Statistiques
| Révision :

root / ETSN / MySteps_1_openacc.c @ 299

Historique | Voir | Annoter | Télécharger (2,79 ko)

1 296 equemene
/* Simple Sum function in C and OpenACC/C */
2 296 equemene
/* compilation with sequential compute : gcc -O3 -fopenacc -foffload=nvptx-none -foffload="-O3 -misa=sm_35 -lm" -o MySteps_1_openacc MySteps_1_openacc.c -lm */
3 296 equemene
/* compilation without sequential compute : gcc -DNOSERIAL -O3 -fopenacc -foffload=nvptx-none -foffload="-O3 -misa=sm_35 -lm" -o MySteps_1_openacc_NoSerial MySteps_1_openacc.c -lm */
4 296 equemene
5 296 equemene
#include <math.h>
6 296 equemene
#include <stdio.h>
7 296 equemene
#include <stdlib.h>
8 296 equemene
#include <sys/time.h>
9 296 equemene
10 296 equemene
#define PI 3.141592653589793
11 296 equemene
12 296 equemene
#define MYFLOAT float
13 296 equemene
14 296 equemene
void MySum(MYFLOAT *res, MYFLOAT *a, MYFLOAT *b, int size)
15 296 equemene
{
16 296 equemene
  for (uint i=0; i<size;i++)
17 296 equemene
    {
18 296 equemene
      res[i] = a[i] + b[i];
19 296 equemene
    }
20 296 equemene
}
21 296 equemene
22 296 equemene
void MySumOpenACC(MYFLOAT *res, MYFLOAT *a, MYFLOAT *b, int size)
23 296 equemene
{
24 296 equemene
  #pragma acc data copyin(a[0:size],b[0:size]),copyout(res[0:size])
25 296 equemene
  #pragma acc parallel loop
26 296 equemene
  for (uint i=0; i<size;i++)
27 296 equemene
    {
28 296 equemene
      res[i] = a[i] + b[i];
29 296 equemene
    }
30 296 equemene
}
31 296 equemene
32 296 equemene
MYFLOAT MyNorm(MYFLOAT *a,MYFLOAT *b,int size)
33 296 equemene
{
34 296 equemene
  MYFLOAT norm=0.;
35 296 equemene
36 296 equemene
  for (int i=0;i<size;i++)
37 296 equemene
    {
38 296 equemene
      norm+=pow(a[i]-b[i],2);
39 296 equemene
    }
40 296 equemene
41 296 equemene
  return(sqrt(norm));
42 296 equemene
}
43 296 equemene
44 296 equemene
void MyPrint(MYFLOAT *a,int size)
45 296 equemene
{
46 296 equemene
  printf("[");
47 296 equemene
  for (int i=0;i<size;i++)
48 296 equemene
    {
49 296 equemene
      printf(" %.8e ",a[i]);
50 296 equemene
    }
51 296 equemene
  printf("]\n");
52 296 equemene
}
53 296 equemene
54 296 equemene
int main(int argc,char *argv[])
55 296 equemene
{
56 296 equemene
  float *a,*b,*res,*resacc;
57 296 equemene
  int size=1024;
58 296 equemene
  struct timeval tv1,tv2;
59 296 equemene
60 296 equemene
  if (argc > 1) {
61 296 equemene
    size=(int)atoll(argv[1]);
62 296 equemene
  }
63 296 equemene
  else {
64 296 equemene
    printf("\n\tPi : Estimate SillySum\n\n\t\t#1 : size (default 1024)\n\n");
65 296 equemene
  }
66 296 equemene
67 296 equemene
  printf("%i\n",size);
68 296 equemene
69 296 equemene
  a=(float*)malloc(size*sizeof(MYFLOAT));
70 296 equemene
  b=(float*)malloc(size*sizeof(MYFLOAT));
71 296 equemene
  res=(float*)malloc(size*sizeof(MYFLOAT));
72 296 equemene
  resacc=(float*)malloc(size*sizeof(MYFLOAT));
73 296 equemene
74 296 equemene
  srand(110271);
75 296 equemene
76 296 equemene
  for (int i=0;i<size;i++)
77 296 equemene
    {
78 296 equemene
      a[i]=(MYFLOAT)rand()/(MYFLOAT)RAND_MAX;
79 296 equemene
      b[i]=(MYFLOAT)rand()/(MYFLOAT)RAND_MAX;
80 296 equemene
      res[i]=0.;
81 296 equemene
      resacc[i]=0.;
82 296 equemene
    }
83 296 equemene
84 296 equemene
#ifndef NOSERIAL
85 296 equemene
  gettimeofday(&tv1, NULL);
86 296 equemene
  MySum(res,a,b,size);
87 296 equemene
  gettimeofday(&tv2, NULL);
88 296 equemene
#endif
89 296 equemene
90 296 equemene
  MYFLOAT elapsed=(MYFLOAT)((tv2.tv_sec-tv1.tv_sec) * 1000000L +
91 296 equemene
                            (tv2.tv_usec-tv1.tv_usec))/1000000;
92 296 equemene
93 296 equemene
  gettimeofday(&tv1, NULL);
94 296 equemene
  MySumOpenACC(resacc,a,b,size);
95 296 equemene
  gettimeofday(&tv2, NULL);
96 296 equemene
97 296 equemene
  MYFLOAT elapsedAcc=(MYFLOAT)((tv2.tv_sec-tv1.tv_sec) * 1000000L +
98 296 equemene
                               (tv2.tv_usec-tv1.tv_usec))/1000000;
99 296 equemene
100 296 equemene
#ifndef NOSERIAL
101 296 equemene
  MYFLOAT MyChecker=MyNorm(res,resacc,size);
102 296 equemene
  printf("Norm: %.8e\n",MyChecker);
103 296 equemene
#endif
104 296 equemene
105 296 equemene
#ifdef VERBOSE
106 296 equemene
  MyPrint(res,size);
107 296 equemene
  MyPrint(resacc,size);
108 296 equemene
#endif
109 296 equemene
110 296 equemene
#ifndef NOSERIAL
111 296 equemene
  printf("Elapsed Time: %.3f\n",elapsed);
112 296 equemene
  printf("OpenACC Elapsed Time: %.3f\n",elapsedAcc);
113 296 equemene
#endif
114 296 equemene
115 296 equemene
#ifndef NOSERIAL
116 296 equemene
  printf("NaiveRate: %.lld\n",(unsigned long)((float)size/elapsed));
117 296 equemene
#endif
118 296 equemene
  printf("OpenACCRate: %.lld\n",(unsigned long)((float)size/elapsedAcc));
119 296 equemene
120 296 equemene
#ifndef NOSERIAL
121 296 equemene
  printf("OpenACCRatio: %.3f\n",elapsed/elapsedAcc);
122 296 equemene
#endif
123 296 equemene
124 296 equemene
  free(a);
125 296 equemene
  free(b);
126 296 equemene
  free(res);
127 296 equemene
  free(resacc);
128 296 equemene
}