annotate example/fft/main.cc @ 2014:8c618e912c88 draft

optimization data transfer. wrong result
author Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
date Tue, 01 Jul 2014 17:04:01 +0900
parents d43c2b7932ea
children 1d7d1e398833
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
1 #include <stdio.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
2 #include <stdlib.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
3 #include <math.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
4 #include <sys/stat.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
5 #include <fcntl.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
6 #include <sys/time.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
7 #include "TaskManager.h"
1727
e99dc86d39e2 minor fix
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1717
diff changeset
8 #include "GpuScheduler.h"
1560
3df1868130cb fix fft ppe example
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1555
diff changeset
9 #include "SchedTask.h"
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
10 #include "Func.h"
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
11 #ifdef __APPLE__
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
12 #include <OpenCL/opencl.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
13 #else
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
14 #include <CL/cl.h>
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
15 #endif
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
16 #include "pgm.h"
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
17 extern void task_init();
1815
9de57d13a386 ChangeLog
kkb
parents: 1779
diff changeset
18 #ifdef GPU
1727
e99dc86d39e2 minor fix
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1717
diff changeset
19 extern void gpu_task_init();
1779
0dd39c19dbb2 minor change
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1752
diff changeset
20 #endif
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
21 #define PI 3.14159265358979
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
22
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
23 #define MAX_SOURCE_SIZE (0x100000)
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
24
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
25 #define AMP(a, b) (sqrt((a)*(a)+(b)*(b)))
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
26
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
27 static double st_time;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
28 static double ed_time;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
29 void TMend(TaskManager *);
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
30 cl_device_id device_id = NULL;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
31 cl_context context = NULL;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
32 cl_command_queue queue = NULL;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
33 cl_program program = NULL;
1752
537707f376c8 fix profile
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1734
diff changeset
34 CPU_TYPE spe_cpu = SPE_ANY;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
35
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
36 cl_float2* xm;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
37 cl_float2* rm;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
38 cl_float2* wm;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
39 pgm_t ipgm;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
40
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
41 enum Mode {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
42 forward = 0,
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
43 inverse = 1
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
44 };
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
45
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
46 static double
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
47 getTime()
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
48 {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
49 struct timeval tv;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
50 gettimeofday(&tv, NULL);
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
51 return tv.tv_sec + (double)tv.tv_usec*1e-6;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
52 }
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
53
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
54 void
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
55 output()
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
56 {
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
57 int n = ipgm.width;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
58 float* ampd;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
59 ampd = (float*)malloc(n*n*sizeof(float));
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
60 for (int i=0; i < n; i++) {
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
61 for (int j=0; j < n; j++) {
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
62 ampd[n*((i))+((j))] = (AMP(((float*)xm)[(2*n*i)+2*j], ((float*)xm)[(2*n*i)+2*j+1]));
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
63 }
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
64 }
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
65 pgm_t opgm;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
66 opgm.width = n;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
67 opgm.height = n;
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
68 normalizeF2PGM(&opgm, ampd);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
69 free(ampd);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
70
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
71 // Write out image
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
72 writePGM(&opgm, "output.pgm");
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
73
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
74 // Finalizations
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
75 destroyPGM(&ipgm);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
76 destroyPGM(&opgm);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
77
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
78 free(wm);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
79 free(rm);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
80 free(xm);
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
81 }
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
82
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
83 const char *usr_help_str = "Usage: ./fft [option]\n \
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
84 options\n\
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
85 -cpu Number of SPE used (default 1)\n\
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
86 -l, --length Sorted number of data (default 1200)\n\
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
87 -h, --help Print this message";
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
88
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
89 int setWorkSize(size_t* gws, size_t* lws, cl_int x, cl_int y)
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
90 {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
91 switch(y) {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
92 case 1:
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
93 gws[0] = x;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
94 gws[1] = 1;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
95 lws[0] = 1;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
96 lws[1] = 1;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
97 break;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
98 default:
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
99 gws[0] = x;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
100 gws[1] = y;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
101 lws[0] = 1;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
102 lws[1] = 1;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
103 break;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
104 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
105
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
106 return 0;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
107 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
108
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
109 HTask*
1663
ce031df3dd32 fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1661
diff changeset
110 fftCore(TaskManager *manager,cl_float2 *dst, cl_float2 *src, cl_float2 *spin, long m, enum Mode direction,HTask* waitTask)
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
111 {
1673
2c3adce7eb40 fix fft on gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1669
diff changeset
112 long direction_flag;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
113 switch (direction) {
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
114 case forward:direction_flag = 0x00000000; break;
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
115 case inverse:direction_flag = 0x80000000; break;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
116 }
1663
ce031df3dd32 fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1661
diff changeset
117 long n = 1<<m;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
118 size_t gws[2],lws[2];
1633
fbb4757d82ee refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1627
diff changeset
119 int length_dst = n*n;
fbb4757d82ee refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1627
diff changeset
120 int length_src = n*n;
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
121
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
122 HTask* brev = manager->create_task(BIT_REVERSE);
1633
fbb4757d82ee refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1627
diff changeset
123 setWorkSize(gws,lws,n,n);
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
124 brev->set_param(0,m);
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
125 brev->set_param(1,n);
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
126 brev->set_inData(0, src, length_src*sizeof(cl_float2));
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
127 brev->set_outData(0, dst, length_dst*sizeof(cl_float2));
1975
4cf85b48ab9e running fft with CudaScheduler, but wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1971
diff changeset
128 brev->set_cpu(spe_cpu);
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
129 brev->flip();
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
130 brev->wait_for(waitTask);
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
131 brev->iterate(gws[0],gws[1]);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
132
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
133 waitTask = brev;
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
134
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
135 setWorkSize(gws,lws,n/2,n);
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
136 for(int iter=1;iter<=m;iter++) {
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
137 HTask* bfly = manager->create_task(BUTTERFLY);
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
138 bfly->set_param(0,n);
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
139 bfly->set_param(1,direction_flag);
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
140 bfly->set_param(2,(long)iter);
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
141 bfly->set_inData(0, dst, length_dst*sizeof(cl_float2));
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
142 bfly->set_inData(1, spin, sizeof(cl_float2)*(n/2));
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
143 bfly->set_outData(0,dst,length_dst*sizeof(cl_float2));
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
144 bfly->set_cpu(spe_cpu);
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
145 bfly->flip();
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
146 bfly->wait_for(waitTask);
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
147 bfly->iterate(gws[0],gws[1]);
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
148 waitTask = bfly;
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
149 }
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
150
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
151 if (direction == inverse) {
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
152 setWorkSize(gws,lws,n,n);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
153 HTask *norm = manager->create_task(NORMALIZATION);
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
154 norm->set_inData(0,dst,length_dst*sizeof(cl_float2));
1571
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1566
diff changeset
155 norm->set_outData(0, dst, length_dst*sizeof(cl_float2));
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
156 norm->set_param(0,n);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
157 norm->set_cpu(spe_cpu);
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
158 norm->wait_for(waitTask);
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
159 norm->iterate(gws[0],gws[1]);
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
160
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
161 waitTask = norm;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
162 }
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
163 return waitTask;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
164 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
165
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
166 char *
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
167 init(int argc, char**argv){
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
168
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
169 char *filename = 0;
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
170
1689
c595245aa59e fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1687
diff changeset
171 // printf("%s ",argv[4]);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
172 for (int i = 1; argv[i]; ++i) {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
173 if (strcmp(argv[i], "-file") == 0) {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
174 filename = argv[i+1];
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
175 } else if (strcmp(argv[i], "-g") == 0) {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
176 spe_cpu = GPU_0;
1702
f52904f8f03e fix example.run ANY_ANY
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1689
diff changeset
177 } else if (strcmp(argv[i], "-any") == 0) {
f52904f8f03e fix example.run ANY_ANY
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1689
diff changeset
178 spe_cpu = ANY_ANY;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
179 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
180 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
181 if ( (argc == 1)||(filename==0)) {
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
182 printf("Usage: ./fft -file [image filename] -cpu or -gpu\n");
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
183 exit(-1);
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
184 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
185
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
186 return filename;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
187 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
188
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
189 void
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
190 run_start(TaskManager *manager,pgm_t ipgm)
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
191 {
1663
ce031df3dd32 fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1661
diff changeset
192 long n = ipgm.width;
ce031df3dd32 fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents: 1661
diff changeset
193 long m = (cl_int)(log((double)n)/log(2.0));
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
194 size_t *gws = new size_t[2];
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
195 size_t *lws = new size_t[2];
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
196
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
197 xm = (cl_float2 *)malloc(n * n * sizeof(cl_float2));
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
198 rm = (cl_float2 *)malloc(n * n * sizeof(cl_float2));
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
199 wm = (cl_float2 *)malloc(n / 2 * sizeof(cl_float2));
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
200
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
201 HTask* waitTask;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
202 /*
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
203 * [cl_float2]
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
204 * typedef union
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
205 * {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
206 * cl_float CL_ALIGNED(8) s[2];
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
207 * #if defined( __GNUC__) && ! defined( __STRICT_ANSI__ )
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
208 * __extension__ struct{ cl_float x, y; };
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
209 * __extension__ struct{ cl_float s0, s1; };
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
210 * __extension__ struct{ cl_float lo, hi; };
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
211 * #endif
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
212 * #if defined( __CL_FLOAT2__)
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
213 * __cl_float2 v2;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
214 * #endif
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
215 * } cl_float2;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
216 */
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
217 for (int i=0; i<n; i++) {
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
218 for (int j=0; j < n; j++) {
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
219 ((float*)xm)[(2*n*j)+2*i+0] = (float)ipgm.buf[n*j+i];
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
220 ((float*)xm)[(2*n*j)+2*i+1] = (float)0;
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
221 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
222 }
1579
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1571
diff changeset
223
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1571
diff changeset
224 // Create spin factor
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
225 setWorkSize(gws,lws,n/2,1);
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
226 int length_w = n / 2;
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
227 HTask* sfac = manager->create_task(SPIN_FACT);
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
228 sfac->set_outData(0, wm, length_w*sizeof(cl_float2));
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
229 sfac->set_param(0,n);
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
230 sfac->set_cpu(spe_cpu);
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
231 sfac->flip();
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
232 sfac->iterate(gws[0]);
1625
6ff0c34c8a3c fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1581
diff changeset
233
1581
8ee897303cd0 fix multi_dimention
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1579
diff changeset
234 // Butterfly Operation
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
235 waitTask = fftCore(manager, rm, xm, wm, m, forward,sfac);
1661
19ab54c76d6f success run to fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1658
diff changeset
236
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
237 // Transpose matrix
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
238 int length_r =n*n;
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
239 setWorkSize(gws,lws,n,n);
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
240 HTask* first_trns = manager->create_task(TRANSPOSE);
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
241 first_trns->set_inData(0,rm,length_r*sizeof(cl_float2));
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
242 first_trns->set_outData(0,xm,length_r*sizeof(cl_float2));
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
243 first_trns->set_param(0,n);
1971
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1850
diff changeset
244 first_trns->set_cpu(spe_cpu);
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
245 first_trns->flip();
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
246 first_trns->wait_for(waitTask);
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
247 first_trns->iterate(gws[0],gws[1]);
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
248
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
249 // Butterfly Operation
1661
19ab54c76d6f success run to fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1658
diff changeset
250 waitTask = fftCore(manager, rm, xm, wm, m, forward,first_trns);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
251
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
252 // Apply high-pass filter
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
253 HTask *hpfl = manager->create_task(HIGH_PASS_FILTER);
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
254 cl_int radius = n/8;
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
255 setWorkSize(gws,lws,n,n);
1658
ed2fa50168ef fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1656
diff changeset
256 hpfl->set_inData(0,rm,length_r*sizeof(cl_float2));
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
257 hpfl->set_outData(0, rm, length_r*sizeof(cl_float2));
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
258 hpfl->set_param(0,n);
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
259 hpfl->set_param(1,(long)radius);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
260 hpfl->set_cpu(spe_cpu);
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
261 hpfl->flip();
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
262 hpfl->wait_for(waitTask);
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
263 hpfl->iterate(gws[0],gws[1]);
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
264
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
265 // Inverse FFT
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
266
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
267 // Butterfly Operation
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
268 waitTask = fftCore(manager,xm, rm, wm, m, inverse,hpfl);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
269
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
270 // Transpose matrix
1643
6c0b6947c231 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1633
diff changeset
271 setWorkSize(gws,lws,n,n);
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
272 HTask* second_trns = manager->create_task(TRANSPOSE);
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
273 second_trns->set_inData(0,xm,length_r*sizeof(cl_float2));
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
274 second_trns->set_outData(0,rm,length_r*sizeof(cl_float2));
1835
144e573b030b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1815
diff changeset
275 second_trns->set_param(0,n);
1971
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1850
diff changeset
276 second_trns->set_cpu(spe_cpu);
2014
8c618e912c88 optimization data transfer. wrong result
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 2013
diff changeset
277 second_trns->flip();
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
278 second_trns->wait_for(waitTask);
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
279 second_trns->iterate(gws[0],gws[1]);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
280
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
281 // Butterfly Operation
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
282
1656
05adaf6f7afb fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1646
diff changeset
283 waitTask = fftCore(manager,xm, rm, wm, m, inverse,second_trns);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
284 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
285
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
286 int TMmain(TaskManager *manager, int argc, char** argv) {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
287 task_init();
1779
0dd39c19dbb2 minor change
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1752
diff changeset
288 #ifdef GPU
1727
e99dc86d39e2 minor fix
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1717
diff changeset
289 gpu_task_init();
1779
0dd39c19dbb2 minor change
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1752
diff changeset
290 #endif
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
291 char * pgm_file = init(argc,argv);
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
292 /* Read image */
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
293 int err = readPGM(&ipgm, pgm_file);
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
294 if (err<0) {
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
295 fprintf(stderr, "Failed to read image file.\n");
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
296 exit(1);
1566
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1560
diff changeset
297 }
1689
c595245aa59e fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1687
diff changeset
298 run_start(manager, ipgm);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
299 st_time = getTime();
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
300 manager->set_TMend(TMend);
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
301 return 0;
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
302 }
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
303
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
304 void
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
305 TMend(TaskManager *manager)
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
306 {
1669
724fd3c1ac92 debugging iterate
kono
parents: 1663
diff changeset
307 ed_time = getTime();
1646
ab6b11476e02 fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1643
diff changeset
308 output();
1675
07ff7ec2a61b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1673
diff changeset
309 // fprintf(stdout, "image out put succeeded.\n");
07ff7ec2a61b fix fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents: 1673
diff changeset
310 printf("%0.6f\n",ed_time-st_time);
1551
57317332f6ef create fft example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
diff changeset
311 }