Statistiques
| Révision :

root / ETSN / MySteps_1_openacc.c @ 296

Historique | Voir | Annoter | Télécharger (2,99 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
/* #pragma acc routine */
15 296 equemene
/* MYFLOAT MySillyFunction(MYFLOAT x) */
16 296 equemene
/* { */
17 296 equemene
/*     return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2));  */
18 296 equemene
/* } */
19 296 equemene
20 296 equemene
void MySum(MYFLOAT *res, MYFLOAT *a, MYFLOAT *b, int size)
21 296 equemene
{
22 296 equemene
  for (uint i=0; i<size;i++)
23 296 equemene
    {
24 296 equemene
      res[i] = a[i] + b[i];
25 296 equemene
    }
26 296 equemene
}
27 296 equemene
28 296 equemene
void MySumOpenACC(MYFLOAT *res, MYFLOAT *a, MYFLOAT *b, int size)
29 296 equemene
{
30 296 equemene
  #pragma acc data copyin(a[0:size],b[0:size]),copyout(res[0:size])
31 296 equemene
  #pragma acc parallel loop
32 296 equemene
  for (uint i=0; i<size;i++)
33 296 equemene
    {
34 296 equemene
      res[i] = a[i] + b[i];
35 296 equemene
    }
36 296 equemene
}
37 296 equemene
38 296 equemene
MYFLOAT MyNorm(MYFLOAT *a,MYFLOAT *b,int size)
39 296 equemene
{
40 296 equemene
  MYFLOAT norm=0.;
41 296 equemene
42 296 equemene
  for (int i=0;i<size;i++)
43 296 equemene
    {
44 296 equemene
      norm+=pow(a[i]-b[i],2);
45 296 equemene
    }
46 296 equemene
47 296 equemene
  return(sqrt(norm));
48 296 equemene
}
49 296 equemene
50 296 equemene
void MyPrint(MYFLOAT *a,int size)
51 296 equemene
{
52 296 equemene
  printf("[");
53 296 equemene
  for (int i=0;i<size;i++)
54 296 equemene
    {
55 296 equemene
      printf(" %.8e ",a[i]);
56 296 equemene
    }
57 296 equemene
  printf("]\n");
58 296 equemene
}
59 296 equemene
60 296 equemene
int main(int argc,char *argv[])
61 296 equemene
{
62 296 equemene
  float *a,*b,*res,*resacc;
63 296 equemene
  int size=1024;
64 296 equemene
  struct timeval tv1,tv2;
65 296 equemene
66 296 equemene
  if (argc > 1) {
67 296 equemene
    size=(int)atoll(argv[1]);
68 296 equemene
  }
69 296 equemene
  else {
70 296 equemene
    printf("\n\tPi : Estimate SillySum\n\n\t\t#1 : size (default 1024)\n\n");
71 296 equemene
  }
72 296 equemene
73 296 equemene
  printf("%i\n",size);
74 296 equemene
75 296 equemene
  a=(float*)malloc(size*sizeof(MYFLOAT));
76 296 equemene
  b=(float*)malloc(size*sizeof(MYFLOAT));
77 296 equemene
  res=(float*)malloc(size*sizeof(MYFLOAT));
78 296 equemene
  resacc=(float*)malloc(size*sizeof(MYFLOAT));
79 296 equemene
80 296 equemene
  srand(110271);
81 296 equemene
82 296 equemene
  for (int i=0;i<size;i++)
83 296 equemene
    {
84 296 equemene
      a[i]=(MYFLOAT)rand()/(MYFLOAT)RAND_MAX;
85 296 equemene
      b[i]=(MYFLOAT)rand()/(MYFLOAT)RAND_MAX;
86 296 equemene
      res[i]=0.;
87 296 equemene
      resacc[i]=0.;
88 296 equemene
    }
89 296 equemene
90 296 equemene
#ifndef NOSERIAL
91 296 equemene
  gettimeofday(&tv1, NULL);
92 296 equemene
  MySum(res,a,b,size);
93 296 equemene
  gettimeofday(&tv2, NULL);
94 296 equemene
#endif
95 296 equemene
96 296 equemene
  MYFLOAT elapsed=(MYFLOAT)((tv2.tv_sec-tv1.tv_sec) * 1000000L +
97 296 equemene
                            (tv2.tv_usec-tv1.tv_usec))/1000000;
98 296 equemene
99 296 equemene
  gettimeofday(&tv1, NULL);
100 296 equemene
  MySumOpenACC(resacc,a,b,size);
101 296 equemene
  gettimeofday(&tv2, NULL);
102 296 equemene
103 296 equemene
  MYFLOAT elapsedAcc=(MYFLOAT)((tv2.tv_sec-tv1.tv_sec) * 1000000L +
104 296 equemene
                               (tv2.tv_usec-tv1.tv_usec))/1000000;
105 296 equemene
106 296 equemene
#ifndef NOSERIAL
107 296 equemene
  MYFLOAT MyChecker=MyNorm(res,resacc,size);
108 296 equemene
  printf("Norm: %.8e\n",MyChecker);
109 296 equemene
#endif
110 296 equemene
111 296 equemene
#ifdef VERBOSE
112 296 equemene
  MyPrint(res,size);
113 296 equemene
  MyPrint(resacc,size);
114 296 equemene
#endif
115 296 equemene
116 296 equemene
#ifndef NOSERIAL
117 296 equemene
  printf("Elapsed Time: %.3f\n",elapsed);
118 296 equemene
  printf("OpenACC Elapsed Time: %.3f\n",elapsedAcc);
119 296 equemene
#endif
120 296 equemene
121 296 equemene
#ifndef NOSERIAL
122 296 equemene
  printf("NaiveRate: %.lld\n",(unsigned long)((float)size/elapsed));
123 296 equemene
#endif
124 296 equemene
  printf("OpenACCRate: %.lld\n",(unsigned long)((float)size/elapsedAcc));
125 296 equemene
126 296 equemene
#ifndef NOSERIAL
127 296 equemene
  printf("OpenACCRatio: %.3f\n",elapsed/elapsedAcc);
128 296 equemene
#endif
129 296 equemene
130 296 equemene
  free(a);
131 296 equemene
  free(b);
132 296 equemene
  free(res);
133 296 equemene
  free(resacc);
134 296 equemene
}