Mercurial > hg > Game > Cerium
annotate example/fft/main.cc @ 1752:537707f376c8 draft
fix profile
author | Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp> |
---|---|
date | Fri, 22 Nov 2013 02:49:43 +0900 |
parents | 4206f70691df |
children | 0dd39c19dbb2 |
rev | line source |
---|---|
1551 | 1 #include <stdio.h> |
2 #include <stdlib.h> | |
3 #include <math.h> | |
4 #include <sys/stat.h> | |
5 #include <fcntl.h> | |
6 #include <sys/time.h> | |
7 #include "TaskManager.h" | |
1727 | 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 | 10 #include "Func.h" |
11 #ifdef __APPLE__ | |
12 #include <OpenCL/opencl.h> | |
13 #else | |
14 #include <CL/cl.h> | |
15 #endif | |
16 #include "pgm.h" | |
17 extern void task_init(); | |
1727 | 18 extern void gpu_task_init(); |
1551 | 19 |
20 #define PI 3.14159265358979 | |
21 | |
22 #define MAX_SOURCE_SIZE (0x100000) | |
23 | |
24 #define AMP(a, b) (sqrt((a)*(a)+(b)*(b))) | |
25 | |
26 static double st_time; | |
27 static double ed_time; | |
28 void TMend(TaskManager *); | |
29 cl_device_id device_id = NULL; | |
30 cl_context context = NULL; | |
31 cl_command_queue queue = NULL; | |
32 cl_program program = NULL; | |
1752 | 33 CPU_TYPE spe_cpu = SPE_ANY; |
1551 | 34 |
1646 | 35 cl_float2* xm; |
36 cl_float2* rm; | |
37 cl_float2* wm; | |
38 pgm_t ipgm; | |
39 | |
1551 | 40 enum Mode { |
41 forward = 0, | |
42 inverse = 1 | |
43 }; | |
44 | |
45 static double | |
46 getTime() | |
47 { | |
48 struct timeval tv; | |
49 gettimeofday(&tv, NULL); | |
50 return tv.tv_sec + (double)tv.tv_usec*1e-6; | |
51 } | |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
52 |
1646 | 53 void |
54 output() | |
55 { | |
56 int n = ipgm.width; | |
57 float* ampd; | |
58 ampd = (float*)malloc(n*n*sizeof(float)); | |
59 for (int i=0; i < n; i++) { | |
60 for (int j=0; j < n; j++) { | |
61 ampd[n*((i))+((j))] = (AMP(((float*)xm)[(2*n*i)+2*j], ((float*)xm)[(2*n*i)+2*j+1])); | |
62 } | |
63 } | |
64 pgm_t opgm; | |
65 opgm.width = n; | |
66 opgm.height = n; | |
67 normalizeF2PGM(&opgm, ampd); | |
68 free(ampd); | |
69 | |
70 // Write out image | |
71 writePGM(&opgm, "output.pgm"); | |
72 | |
73 // Finalizations | |
74 destroyPGM(&ipgm); | |
75 destroyPGM(&opgm); | |
76 | |
77 free(wm); | |
78 free(rm); | |
79 free(xm); | |
80 } | |
81 | |
1551 | 82 const char *usr_help_str = "Usage: ./fft [option]\n \ |
83 options\n\ | |
84 -cpu Number of SPE used (default 1)\n\ | |
85 -l, --length Sorted number of data (default 1200)\n\ | |
86 -h, --help Print this message"; | |
87 | |
88 int setWorkSize(size_t* gws, size_t* lws, cl_int x, cl_int y) | |
89 { | |
90 switch(y) { | |
91 case 1: | |
92 gws[0] = x; | |
93 gws[1] = 1; | |
94 lws[0] = 1; | |
95 lws[1] = 1; | |
96 break; | |
97 default: | |
98 gws[0] = x; | |
99 gws[1] = y; | |
100 lws[0] = 1; | |
101 lws[1] = 1; | |
102 break; | |
103 } | |
104 | |
105 return 0; | |
106 } | |
107 | |
1646 | 108 HTask* |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
109 fftCore(TaskManager *manager,cl_float2 *dst, cl_float2 *src, cl_float2 *spin, long m, enum Mode direction,HTask* waitTask) |
1551 | 110 { |
1673 | 111 long direction_flag; |
1551 | 112 switch (direction) { |
1643 | 113 case forward:direction_flag = 0x00000000; break; |
114 case inverse:direction_flag = 0x80000000; break; | |
1551 | 115 } |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
116 long n = 1<<m; |
1551 | 117 size_t gws[2],lws[2]; |
1633
fbb4757d82ee
refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1627
diff
changeset
|
118 int length_dst = n*n; |
fbb4757d82ee
refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1627
diff
changeset
|
119 int length_src = n*n; |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
120 |
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
121 HTask* brev = manager->create_task(BIT_REVERSE); |
1633
fbb4757d82ee
refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1627
diff
changeset
|
122 setWorkSize(gws,lws,n,n); |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
123 brev->set_param(3,m); |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
124 brev->set_param(4,n); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
125 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
|
126 brev->set_outData(0, dst, length_dst*sizeof(cl_float2)); |
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
127 brev->set_cpu(spe_cpu); |
1646 | 128 brev->wait_for(waitTask); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
129 brev->iterate(gws[0],gws[1]); |
1551 | 130 |
1658 | 131 waitTask = brev; |
132 | |
1643 | 133 setWorkSize(gws,lws,n/2,n); |
1658 | 134 for(int iter=1;iter<=m;iter++) { |
135 HTask* bfly = manager->create_task(BUTTERFLY); | |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
136 bfly->set_param(3,n); |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
137 bfly->set_param(4,direction_flag); |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
138 bfly->set_param(5,(long)iter); |
1658 | 139 bfly->set_inData(0, dst, length_dst*sizeof(cl_float2)); |
140 bfly->set_inData(1, spin, sizeof(cl_float2)*(n/2)); | |
141 bfly->set_outData(0,dst,length_dst*sizeof(cl_float2)); | |
142 bfly->set_cpu(spe_cpu); | |
143 bfly->wait_for(waitTask); | |
144 bfly->iterate(gws[0],gws[1]); | |
145 waitTask = bfly; | |
146 } | |
1551 | 147 |
148 if (direction == inverse) { | |
1658 | 149 setWorkSize(gws,lws,n,n); |
1551 | 150 HTask *norm = manager->create_task(NORMALIZATION); |
1658 | 151 norm->set_inData(0,dst,length_dst*sizeof(cl_float2)); |
1571 | 152 norm->set_outData(0, dst, length_dst*sizeof(cl_float2)); |
1643 | 153 norm->set_param(3,n); |
1551 | 154 norm->set_cpu(spe_cpu); |
1658 | 155 norm->wait_for(waitTask); |
1656 | 156 norm->iterate(gws[0],gws[1]); |
1646 | 157 |
158 waitTask = norm; | |
1551 | 159 } |
1646 | 160 return waitTask; |
1551 | 161 } |
162 | |
163 char * | |
164 init(int argc, char**argv){ | |
165 | |
166 char *filename = 0; | |
167 | |
1689 | 168 // printf("%s ",argv[4]); |
1551 | 169 for (int i = 1; argv[i]; ++i) { |
170 if (strcmp(argv[i], "-file") == 0) { | |
171 filename = argv[i+1]; | |
172 } else if (strcmp(argv[i], "-g") == 0) { | |
173 spe_cpu = GPU_0; | |
1702
f52904f8f03e
fix example.run ANY_ANY
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1689
diff
changeset
|
174 } 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
|
175 spe_cpu = ANY_ANY; |
1551 | 176 } |
177 } | |
178 if ( (argc == 1)||(filename==0)) { | |
1643 | 179 printf("Usage: ./fft -file [image filename] -cpu or -gpu\n"); |
1551 | 180 exit(-1); |
181 } | |
182 | |
183 return filename; | |
184 } | |
185 | |
186 void | |
187 run_start(TaskManager *manager,pgm_t ipgm) | |
188 { | |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
189 long n = ipgm.width; |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
190 long m = (cl_int)(log((double)n)/log(2.0)); |
1643 | 191 size_t *gws = new size_t[2]; |
192 size_t *lws = new size_t[2]; | |
1555
096412ad80fb
add opencl example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1551
diff
changeset
|
193 |
1646 | 194 xm = (cl_float2 *)malloc(n * n * sizeof(cl_float2)); |
195 rm = (cl_float2 *)malloc(n * n * sizeof(cl_float2)); | |
196 wm = (cl_float2 *)malloc(n / 2 * sizeof(cl_float2)); | |
197 | |
198 HTask* waitTask; | |
1551 | 199 /* |
200 * [cl_float2] | |
201 * typedef union | |
202 * { | |
203 * cl_float CL_ALIGNED(8) s[2]; | |
204 * #if defined( __GNUC__) && ! defined( __STRICT_ANSI__ ) | |
205 * __extension__ struct{ cl_float x, y; }; | |
206 * __extension__ struct{ cl_float s0, s1; }; | |
207 * __extension__ struct{ cl_float lo, hi; }; | |
208 * #endif | |
209 * #if defined( __CL_FLOAT2__) | |
210 * __cl_float2 v2; | |
211 * #endif | |
212 * } cl_float2; | |
213 */ | |
1643 | 214 for (int i=0; i<n; i++) { |
215 for (int j=0; j < n; j++) { | |
216 ((float*)xm)[(2*n*j)+2*i+0] = (float)ipgm.buf[n*j+i]; | |
217 ((float*)xm)[(2*n*j)+2*i+1] = (float)0; | |
1551 | 218 } |
219 } | |
1579 | 220 |
221 // Create spin factor | |
1658 | 222 setWorkSize(gws,lws,n/2,1); |
1643 | 223 int length_w = n / 2; |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
224 HTask* sfac = manager->create_task(SPIN_FACT); |
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
225 sfac->set_outData(0, wm, length_w*sizeof(cl_float2)); |
1643 | 226 sfac->set_param(3,n); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
227 sfac->set_cpu(spe_cpu); |
1656 | 228 sfac->iterate(gws[0]); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
229 |
1581
8ee897303cd0
fix multi_dimention
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1579
diff
changeset
|
230 // Butterfly Operation |
1646 | 231 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
|
232 |
1551 | 233 // Transpose matrix |
1643 | 234 int length_r =n*n; |
1656 | 235 setWorkSize(gws,lws,n,n); |
236 HTask* first_trns = manager->create_task(TRANSPOSE); | |
237 first_trns->set_inData(0,rm,length_r*sizeof(cl_float2)); | |
238 first_trns->set_outData(0,xm,length_r*sizeof(cl_float2)); | |
239 first_trns->set_param(3,n); | |
240 first_trns->set_cpu(spe_cpu); | |
241 first_trns->wait_for(waitTask); | |
242 first_trns->iterate(gws[0],gws[1]); | |
243 | |
1551 | 244 // Butterfly Operation |
1661
19ab54c76d6f
success run to fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1658
diff
changeset
|
245 waitTask = fftCore(manager, rm, xm, wm, m, forward,first_trns); |
1551 | 246 |
247 // Apply high-pass filter | |
248 HTask *hpfl = manager->create_task(HIGH_PASS_FILTER); | |
1643 | 249 cl_int radius = n/8; |
1656 | 250 setWorkSize(gws,lws,n,n); |
1658 | 251 hpfl->set_inData(0,rm,length_r*sizeof(cl_float2)); |
1551 | 252 hpfl->set_outData(0, rm, length_r*sizeof(cl_float2)); |
1643 | 253 hpfl->set_param(3,n); |
1733 | 254 hpfl->set_param(4,(long)radius); |
1551 | 255 hpfl->set_cpu(spe_cpu); |
1656 | 256 hpfl->wait_for(waitTask); |
1643 | 257 hpfl->iterate(gws[0],gws[1]); |
1656 | 258 |
1551 | 259 // Inverse FFT |
260 | |
261 // Butterfly Operation | |
1646 | 262 waitTask = fftCore(manager,xm, rm, wm, m, inverse,hpfl); |
1551 | 263 |
264 // Transpose matrix | |
1643 | 265 setWorkSize(gws,lws,n,n); |
1656 | 266 HTask* second_trns = manager->create_task(TRANSPOSE); |
267 second_trns->set_inData(0,xm,length_r*sizeof(cl_float2)); | |
268 second_trns->set_outData(0,rm,length_r*sizeof(cl_float2)); | |
269 second_trns->set_param(3,n); | |
270 second_trns->set_cpu(spe_cpu); | |
271 second_trns->wait_for(waitTask); | |
272 second_trns->iterate(gws[0],gws[1]); | |
1551 | 273 |
274 // Butterfly Operation | |
275 | |
1656 | 276 waitTask = fftCore(manager,xm, rm, wm, m, inverse,second_trns); |
1551 | 277 } |
278 | |
279 int TMmain(TaskManager *manager, int argc, char** argv) { | |
280 task_init(); | |
1727 | 281 gpu_task_init(); |
1551 | 282 char * pgm_file = init(argc,argv); |
283 /* Read image */ | |
284 int err = readPGM(&ipgm, pgm_file); | |
285 if (err<0) { | |
286 fprintf(stderr, "Failed to read image file.\n"); | |
287 exit(1); | |
1566 | 288 } |
1689 | 289 run_start(manager, ipgm); |
1551 | 290 st_time = getTime(); |
291 manager->set_TMend(TMend); | |
292 return 0; | |
293 } | |
294 | |
295 void | |
296 TMend(TaskManager *manager) | |
297 { | |
1669 | 298 ed_time = getTime(); |
1646 | 299 output(); |
1675 | 300 // fprintf(stdout, "image out put succeeded.\n"); |
301 printf("%0.6f\n",ed_time-st_time); | |
1551 | 302 } |