root / ETSN / MySteps_6_openmp.c @ 301
Historique | Voir | Annoter | Télécharger (3,18 ko)
1 |
/* Simple SillySum function in C and OpenMP/C */
|
---|---|
2 |
/* compilation with sequential compute : gcc -fopenmp -O3 -o MySteps_6_openmp MySteps_6_openmp.c -lm -lgomp */
|
3 |
/* compilation without sequential compute : gcc -DNOSERIAL -fopenmp -O3 -o MySteps_6_openmp_NoSerial MySteps_6_openmp.c -lm -lgomp */
|
4 |
|
5 |
#include <math.h> |
6 |
#include <stdio.h> |
7 |
#include <stdlib.h> |
8 |
#include <sys/time.h> |
9 |
|
10 |
#define PI 3.141592653589793 |
11 |
|
12 |
#define MYFLOAT float |
13 |
|
14 |
MYFLOAT MySillyFunction(MYFLOAT x) |
15 |
{ |
16 |
return(pow(sqrt(log(exp(atanh(tanh(asinh(sinh(acosh(cosh(atan(tan(asin(sin(acos(cos(x))))))))))))))),2)); |
17 |
} |
18 |
|
19 |
void MySillySum(MYFLOAT *res, MYFLOAT *a, MYFLOAT *b,int calls, int size) |
20 |
{ |
21 |
for (uint i=0; i<size;i++) |
22 |
{ |
23 |
MYFLOAT ai=a[i]; |
24 |
MYFLOAT bi=b[i]; |
25 |
|
26 |
for (int c=0;c<calls;c++) |
27 |
{ |
28 |
ai=MySillyFunction(ai); |
29 |
bi=MySillyFunction(bi); |
30 |
} |
31 |
|
32 |
res[i] = ai + bi; |
33 |
} |
34 |
} |
35 |
|
36 |
void MySillySumOMP(MYFLOAT *res, MYFLOAT *a, MYFLOAT *b,int calls, int size) |
37 |
{ |
38 |
#pragma omp parallel for |
39 |
for (uint i=0; i<size;i++) |
40 |
{ |
41 |
MYFLOAT ai=a[i]; |
42 |
MYFLOAT bi=b[i]; |
43 |
|
44 |
for (int c=0;c<calls;c++) |
45 |
{ |
46 |
ai=MySillyFunction(ai); |
47 |
bi=MySillyFunction(bi); |
48 |
} |
49 |
|
50 |
res[i] = ai + bi; |
51 |
} |
52 |
} |
53 |
|
54 |
MYFLOAT MyNorm(MYFLOAT *a,MYFLOAT *b,int size)
|
55 |
{ |
56 |
MYFLOAT norm=0.;
|
57 |
|
58 |
for (int i=0;i<size;i++) |
59 |
{ |
60 |
norm+=pow(a[i]-b[i],2);
|
61 |
} |
62 |
|
63 |
return(sqrt(norm));
|
64 |
} |
65 |
|
66 |
void MyPrint(MYFLOAT *a,int size) |
67 |
{ |
68 |
printf("[");
|
69 |
for (int i=0;i<size;i++) |
70 |
{ |
71 |
printf(" %.8e ",a[i]);
|
72 |
} |
73 |
printf("]\n");
|
74 |
} |
75 |
|
76 |
int main(int argc,char *argv[]) |
77 |
{ |
78 |
float *a,*b,*res,*resacc;
|
79 |
int size=1024; |
80 |
int calls=1; |
81 |
struct timeval tv1,tv2;
|
82 |
|
83 |
if (argc > 1) { |
84 |
size=(int)atoll(argv[1]); |
85 |
calls=(int)atoll(argv[2]); |
86 |
} |
87 |
else {
|
88 |
printf("\n\tPi : Estimate SillySum\n\n\t\t#1 : size (default 1024)\n\t\t#2 : calls (default 1)\n\n");
|
89 |
} |
90 |
|
91 |
printf("%i %i\n",size,calls);
|
92 |
|
93 |
a=(float*)malloc(size*sizeof(MYFLOAT)); |
94 |
b=(float*)malloc(size*sizeof(MYFLOAT)); |
95 |
res=(float*)malloc(size*sizeof(MYFLOAT)); |
96 |
resacc=(float*)malloc(size*sizeof(MYFLOAT)); |
97 |
|
98 |
srand(110271);
|
99 |
|
100 |
for (int i=0;i<size;i++) |
101 |
{ |
102 |
a[i]=(MYFLOAT)rand()/(MYFLOAT)RAND_MAX; |
103 |
b[i]=(MYFLOAT)rand()/(MYFLOAT)RAND_MAX; |
104 |
res[i]=0.;
|
105 |
resacc[i]=0.;
|
106 |
} |
107 |
|
108 |
#ifndef NOSERIAL
|
109 |
gettimeofday(&tv1, NULL);
|
110 |
MySillySum(res,a,b,calls,size); |
111 |
gettimeofday(&tv2, NULL);
|
112 |
#endif
|
113 |
|
114 |
MYFLOAT elapsed=(MYFLOAT)((tv2.tv_sec-tv1.tv_sec) * 1000000L +
|
115 |
(tv2.tv_usec-tv1.tv_usec))/1000000;
|
116 |
|
117 |
gettimeofday(&tv1, NULL);
|
118 |
MySillySumOMP(resacc,a,b,calls,size); |
119 |
gettimeofday(&tv2, NULL);
|
120 |
|
121 |
MYFLOAT elapsedAcc=(MYFLOAT)((tv2.tv_sec-tv1.tv_sec) * 1000000L +
|
122 |
(tv2.tv_usec-tv1.tv_usec))/1000000;
|
123 |
|
124 |
#ifndef NOSERIAL
|
125 |
MYFLOAT MyChecker=MyNorm(res,resacc,size); |
126 |
printf("Norm: %.8e\n",MyChecker);
|
127 |
#endif
|
128 |
|
129 |
#ifdef VERBOSE
|
130 |
MyPrint(res,size); |
131 |
MyPrint(resacc,size); |
132 |
#endif
|
133 |
|
134 |
#ifndef NOSERIAL
|
135 |
printf("Elapsed Time: %.3f\n",elapsed);
|
136 |
#endif
|
137 |
|
138 |
printf("OMP Elapsed Time: %.3f\n",elapsedAcc);
|
139 |
|
140 |
#ifndef NOSERIAL
|
141 |
printf("NativeRate: %.lld\n",(unsigned long)((float)size/elapsed)); |
142 |
#endif
|
143 |
printf("OMPRate: %.lld\n",(unsigned long)((float)size/elapsedAcc)); |
144 |
|
145 |
#ifndef NOSERIAL
|
146 |
printf("AccRatio: %.3f\n",elapsed/elapsedAcc);
|
147 |
#endif
|
148 |
|
149 |
free(a); |
150 |
free(b); |
151 |
free(res); |
152 |
free(resacc); |
153 |
} |
154 |
|