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
saxpy(int n,float a,float * x,float * y)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
context_check(CUcontext ctx1)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
main(int argc,char ** argv)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 3 - 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 d_X = (float *) acc_copyin (&h_X[0], N * sizeof (float));
108 if (d_X == NULL)
109 {
110 fprintf (stderr, "copyin error h_X\n");
111 exit (EXIT_FAILURE);
112 }
113
114 d_Y = (float *) acc_copyin (&h_Y1[0], N * sizeof (float));
115 if (d_Y == NULL)
116 {
117 fprintf (stderr, "copyin error h_Y1\n");
118 exit (EXIT_FAILURE);
119 }
120
121 context_check (pctx);
122
123 s = cublasCreate (&h);
124 if (s != CUBLAS_STATUS_SUCCESS)
125 {
126 fprintf (stderr, "cublasCreate failed: %d\n", s);
127 exit (EXIT_FAILURE);
128 }
129
130 context_check (pctx);
131
132 s = cublasSaxpy (h, N, &alpha, d_X, 1, d_Y, 1);
133 if (s != CUBLAS_STATUS_SUCCESS)
134 {
135 fprintf (stderr, "cublasSaxpy failed: %d\n", s);
136 exit (EXIT_FAILURE);
137 }
138
139 context_check (pctx);
140
141 acc_memcpy_from_device (&h_Y1[0], d_Y, N * sizeof (float));
142
143 context_check (pctx);
144
145 saxpy (N, alpha, h_X, h_Y2);
146
147 error_norm = 0;
148 ref_norm = 0;
149
150 for (i = 0; i < N; ++i)
151 {
152 float diff;
153
154 diff = h_Y1[i] - h_Y2[i];
155 error_norm += diff * diff;
156 ref_norm += h_Y2[i] * h_Y2[i];
157 }
158
159 error_norm = (float) sqrt ((double) error_norm);
160 ref_norm = (float) sqrt ((double) ref_norm);
161
162 if ((fabs (ref_norm) < 1e-7) || ((error_norm / ref_norm) >= 1e-6f))
163 {
164 fprintf (stderr, "math error\n");
165 exit (EXIT_FAILURE);
166 }
167
168 acc_delete (&h_X[0], N * sizeof (float));
169 acc_delete (&h_Y1[0], N * sizeof (float));
170
171 free (h_X);
172 free (h_Y1);
173 free (h_Y2);
174
175 context_check (pctx);
176
177 s = cublasDestroy (h);
178 if (s != CUBLAS_STATUS_SUCCESS)
179 {
180 fprintf (stderr, "cublasDestroy failed: %d\n", s);
181 exit (EXIT_FAILURE);
182 }
183
184 context_check (pctx);
185
186 acc_shutdown (acc_device_nvidia);
187
188 r = cuCtxGetCurrent (&pctx);
189 if (r != CUDA_SUCCESS)
190 {
191 fprintf (stderr, "cuCtxGetCurrent failed: %d\n", r);
192 exit (EXIT_FAILURE);
193 }
194
195 if (pctx)
196 {
197 fprintf (stderr, "Unexpected context\n");
198 exit (EXIT_FAILURE);
199 }
200
201 return EXIT_SUCCESS;
202 }
203