Mercurial > hg > Game > Cerium
annotate example/fft/main.cc @ 1673:2c3adce7eb40 draft
fix fft on gpu
author | Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp> |
---|---|
date | Thu, 18 Jul 2013 17:16:43 +0900 |
parents | 724fd3c1ac92 |
children | 07ff7ec2a61b |
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" | |
1555
096412ad80fb
add opencl example
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1551
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 | 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(); | |
18 | |
19 #define PI 3.14159265358979 | |
20 | |
21 #define MAX_SOURCE_SIZE (0x100000) | |
22 | |
23 #define AMP(a, b) (sqrt((a)*(a)+(b)*(b))) | |
24 | |
25 static double st_time; | |
26 static double ed_time; | |
27 void TMend(TaskManager *); | |
28 cl_device_id device_id = NULL; | |
29 cl_context context = NULL; | |
30 cl_command_queue queue = NULL; | |
31 cl_program program = NULL; | |
32 CPU_TYPE spe_cpu = SPE_ANY; | |
33 | |
1646 | 34 cl_float2* xm; |
35 cl_float2* rm; | |
36 cl_float2* wm; | |
37 pgm_t ipgm; | |
38 | |
1551 | 39 enum Mode { |
40 forward = 0, | |
41 inverse = 1 | |
42 }; | |
43 | |
44 static double | |
45 getTime() | |
46 { | |
47 struct timeval tv; | |
48 gettimeofday(&tv, NULL); | |
49 return tv.tv_sec + (double)tv.tv_usec*1e-6; | |
50 } | |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
51 |
1646 | 52 void |
53 output() | |
54 { | |
55 int n = ipgm.width; | |
56 float* ampd; | |
57 ampd = (float*)malloc(n*n*sizeof(float)); | |
58 for (int i=0; i < n; i++) { | |
59 for (int j=0; j < n; j++) { | |
60 ampd[n*((i))+((j))] = (AMP(((float*)xm)[(2*n*i)+2*j], ((float*)xm)[(2*n*i)+2*j+1])); | |
61 } | |
62 } | |
63 pgm_t opgm; | |
64 opgm.width = n; | |
65 opgm.height = n; | |
66 normalizeF2PGM(&opgm, ampd); | |
67 free(ampd); | |
68 | |
69 // Write out image | |
70 writePGM(&opgm, "output.pgm"); | |
71 | |
72 // Finalizations | |
73 destroyPGM(&ipgm); | |
74 destroyPGM(&opgm); | |
75 | |
76 free(wm); | |
77 free(rm); | |
78 free(xm); | |
79 } | |
80 | |
1551 | 81 const char *usr_help_str = "Usage: ./fft [option]\n \ |
82 options\n\ | |
83 -cpu Number of SPE used (default 1)\n\ | |
84 -l, --length Sorted number of data (default 1200)\n\ | |
85 -h, --help Print this message"; | |
86 | |
87 int setWorkSize(size_t* gws, size_t* lws, cl_int x, cl_int y) | |
88 { | |
89 switch(y) { | |
90 case 1: | |
91 gws[0] = x; | |
92 gws[1] = 1; | |
93 lws[0] = 1; | |
94 lws[1] = 1; | |
95 break; | |
96 default: | |
97 gws[0] = x; | |
98 gws[1] = y; | |
99 lws[0] = 1; | |
100 lws[1] = 1; | |
101 break; | |
102 } | |
103 | |
104 return 0; | |
105 } | |
106 | |
1646 | 107 HTask* |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
108 fftCore(TaskManager *manager,cl_float2 *dst, cl_float2 *src, cl_float2 *spin, long m, enum Mode direction,HTask* waitTask) |
1551 | 109 { |
1673 | 110 long direction_flag; |
1551 | 111 switch (direction) { |
1643 | 112 case forward:direction_flag = 0x00000000; break; |
113 case inverse:direction_flag = 0x80000000; break; | |
1551 | 114 } |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
115 long n = 1<<m; |
1551 | 116 size_t gws[2],lws[2]; |
1633
fbb4757d82ee
refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1627
diff
changeset
|
117 int length_dst = n*n; |
fbb4757d82ee
refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1627
diff
changeset
|
118 int length_src = n*n; |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
119 |
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
120 HTask* brev = manager->create_task(BIT_REVERSE); |
1633
fbb4757d82ee
refactor GpuScheduler
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1627
diff
changeset
|
121 setWorkSize(gws,lws,n,n); |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
122 brev->set_param(3,m); |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
123 brev->set_param(4,n); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
124 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
|
125 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
|
126 brev->set_cpu(spe_cpu); |
1646 | 127 brev->wait_for(waitTask); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
128 brev->iterate(gws[0],gws[1]); |
1551 | 129 |
1658 | 130 waitTask = brev; |
131 | |
1643 | 132 setWorkSize(gws,lws,n/2,n); |
1658 | 133 for(int iter=1;iter<=m;iter++) { |
134 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
|
135 bfly->set_param(3,n); |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
136 bfly->set_param(4,direction_flag); |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
137 bfly->set_param(5,(long)iter); |
1658 | 138 bfly->set_inData(0, dst, length_dst*sizeof(cl_float2)); |
139 bfly->set_inData(1, spin, sizeof(cl_float2)*(n/2)); | |
140 bfly->set_outData(0,dst,length_dst*sizeof(cl_float2)); | |
141 bfly->set_cpu(spe_cpu); | |
142 bfly->wait_for(waitTask); | |
143 bfly->iterate(gws[0],gws[1]); | |
144 waitTask = bfly; | |
145 } | |
1551 | 146 |
147 if (direction == inverse) { | |
1658 | 148 setWorkSize(gws,lws,n,n); |
1551 | 149 HTask *norm = manager->create_task(NORMALIZATION); |
1658 | 150 norm->set_inData(0,dst,length_dst*sizeof(cl_float2)); |
1571 | 151 norm->set_outData(0, dst, length_dst*sizeof(cl_float2)); |
1643 | 152 norm->set_param(3,n); |
1551 | 153 norm->set_cpu(spe_cpu); |
1658 | 154 norm->wait_for(waitTask); |
1656 | 155 norm->iterate(gws[0],gws[1]); |
1646 | 156 |
157 waitTask = norm; | |
1551 | 158 } |
1646 | 159 return waitTask; |
1551 | 160 } |
161 | |
162 char * | |
163 init(int argc, char**argv){ | |
164 | |
165 char *filename = 0; | |
166 | |
167 for (int i = 1; argv[i]; ++i) { | |
168 if (strcmp(argv[i], "-file") == 0) { | |
169 filename = argv[i+1]; | |
170 } else if (strcmp(argv[i], "-g") == 0) { | |
171 spe_cpu = GPU_0; | |
172 } | |
173 } | |
174 if ( (argc == 1)||(filename==0)) { | |
1643 | 175 printf("Usage: ./fft -file [image filename] -cpu or -gpu\n"); |
1551 | 176 exit(-1); |
177 } | |
178 | |
179 return filename; | |
180 } | |
181 | |
182 void | |
183 run_start(TaskManager *manager,pgm_t ipgm) | |
184 { | |
1663
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
185 long n = ipgm.width; |
ce031df3dd32
fix syntax fft gpu
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
1661
diff
changeset
|
186 long m = (cl_int)(log((double)n)/log(2.0)); |
1643 | 187 size_t *gws = new size_t[2]; |
188 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
|
189 |
1646 | 190 xm = (cl_float2 *)malloc(n * n * sizeof(cl_float2)); |
191 rm = (cl_float2 *)malloc(n * n * sizeof(cl_float2)); | |
192 wm = (cl_float2 *)malloc(n / 2 * sizeof(cl_float2)); | |
193 | |
194 HTask* waitTask; | |
1551 | 195 /* |
196 * [cl_float2] | |
197 * typedef union | |
198 * { | |
199 * cl_float CL_ALIGNED(8) s[2]; | |
200 * #if defined( __GNUC__) && ! defined( __STRICT_ANSI__ ) | |
201 * __extension__ struct{ cl_float x, y; }; | |
202 * __extension__ struct{ cl_float s0, s1; }; | |
203 * __extension__ struct{ cl_float lo, hi; }; | |
204 * #endif | |
205 * #if defined( __CL_FLOAT2__) | |
206 * __cl_float2 v2; | |
207 * #endif | |
208 * } cl_float2; | |
209 */ | |
1643 | 210 for (int i=0; i<n; i++) { |
211 for (int j=0; j < n; j++) { | |
212 ((float*)xm)[(2*n*j)+2*i+0] = (float)ipgm.buf[n*j+i]; | |
213 ((float*)xm)[(2*n*j)+2*i+1] = (float)0; | |
1551 | 214 } |
215 } | |
1579 | 216 |
217 // Create spin factor | |
1658 | 218 setWorkSize(gws,lws,n/2,1); |
1643 | 219 int length_w = n / 2; |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
220 HTask* sfac = manager->create_task(SPIN_FACT); |
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
221 sfac->set_outData(0, wm, length_w*sizeof(cl_float2)); |
1643 | 222 sfac->set_param(3,n); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
223 sfac->set_cpu(spe_cpu); |
1656 | 224 sfac->iterate(gws[0]); |
1625
6ff0c34c8a3c
fix fft , used iterate
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1581
diff
changeset
|
225 |
1581
8ee897303cd0
fix multi_dimention
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1579
diff
changeset
|
226 // Butterfly Operation |
1646 | 227 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
|
228 |
1551 | 229 // Transpose matrix |
1643 | 230 int length_r =n*n; |
1656 | 231 setWorkSize(gws,lws,n,n); |
232 HTask* first_trns = manager->create_task(TRANSPOSE); | |
233 first_trns->set_inData(0,rm,length_r*sizeof(cl_float2)); | |
234 first_trns->set_outData(0,xm,length_r*sizeof(cl_float2)); | |
235 first_trns->set_param(3,n); | |
236 first_trns->set_cpu(spe_cpu); | |
237 first_trns->wait_for(waitTask); | |
238 first_trns->iterate(gws[0],gws[1]); | |
239 | |
1551 | 240 // Butterfly Operation |
1661
19ab54c76d6f
success run to fft
Shohei KOKUBO <e105744@ie.u-ryukyu.ac.jp>
parents:
1658
diff
changeset
|
241 waitTask = fftCore(manager, rm, xm, wm, m, forward,first_trns); |
1551 | 242 |
243 // Apply high-pass filter | |
244 HTask *hpfl = manager->create_task(HIGH_PASS_FILTER); | |
1643 | 245 cl_int radius = n/8; |
1656 | 246 setWorkSize(gws,lws,n,n); |
1658 | 247 hpfl->set_inData(0,rm,length_r*sizeof(cl_float2)); |
1551 | 248 hpfl->set_outData(0, rm, length_r*sizeof(cl_float2)); |
1643 | 249 hpfl->set_param(3,n); |
250 hpfl->set_param(4,radius); | |
1551 | 251 hpfl->set_cpu(spe_cpu); |
1656 | 252 hpfl->wait_for(waitTask); |
1643 | 253 hpfl->iterate(gws[0],gws[1]); |
1656 | 254 |
1551 | 255 // Inverse FFT |
256 | |
257 // Butterfly Operation | |
1646 | 258 waitTask = fftCore(manager,xm, rm, wm, m, inverse,hpfl); |
1551 | 259 |
260 // Transpose matrix | |
1643 | 261 setWorkSize(gws,lws,n,n); |
1656 | 262 HTask* second_trns = manager->create_task(TRANSPOSE); |
263 second_trns->set_inData(0,xm,length_r*sizeof(cl_float2)); | |
264 second_trns->set_outData(0,rm,length_r*sizeof(cl_float2)); | |
265 second_trns->set_param(3,n); | |
266 second_trns->set_cpu(spe_cpu); | |
267 second_trns->wait_for(waitTask); | |
268 second_trns->iterate(gws[0],gws[1]); | |
1551 | 269 |
270 // Butterfly Operation | |
271 | |
1656 | 272 waitTask = fftCore(manager,xm, rm, wm, m, inverse,second_trns); |
1551 | 273 } |
274 | |
275 int TMmain(TaskManager *manager, int argc, char** argv) { | |
276 task_init(); | |
277 char * pgm_file = init(argc,argv); | |
278 /* Read image */ | |
279 int err = readPGM(&ipgm, pgm_file); | |
280 if (err<0) { | |
281 fprintf(stderr, "Failed to read image file.\n"); | |
282 exit(1); | |
1566 | 283 } |
1551 | 284 st_time = getTime(); |
285 run_start(manager, ipgm); | |
286 manager->set_TMend(TMend); | |
287 return 0; | |
288 } | |
289 | |
290 void | |
291 TMend(TaskManager *manager) | |
292 { | |
1669 | 293 ed_time = getTime(); |
1646 | 294 output(); |
1551 | 295 fprintf(stdout, "image out put succeeded.\n"); |
296 printf("Time: %0.6f\n",ed_time-st_time); | |
297 } |