Mercurial > hg > Members > yuuhi > OpenCL
annotate fft_fixstart/main.cc @ 5:3602b23914ad
fix device type bug
| author | Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp> |
|---|---|
| date | Tue, 05 Feb 2013 13:32:08 +0900 |
| parents | 8df0d3128672 |
| children | db074091ed0b |
| rev | line source |
|---|---|
| 3 | 1 #include <stdio.h> |
| 2 #include <stdlib.h> | |
| 3 #include <math.h> | |
| 4 #include <sys/stat.h> | |
| 5 #include <fcntl.h> | |
|
4
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
6 #include <sys/time.h> |
| 3 | 7 |
| 8 #ifdef __APPLE__ | |
| 9 #include <OpenCL/opencl.h> | |
| 10 #else | |
| 11 #include <CL/cl.h> | |
| 12 #endif | |
| 13 | |
| 14 #include "pgm.h" | |
| 15 | |
| 16 #define PI 3.14159265358979 | |
| 17 | |
| 18 #define MAX_SOURCE_SIZE (0x100000) | |
| 19 | |
| 20 #define AMP(a, b) (sqrt((a)*(a)+(b)*(b))) | |
| 21 | |
|
4
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
22 static double st_time; |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
23 static double ed_time; |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
24 |
| 3 | 25 cl_device_id device_id = NULL; |
| 26 cl_context context = NULL; | |
| 27 cl_command_queue queue = NULL; | |
| 28 cl_program program = NULL; | |
| 29 cl_device_type device_type = CL_DEVICE_TYPE_GPU; | |
| 30 | |
| 31 enum Mode { | |
| 32 forward = 0, | |
| 33 inverse = 1 | |
| 34 }; | |
| 35 | |
|
4
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
36 static double |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
37 getTime() |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
38 { |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
39 struct timeval tv; |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
40 gettimeofday(&tv, NULL); |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
41 return tv.tv_sec + (double)tv.tv_usec*1e-6; |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
42 } |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
43 |
| 3 | 44 int setWorkSize(size_t* gws, size_t* lws, cl_int x, cl_int y) |
| 45 { | |
| 46 switch(y) { | |
| 47 case 1: | |
| 48 gws[0] = x; | |
| 49 gws[1] = 1; | |
| 50 lws[0] = 1; | |
| 51 lws[1] = 1; | |
| 52 break; | |
| 53 default: | |
| 54 gws[0] = x; | |
| 55 gws[1] = y; | |
| 56 lws[0] = 1; | |
| 57 lws[1] = 1; | |
| 58 break; | |
| 59 } | |
| 60 | |
| 61 return 0; | |
| 62 } | |
| 63 | |
| 64 int fftCore(cl_mem dst, cl_mem src, cl_mem spin, cl_int m, enum Mode direction) | |
| 65 { | |
| 66 cl_int ret; | |
| 67 | |
| 68 cl_int iter; | |
| 69 cl_uint flag; | |
| 70 | |
| 71 cl_int n = 1<<m; | |
| 72 | |
| 73 cl_event kernelDone; | |
| 74 | |
| 75 cl_kernel brev = NULL; | |
| 76 cl_kernel bfly = NULL; | |
| 77 cl_kernel norm = NULL; | |
| 78 | |
| 79 brev = clCreateKernel(program, "bitReverse", &ret); | |
| 80 bfly = clCreateKernel(program, "butterfly", &ret); | |
| 81 norm = clCreateKernel(program, "norm", &ret); | |
| 82 | |
| 83 size_t gws[2]; | |
| 84 size_t lws[2]; | |
| 85 | |
| 86 switch (direction) { | |
| 87 case forward:flag = 0x00000000; break; | |
| 88 case inverse:flag = 0x80000000; break; | |
| 89 } | |
| 90 | |
| 91 ret = clSetKernelArg(brev, 0, sizeof(cl_mem), (void *)&dst); | |
| 92 ret = clSetKernelArg(brev, 1, sizeof(cl_mem), (void *)&src); | |
| 93 ret = clSetKernelArg(brev, 2, sizeof(cl_int), (void *)&m); | |
| 94 ret = clSetKernelArg(brev, 3, sizeof(cl_int), (void *)&n); | |
| 95 | |
| 96 ret = clSetKernelArg(bfly, 0, sizeof(cl_mem), (void *)&dst); | |
| 97 ret = clSetKernelArg(bfly, 1, sizeof(cl_mem), (void *)&spin); | |
| 98 ret = clSetKernelArg(bfly, 2, sizeof(cl_int), (void *)&m); | |
| 99 ret = clSetKernelArg(bfly, 3, sizeof(cl_int), (void *)&n); | |
| 100 ret = clSetKernelArg(bfly, 5, sizeof(cl_uint), (void *)&flag); | |
| 101 | |
| 102 ret = clSetKernelArg(norm, 0, sizeof(cl_mem), (void *)&dst); | |
| 103 ret = clSetKernelArg(norm, 1, sizeof(cl_int), (void *)&n); | |
| 104 | |
| 105 /* Reverse bit ordering */ | |
| 106 setWorkSize(gws, lws, n, n); | |
| 5 | 107 ret = clEnqueueTask(queue, brev, 0, NULL, NULL); |
| 3 | 108 |
| 109 /* Perform Butterfly Operations*/ | |
| 110 setWorkSize(gws, lws, n/2, n); | |
| 111 for (iter=1; iter <= m; iter++) { | |
| 112 ret = clSetKernelArg(bfly, 4, sizeof(cl_int), (void *)&iter); | |
| 5 | 113 ret = clEnqueueTask(queue, bfly, 0, NULL, NULL); |
| 3 | 114 ret = clWaitForEvents(1, &kernelDone); |
| 115 } | |
| 116 | |
| 117 if (direction == inverse) { | |
| 118 setWorkSize(gws, lws, n, n); | |
| 5 | 119 ret = clEnqueueTask(queue, norm, 0, NULL, NULL); |
| 3 | 120 ret = clWaitForEvents(1, &kernelDone); |
| 121 } | |
| 122 | |
| 123 ret = clReleaseKernel(bfly); | |
| 124 ret = clReleaseKernel(brev); | |
| 125 ret = clReleaseKernel(norm); | |
| 126 | |
| 127 return 0; | |
| 128 } | |
| 129 | |
| 130 char * | |
| 131 init(int argc, char**argv){ | |
| 132 | |
| 133 char *filename = 0; | |
| 134 | |
| 135 for (int i = 1; argv[i]; ++i) { | |
| 136 if (strcmp(argv[i], "-file") == 0) { | |
| 137 filename = argv[i+1]; | |
| 138 } else if (strcmp(argv[i], "-cpu") == 0) { | |
| 139 device_type = CL_DEVICE_TYPE_CPU; | |
| 5 | 140 printf("cpu!\n"); |
| 3 | 141 } else if (strcmp(argv[i], "-gpu") == 0) { |
| 142 device_type = CL_DEVICE_TYPE_GPU; | |
| 5 | 143 printf("gpu!\n"); |
| 3 | 144 } |
| 145 } | |
| 146 if ( (argc == 1)||(filename==0)) { | |
| 147 printf("Usage: ./fft [image filename]\n"); | |
| 148 exit(-1); | |
| 149 } | |
| 150 | |
| 151 return filename; | |
| 152 } | |
| 153 | |
| 154 int main(int argc, char** argv) { | |
| 155 cl_mem xmobj = NULL; | |
| 156 cl_mem rmobj = NULL; | |
| 157 cl_mem wmobj = NULL; | |
| 158 cl_kernel sfac = NULL; | |
| 159 cl_kernel trns = NULL; | |
| 160 cl_kernel hpfl = NULL; | |
| 161 | |
| 162 cl_platform_id platform_id = NULL; | |
| 163 | |
| 164 cl_uint ret_num_devices; | |
| 165 cl_uint ret_num_platforms; | |
| 166 | |
| 167 cl_int ret; | |
| 168 | |
| 169 cl_float2 *xm; | |
| 170 cl_float2 *rm; | |
| 171 cl_float2 *wm; | |
| 172 | |
| 173 pgm_t ipgm; | |
| 174 pgm_t opgm; | |
| 175 | |
| 176 const char fileName[] = "./fft.cl"; | |
| 177 size_t source_size; | |
| 178 char *source_str; | |
| 179 cl_int i, j; | |
| 180 cl_int n; | |
| 181 cl_int m; | |
| 182 | |
| 183 size_t gws[2]; | |
| 184 size_t lws[2]; | |
| 185 | |
| 186 /* Load kernel source code */ | |
| 187 int fd = open(fileName, O_RDONLY); | |
| 188 | |
| 189 if (fd<0) { | |
| 190 fprintf(stderr, "Failed to load kernel %s.\n",fileName); | |
| 191 exit(1); | |
| 192 } | |
| 193 struct stat stats; | |
| 194 fstat(fd, &stats); | |
| 195 off_t size = stats.st_size; | |
| 196 if (size<=0) { | |
| 197 fprintf(stderr, "Failed to load kernel.\n"); | |
| 198 exit(1); | |
| 199 } | |
| 200 source_str = (char*)alloca(size); | |
| 201 source_size = read(fd, source_str, size); | |
| 202 close( fd ); | |
| 203 | |
| 204 char * pgm_file = init(argc,argv); | |
| 205 | |
| 206 /* Read image */ | |
| 207 int err = readPGM(&ipgm, pgm_file); | |
| 208 if (err<0) { | |
| 209 fprintf(stderr, "Failed to read image file.\n"); | |
| 210 exit(1); | |
| 211 } | |
| 212 | |
| 213 n = ipgm.width; | |
| 214 m = (cl_int)(log((double)n)/log(2.0)); | |
| 215 | |
| 216 xm = (cl_float2 *)malloc(n * n * sizeof(cl_float2)); | |
| 217 rm = (cl_float2 *)malloc(n * n * sizeof(cl_float2)); | |
| 218 wm = (cl_float2 *)malloc(n / 2 * sizeof(cl_float2)); | |
| 219 | |
| 220 for (i=0; i < n; i++) { | |
| 221 for (j=0; j < n; j++) { | |
| 222 ((float*)xm)[(2*n*j)+2*i+0] = (float)ipgm.buf[n*j+i]; | |
| 223 ((float*)xm)[(2*n*j)+2*i+1] = (float)0; | |
| 224 } | |
| 225 } | |
| 226 | |
|
4
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
227 st_time = getTime(); |
| 3 | 228 /* Get platform/device */ |
| 229 ret = clGetPlatformIDs(1, &platform_id, &ret_num_platforms); | |
| 5 | 230 |
| 231 ret = clGetDeviceIDs( platform_id, device_type, 1, &device_id, &ret_num_devices); | |
| 3 | 232 |
| 233 /* Create OpenCL context */ | |
| 234 context = clCreateContext(NULL, 1, &device_id, NULL, NULL, &ret); | |
| 235 | |
| 236 /* Create Command queue */ | |
| 237 queue = clCreateCommandQueue(context, device_id, 0, &ret); | |
| 238 | |
| 239 /* Create Buffer Objects */ | |
| 240 xmobj = clCreateBuffer(context, CL_MEM_READ_WRITE, n*n*sizeof(cl_float2), NULL, &ret); | |
| 241 rmobj = clCreateBuffer(context, CL_MEM_READ_WRITE, n*n*sizeof(cl_float2), NULL, &ret); | |
| 242 wmobj = clCreateBuffer(context, CL_MEM_READ_WRITE, (n/2)*sizeof(cl_float2), NULL, &ret); | |
| 243 | |
| 244 /* Transfer data to memory buffer */ | |
| 245 ret = clEnqueueWriteBuffer(queue, xmobj, CL_TRUE, 0, n*n*sizeof(cl_float2), xm, 0, NULL, NULL); | |
| 246 | |
| 247 /* Create kernel program from source */ | |
| 248 program = clCreateProgramWithSource(context, 1, (const char **)&source_str, (const size_t *)&source_size, &ret); | |
| 249 | |
| 250 /* Build kernel program */ | |
| 251 ret = clBuildProgram(program, 1, &device_id, NULL, NULL, NULL); | |
| 252 | |
| 253 if (ret<0) { | |
| 254 size_t size; | |
| 255 clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, 0, NULL, &size); | |
| 256 | |
| 257 char *log = new char[size]; | |
| 258 clGetProgramBuildInfo(program, device_id, CL_PROGRAM_BUILD_LOG, size, log, NULL); | |
| 259 printf("%s ",log); | |
| 260 exit (ret); | |
| 261 } | |
| 262 | |
| 263 /* Create OpenCL Kernel */ | |
| 264 sfac = clCreateKernel(program, "spinFact", &ret); | |
| 265 trns = clCreateKernel(program, "transpose", &ret); | |
| 266 hpfl = clCreateKernel(program, "highPassFilter", &ret); | |
| 267 | |
| 268 /* Create spin factor */ | |
| 269 ret = clSetKernelArg(sfac, 0, sizeof(cl_mem), (void *)&wmobj); | |
| 270 ret = clSetKernelArg(sfac, 1, sizeof(cl_int), (void *)&n); | |
| 271 setWorkSize(gws, lws, n/2, 1); | |
| 5 | 272 ret = clEnqueueTask(queue, sfac, 0, NULL, NULL); |
| 3 | 273 |
| 274 /* Butterfly Operation */ | |
| 275 fftCore(rmobj, xmobj, wmobj, m, forward); | |
| 276 | |
| 277 /* Transpose matrix */ | |
| 278 ret = clSetKernelArg(trns, 0, sizeof(cl_mem), (void *)&xmobj); | |
| 279 ret = clSetKernelArg(trns, 1, sizeof(cl_mem), (void *)&rmobj); | |
| 280 ret = clSetKernelArg(trns, 2, sizeof(cl_int), (void *)&n); | |
| 281 setWorkSize(gws, lws, n, n); | |
| 5 | 282 ret = clEnqueueTask(queue, trns, 0, NULL, NULL); |
| 3 | 283 |
| 284 /* Butterfly Operation */ | |
| 285 fftCore(rmobj, xmobj, wmobj, m, forward); | |
| 286 | |
| 287 /* Apply high-pass filter */ | |
| 288 cl_int radius = n/8; | |
| 289 ret = clSetKernelArg(hpfl, 0, sizeof(cl_mem), (void *)&rmobj); | |
| 290 ret = clSetKernelArg(hpfl, 1, sizeof(cl_int), (void *)&n); | |
| 291 ret = clSetKernelArg(hpfl, 2, sizeof(cl_int), (void *)&radius); | |
| 292 setWorkSize(gws, lws, n, n); | |
| 5 | 293 ret = clEnqueueTask(queue, hpfl, 0, NULL, NULL); |
| 3 | 294 |
| 295 /* Inverse FFT */ | |
| 296 | |
| 297 /* Butterfly Operation */ | |
| 298 fftCore(xmobj, rmobj, wmobj, m, inverse); | |
| 299 | |
| 300 /* Transpose matrix */ | |
| 301 ret = clSetKernelArg(trns, 0, sizeof(cl_mem), (void *)&rmobj); | |
| 302 ret = clSetKernelArg(trns, 1, sizeof(cl_mem), (void *)&xmobj); | |
| 303 setWorkSize(gws, lws, n, n); | |
| 5 | 304 ret = clEnqueueTask(queue, trns, 0, NULL, NULL); |
| 3 | 305 |
| 306 /* Butterfly Operation */ | |
| 307 | |
| 308 fftCore(xmobj, rmobj, wmobj, m, inverse); | |
| 309 | |
| 310 /* Read data from memory buffer */ | |
| 311 ret = clEnqueueReadBuffer(queue, xmobj, CL_TRUE, 0, n*n*sizeof(cl_float2), xm, 0, NULL, NULL); | |
| 312 | |
| 313 /* */ | |
| 314 float* ampd; | |
| 315 ampd = (float*)malloc(n*n*sizeof(float)); | |
| 316 for (i=0; i < n; i++) { | |
| 317 for (j=0; j < n; j++) { | |
| 318 ampd[n*((i))+((j))] = (AMP(((float*)xm)[(2*n*i)+2*j], ((float*)xm)[(2*n*i)+2*j+1])); | |
| 319 } | |
| 320 } | |
| 321 opgm.width = n; | |
| 322 opgm.height = n; | |
| 323 normalizeF2PGM(&opgm, ampd); | |
| 324 free(ampd); | |
| 325 | |
| 326 /* Write out image */ | |
| 327 writePGM(&opgm, "output.pgm"); | |
| 328 | |
| 329 /* Finalizations*/ | |
| 330 ret = clFlush(queue); | |
| 331 ret = clFinish(queue); | |
|
4
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
332 ed_time = getTime(); |
| 3 | 333 ret = clReleaseKernel(hpfl); |
| 334 ret = clReleaseKernel(trns); | |
| 335 ret = clReleaseKernel(sfac); | |
| 336 ret = clReleaseProgram(program); | |
| 337 ret = clReleaseMemObject(xmobj); | |
| 338 ret = clReleaseMemObject(rmobj); | |
| 339 ret = clReleaseMemObject(wmobj); | |
| 340 ret = clReleaseCommandQueue(queue); | |
| 341 ret = clReleaseContext(context); | |
| 342 | |
| 343 destroyPGM(&ipgm); | |
| 344 destroyPGM(&opgm); | |
| 345 | |
| 346 free(wm); | |
| 347 free(rm); | |
| 348 free(xm); | |
| 349 | |
|
4
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
350 fprintf(stdout, "image out put succeeded.\n"); |
|
8df0d3128672
add time measurement function
Yuhi TOMARI <yuhi@cr.ie.u-ryukyu.ac.jp>
parents:
3
diff
changeset
|
351 printf("Time: %0.6f\n",ed_time-st_time); |
| 3 | 352 return 0; |
| 353 } |