
| 57 |
cublasStatus_t s; |
| 58 |
cublasHandle_t h; |
| 59 |
CUcontext pctx; |
| 60 |
CUresult r; |
| 61 |
int i; |
| 62 |
const int N = 256; |
| 63 |
float *h_X, *h_Y1, *h_Y2; |
| 64 |
float *d_X,*d_Y; |
| 65 |
float alpha = 2.0f; |
| 66 |
float error_norm; |
| 67 |
float ref_norm; |
| 68 |
|
| 69 |
/* Test 4 - OpenACC creates, cuBLAS shares. */ |
| 70 |
|
| 71 |
acc_set_device_num (0, acc_device_nvidia); |
| 72 |
|
| 73 |
r = cuCtxGetCurrent (&pctx); |
| 74 |
if (r != CUDA_SUCCESS) |
| 75 |
{ |
| 76 |
fprintf (stderr, "cuCtxGetCurrent failed: %d\n", r); |
| 77 |
exit (EXIT_FAILURE); |
| 78 |
} |
| 79 |
|
| 80 |
h_X = (float *) malloc (N * sizeof (float)); |
| 81 |
if (h_X == 0) |
| 82 |
{ |
| 83 |
fprintf (stderr, "malloc failed: for h_X\n"); |
| 84 |
exit (EXIT_FAILURE); |
| 85 |
} |
| 86 |
|
| 87 |
h_Y1 = (float *) malloc (N * sizeof (float)); |
| 88 |
if (h_Y1 == 0) |
| 89 |
{ |
| 90 |
fprintf (stderr, "malloc failed: for h_Y1\n"); |
| 91 |
exit (EXIT_FAILURE); |
| 92 |
} |
| 93 |
|
| 94 |
h_Y2 = (float *) malloc (N * sizeof (float)); |
| 95 |
if (h_Y2 == 0) |
| 96 |
{ |
| 97 |
fprintf (stderr, "malloc failed: for h_Y2\n"); |
| 98 |
exit (EXIT_FAILURE); |
| 99 |
} |
| 100 |
|
| 101 |
for (i = 0; i < N; i++) |
| 102 |
{ |
| 103 |
h_X[i] = rand () / (float) RAND_MAX; |
| 104 |
h_Y2[i] = h_Y1[i] = rand () / (float) RAND_MAX; |
| 105 |
} |
| 106 |
|
| 107 |
#pragma acc parallel copyin (h_X[0:N]), copy (h_Y2[0:N]) copy (alpha) |
| 108 |
{ |
| 109 |
int i; |
| 110 |
|
| 111 |
for (i = 0; i < N; i++) |
| 112 |
{ |
| 113 |
h_Y2[i] = alpha * h_X[i] + h_Y2[i]; |
| 114 |
} |
| 115 |
} |
| 116 |
|
| 117 |
r = cuCtxGetCurrent (&pctx); |
| 118 |
if (r != CUDA_SUCCESS) |
| 119 |
{ |
| 120 |
fprintf (stderr, "cuCtxGetCurrent failed: %d\n", r); |
| 121 |
exit (EXIT_FAILURE); |
| 122 |
} |
| 123 |
|
| 124 |
d_X = (float *) acc_copyin (&h_X[0], N * sizeof (float)); |
| 125 |
if (d_X == NULL) |
| 126 |
{ |
| 127 |
fprintf (stderr, "copyin error h_Y1\n"); |
| 128 |
exit (EXIT_FAILURE); |
| 129 |
} |
| 130 |
|
| 131 |
d_Y = (float *) acc_copyin (&h_Y1[0], N * sizeof (float)); |
| 132 |
if (d_Y == NULL) |
| 133 |
{ |
| 134 |
fprintf (stderr, "copyin error h_Y1\n"); |
| 135 |
exit (EXIT_FAILURE); |
| 136 |
} |
| 137 |
|
| 138 |
s = cublasCreate (&h); |
| 139 |
if (s != CUBLAS_STATUS_SUCCESS) |
| 140 |
{ |
| 141 |
fprintf (stderr, "cublasCreate failed: %d\n", s); |
| 142 |
exit (EXIT_FAILURE); |
| 143 |
} |
| 144 |
|
| 145 |
context_check (pctx); |
| 146 |
|
| 147 |
s = cublasSaxpy (h, N, &alpha, d_X, 1, d_Y, 1); |
| 148 |
if (s != CUBLAS_STATUS_SUCCESS) |
| 149 |
{ |
| 150 |
fprintf (stderr, "cublasSaxpy failed: %d\n", s); |
| 151 |
exit (EXIT_FAILURE); |
| 152 |
} |
| 153 |
|
| 154 |
context_check (pctx); |
| 155 |
|
| 156 |
acc_memcpy_from_device (&h_Y1[0], d_Y, N * sizeof (float)); |
| 157 |
|
| 158 |
context_check (pctx); |
| 159 |
|
| 160 |
error_norm = 0; |
| 161 |
ref_norm = 0; |
| 162 |
|
| 163 |
for (i = 0; i < N; ++i) |
| 164 |
{ |
| 165 |
float diff; |
| 166 |
|
| 167 |
diff = h_Y1[i] - h_Y2[i]; |
| 168 |
error_norm += diff * diff; |
| 169 |
ref_norm += h_Y2[i] * h_Y2[i]; |
| 170 |
} |
| 171 |
|
| 172 |
error_norm = (float) sqrt ((double) error_norm); |
| 173 |
ref_norm = (float) sqrt ((double) ref_norm); |
| 174 |
|
| 175 |
if ((fabs (ref_norm) < 1e-7) || ((error_norm / ref_norm) >= 1e-6f)) |
| 176 |
{ |
| 177 |
fprintf (stderr, "math error\n"); |
| 178 |
exit (EXIT_FAILURE); |
| 179 |
} |
| 180 |
|
| 181 |
acc_delete (&h_X[0], N * sizeof (float)); |
| 182 |
acc_delete (&h_Y1[0], N * sizeof (float)); |
| 183 |
|
| 184 |
free (h_X); |
| 185 |
free (h_Y1); |
| 186 |
free (h_Y2); |
| 187 |
|
| 188 |
context_check (pctx); |
| 189 |
|
| 190 |
s = cublasDestroy (h); |
| 191 |
if (s != CUBLAS_STATUS_SUCCESS) |
| 192 |
{ |
| 193 |
fprintf (stderr, "cublasDestroy failed: %d\n", s); |
| 194 |
exit (EXIT_FAILURE); |
| 195 |
} |
| 196 |
|
| 197 |
context_check (pctx); |
| 198 |
|
| 199 |
acc_shutdown (acc_device_nvidia); |
| 200 |
|
| 201 |
r = cuCtxGetCurrent (&pctx); |
| 202 |
if (r != CUDA_SUCCESS) |
| 203 |
{ |
| 204 |
fprintf (stderr, "cuCtxGetCurrent failed: %d\n", r); |
| 205 |
exit (EXIT_FAILURE); |
| 206 |
} |
| 207 |
|
| 208 |
if (pctx) |
| 209 |
{ |
| 210 |
fprintf (stderr, "Unexpected context\n"); |
| 211 |
exit (EXIT_FAILURE); |
| 212 |
} |
| 213 |
|
| 214 |
return EXIT_SUCCESS; |
| 215 |
} |