1 |
1 |
//
|
2 |
2 |
// Estimation of Pi using Monte Carlo exploration process
|
3 |
|
// gcc -std=c99 -O3 -o Pi Pi.c -lm
|
|
3 |
// Exploit OpenMP on Xeon Phi
|
|
4 |
// source /opt/intel/bin/compilervars.sh intel64
|
|
5 |
// icpc -std=c99 -O3 -o Pi Pi.c -lm
|
4 |
6 |
//
|
5 |
7 |
|
6 |
8 |
#include <math.h>
|
... | ... | |
19 |
21 |
|
20 |
22 |
#define MWCfp MWC * 2.328306435454494e-10f
|
21 |
23 |
#define KISSfp KISS * 2.328306435454494e-10f
|
|
24 |
#define SHR3fp SHR3 * 2.328306435454494e-10f
|
|
25 |
#define CONGfp CONG * 2.328306435454494e-10f
|
22 |
26 |
|
23 |
27 |
#define ITERATIONS 1000000000
|
24 |
28 |
|
... | ... | |
35 |
39 |
|
36 |
40 |
LENGTH MainLoopGlobal(LENGTH iterations,unsigned int seed_w,unsigned int seed_z)
|
37 |
41 |
{
|
|
42 |
#if defined TCONG
|
|
43 |
unsigned int jcong=seed_z;
|
|
44 |
#elif defined TSHR3
|
|
45 |
unsigned int jsr=seed_w;
|
|
46 |
#elif defined TMWC
|
38 |
47 |
unsigned int z=seed_z;
|
39 |
48 |
unsigned int w=seed_w;
|
40 |
|
|
|
49 |
#elif defined TKISS
|
|
50 |
unsigned int jcong=seed_z;
|
|
51 |
unsigned int jsr=seed_w;
|
|
52 |
unsigned int z=seed_z;
|
|
53 |
unsigned int w=seed_w;
|
|
54 |
#endif
|
|
55 |
|
41 |
56 |
LENGTH total=0;
|
42 |
57 |
|
43 |
58 |
for (LENGTH i=0;i<iterations;i++) {
|
44 |
59 |
|
45 |
|
float x=MWCfp ;
|
46 |
|
float y=MWCfp ;
|
|
60 |
#if defined TINT32
|
|
61 |
#define THEONE 1073741824
|
|
62 |
#if defined TCONG
|
|
63 |
unsigned int x=CONG>>17 ;
|
|
64 |
unsigned int y=CONG>>17 ;
|
|
65 |
#elif defined TSHR3
|
|
66 |
unsigned int x=SHR3>>17 ;
|
|
67 |
unsigned int y=SHR3>>17 ;
|
|
68 |
#elif defined TMWC
|
|
69 |
unsigned int x=MWC>>17 ;
|
|
70 |
unsigned int y=MWC>>17 ;
|
|
71 |
#elif defined TKISS
|
|
72 |
unsigned int x=KISS>>17 ;
|
|
73 |
unsigned int y=KISS>>17 ;
|
|
74 |
#endif
|
|
75 |
#elif defined TINT64
|
|
76 |
#define THEONE 4611686018427387904
|
|
77 |
#if defined TCONG
|
|
78 |
unsigned long x=(unsigned long)(CONG>>1) ;
|
|
79 |
unsigned long y=(unsigned long)(CONG>>1) ;
|
|
80 |
#elif defined TSHR3
|
|
81 |
unsigned long x=(unsigned long)(SHR3>>1) ;
|
|
82 |
unsigned long y=(unsigned long)(SHR3>>1) ;
|
|
83 |
#elif defined TMWC
|
|
84 |
unsigned long x=(unsigned long)(MWC>>1) ;
|
|
85 |
unsigned long y=(unsigned long)(MWC>>1) ;
|
|
86 |
#elif defined TKISS
|
|
87 |
unsigned long x=(unsigned long)(KISS>>1) ;
|
|
88 |
unsigned long y=(unsigned long)(KISS>>1) ;
|
|
89 |
#endif
|
|
90 |
#elif defined TFP32
|
|
91 |
#define THEONE 1.0f
|
|
92 |
#if defined TCONG
|
|
93 |
float x=CONGfp ;
|
|
94 |
float y=CONGfp ;
|
|
95 |
#elif defined TSHR3
|
|
96 |
float x=SHR3fp ;
|
|
97 |
float y=SHR3fp ;
|
|
98 |
#elif defined TMWC
|
|
99 |
float x=MWCfp ;
|
|
100 |
float y=MWCfp ;
|
|
101 |
#elif defined TKISS
|
|
102 |
float x=KISSfp ;
|
|
103 |
float y=KISSfp ;
|
|
104 |
#endif
|
|
105 |
#elif defined TFP64
|
|
106 |
#define THEONE 1.0f
|
|
107 |
#if defined TCONG
|
|
108 |
double x=(double)CONGfp ;
|
|
109 |
double y=(double)CONGfp ;
|
|
110 |
#elif defined TSHR3
|
|
111 |
double x=(double)SHR3fp ;
|
|
112 |
double y=(double)SHR3fp ;
|
|
113 |
#elif defined TMWC
|
|
114 |
double x=(double)MWCfp ;
|
|
115 |
double y=(double)MWCfp ;
|
|
116 |
#elif defined TKISS
|
|
117 |
double x=(double)KISSfp ;
|
|
118 |
double y=(double)KISSfp ;
|
|
119 |
#endif
|
|
120 |
#endif
|
47 |
121 |
|
48 |
122 |
// Matching test
|
49 |
|
int inside=((x*x+y*y) < 1.0f) ? 1:0;
|
|
123 |
unsigned long inside=((x*x+y*y) < THEONE) ? 1:0;
|
50 |
124 |
total+=inside;
|
|
125 |
|
51 |
126 |
}
|
52 |
127 |
|
53 |
128 |
return(total);
|
... | ... | |
55 |
130 |
|
56 |
131 |
LENGTH splitter(int process,int seed_w,int seed_z,LENGTH iterations) {
|
57 |
132 |
|
58 |
|
LENGTH inside[1024],insides=0;
|
|
133 |
LENGTH inside[4096],insides=0;
|
59 |
134 |
int i;
|
60 |
135 |
|
61 |
136 |
#pragma omp target device(0)
|
62 |
137 |
#pragma omp teams num_teams(60) thread_limit(4)
|
|
138 |
// #pragma omp parallel for
|
63 |
139 |
#pragma omp distribute
|
64 |
140 |
for (int i=0 ; i<process; i++) {
|
65 |
|
inside[i]=MainLoopGlobal(iterations/process,seed_w+process,seed_z+process);
|
66 |
|
// printf("\tFound %lld for process %i\n",(long long)inside[i],i);
|
|
141 |
inside[i]=MainLoopGlobal(iterations/process,seed_w+i,seed_z+i);
|
67 |
142 |
}
|
68 |
|
//printf("\n");
|
69 |
143 |
|
70 |
144 |
for (int i=0 ; i<process; i++) {
|
|
145 |
printf("\tFound %lld for process %i\n",(long long)inside[i],i);
|
71 |
146 |
insides+=inside[i];
|
72 |
147 |
}
|
|
148 |
printf("\n");
|
73 |
149 |
|
74 |
150 |
return(insides);
|
75 |
151 |
}
|
... | ... | |
78 |
154 |
|
79 |
155 |
unsigned int seed_w=10,seed_z=10,process=PROCESS;
|
80 |
156 |
LENGTH iterations=ITERATIONS;
|
81 |
|
LENGTH inside[1024],insides=0;
|
|
157 |
LENGTH insides=0;
|
82 |
158 |
|
83 |
159 |
if (argc > 1) {
|
84 |
160 |
iterations=(LENGTH)atoll(argv[1]);
|