clray

annotate src/ocl.cc @ 11:d9a1bab1c3f5

find changesets by author, revision, files, or words in the commit message
ported to windows
author John Tsiombikas
date Sat, 31 Jul 2010 22:23:57 +0100
parents deaf85acf6af
children 85fd61f374d9
rev   line source
nuclear@8 1 #define OCL_CC_
nuclear@8 2
nuclear@0 3 #include <stdio.h>
nuclear@0 4 #include <stdlib.h>
nuclear@0 5 #include <string.h>
nuclear@8 6 #include <stdarg.h>
nuclear@0 7 #include <errno.h>
John@11 8 #ifndef _MSC_VER
nuclear@0 9 #include <alloca.h>
John@11 10 #else
John@11 11 #include <malloc.h>
John@11 12 #endif
nuclear@0 13 #include <sys/stat.h>
nuclear@0 14 #include "ocl.h"
nuclear@8 15 #include "ocl_errstr.h"
nuclear@0 16
nuclear@0 17
nuclear@0 18 class InitCL {
nuclear@0 19 public:
nuclear@0 20 InitCL();
nuclear@0 21 };
nuclear@0 22
nuclear@0 23 struct device_info {
nuclear@0 24 cl_device_id id;
nuclear@0 25 cl_device_type type;
nuclear@0 26 unsigned int units;
nuclear@0 27 unsigned int clock;
nuclear@0 28
nuclear@0 29 unsigned int dim;
nuclear@0 30 size_t *work_item_sizes;
nuclear@0 31 size_t work_group_size;
nuclear@0 32
nuclear@0 33 unsigned long mem_size;
nuclear@0 34 };
nuclear@0 35
nuclear@0 36 static bool init_opencl(void);
nuclear@0 37 static int select_device(struct device_info *di, int (*devcmp)(struct device_info*, struct device_info*));
nuclear@0 38 static int get_dev_info(cl_device_id dev, struct device_info *di);
nuclear@0 39 static int devcmp(struct device_info *a, struct device_info *b);
nuclear@0 40 static const char *devtypestr(cl_device_type type);
nuclear@0 41 static void print_memsize(FILE *out, unsigned long memsz);
nuclear@8 42 static const char *clstrerror(int err);
nuclear@0 43
nuclear@0 44
nuclear@0 45 static InitCL initcl;
nuclear@0 46 static cl_context ctx;
nuclear@0 47 static cl_command_queue cmdq;
nuclear@0 48 static device_info devinf;
nuclear@0 49
nuclear@0 50 InitCL::InitCL()
nuclear@0 51 {
nuclear@0 52 if(!init_opencl()) {
nuclear@0 53 exit(0);
nuclear@0 54 }
nuclear@0 55 }
nuclear@0 56
nuclear@0 57 static bool init_opencl(void)
nuclear@0 58 {
nuclear@0 59 if(select_device(&devinf, devcmp) == -1) {
nuclear@0 60 return false;
nuclear@0 61 }
nuclear@0 62
nuclear@0 63
nuclear@0 64 if(!(ctx = clCreateContext(0, 1, &devinf.id, 0, 0, 0))) {
nuclear@0 65 fprintf(stderr, "failed to create opencl context\n");
nuclear@0 66 return false;
nuclear@0 67 }
nuclear@0 68
nuclear@0 69 if(!(cmdq = clCreateCommandQueue(ctx, devinf.id, 0, 0))) {
nuclear@0 70 fprintf(stderr, "failed to create command queue\n");
nuclear@0 71 return false;
nuclear@0 72 }
nuclear@0 73 return true;
nuclear@0 74 }
nuclear@0 75
nuclear@0 76
nuclear@0 77 CLMemBuffer *create_mem_buffer(int rdwr, size_t sz, void *buf)
nuclear@0 78 {
nuclear@0 79 int err;
nuclear@0 80 cl_mem mem;
nuclear@0 81
nuclear@0 82
nuclear@0 83 if(!(mem = clCreateBuffer(ctx, rdwr | CL_MEM_USE_HOST_PTR, sz, buf, &err))) {
nuclear@8 84 fprintf(stderr, "failed to create memory buffer: %s\n", clstrerror(err));
nuclear@0 85 return 0;
nuclear@0 86 }
nuclear@0 87
nuclear@0 88 CLMemBuffer *mbuf = new CLMemBuffer;
nuclear@0 89 mbuf->mem = mem;
nuclear@0 90 mbuf->size = sz;
nuclear@0 91 return mbuf;
nuclear@0 92 }
nuclear@0 93
nuclear@0 94 void destroy_mem_buffer(CLMemBuffer *mbuf)
nuclear@0 95 {
nuclear@0 96 if(mbuf) {
nuclear@0 97
nuclear@0 98 clReleaseMemObject(mbuf->mem);
nuclear@0 99 delete mbuf;
nuclear@0 100 }
nuclear@0 101 }
nuclear@0 102
nuclear@0 103 void *map_mem_buffer(CLMemBuffer *mbuf, int rdwr)
nuclear@0 104 {
nuclear@0 105 if(!mbuf) return 0;
nuclear@0 106
nuclear@0 107 int err;
nuclear@0 108 mbuf->ptr = clEnqueueMapBuffer(cmdq, mbuf->mem, 1, rdwr, 0, mbuf->size, 0, 0, 0, &err);
nuclear@0 109 if(!mbuf->ptr) {
nuclear@8 110 fprintf(stderr, "failed to map buffer: %s\n", clstrerror(err));
nuclear@0 111 return 0;
nuclear@0 112 }
nuclear@0 113 return mbuf->ptr;
nuclear@0 114 }
nuclear@0 115
nuclear@0 116 void unmap_mem_buffer(CLMemBuffer *mbuf)
nuclear@0 117 {
nuclear@0 118 if(!mbuf || !mbuf->ptr) return;
nuclear@0 119 clEnqueueUnmapMemObject(cmdq, mbuf->mem, mbuf->ptr, 0, 0, 0);
nuclear@0 120 }
nuclear@0 121
nuclear@0 122 bool write_mem_buffer(CLMemBuffer *mbuf, size_t sz, void *src)
nuclear@0 123 {
nuclear@0 124 if(!mbuf) return false;
nuclear@0 125
nuclear@0 126 int err;
nuclear@0 127 if((err = clEnqueueWriteBuffer(cmdq, mbuf->mem, 1, 0, sz, src, 0, 0, 0)) != 0) {
nuclear@8 128 fprintf(stderr, "failed to write buffer: %s\n", clstrerror(err));
nuclear@0 129 return false;
nuclear@0 130 }
nuclear@0 131 return true;
nuclear@0 132 }
nuclear@0 133
nuclear@0 134 bool read_mem_buffer(CLMemBuffer *mbuf, size_t sz, void *dest)
nuclear@0 135 {
nuclear@0 136 if(!mbuf) return false;
nuclear@0 137
nuclear@0 138 int err;
nuclear@0 139 if((err = clEnqueueReadBuffer(cmdq, mbuf->mem, 1, 0, sz, dest, 0, 0, 0)) != 0) {
nuclear@8 140 fprintf(stderr, "failed to read buffer: %s\n", clstrerror(err));
nuclear@0 141 return false;
nuclear@0 142 }
nuclear@0 143 return true;
nuclear@0 144 }
nuclear@0 145
nuclear@0 146
nuclear@0 147 CLProgram::CLProgram(const char *kname)
nuclear@0 148 {
nuclear@0 149 prog = 0;
nuclear@0 150 kernel = 0;
nuclear@0 151 this->kname = kname;
nuclear@1 152 args.resize(16);
nuclear@0 153 built = false;
nuclear@0 154 }
nuclear@0 155
nuclear@0 156 CLProgram::~CLProgram()
nuclear@0 157 {
nuclear@0 158 if(prog) {
nuclear@0 159
nuclear@0 160 clReleaseProgram(prog);
nuclear@0 161 }
nuclear@0 162 if(kernel) {
nuclear@0 163
nuclear@0 164 clReleaseKernel(kernel);
nuclear@0 165 }
nuclear@1 166 for(size_t i=0; i<args.size(); i++) {
nuclear@1 167 if(args[i].type == ARGTYPE_MEM_BUF) {
nuclear@1 168 destroy_mem_buffer(args[i].v.mbuf);
nuclear@0 169 }
nuclear@0 170 }
nuclear@0 171 }
nuclear@0 172
nuclear@0 173 bool CLProgram::load(const char *fname)
nuclear@0 174 {
nuclear@0 175 FILE *fp;
nuclear@0 176 char *src;
nuclear@0 177 struct stat st;
nuclear@0 178
nuclear@0 179 printf("loading opencl program (%s)\n", fname);
nuclear@0 180
nuclear@0 181 if(!(fp = fopen(fname, "rb"))) {
nuclear@0 182 fprintf(stderr, "failed to open %s: %s\n", fname, strerror(errno));
nuclear@0 183 return false;
nuclear@0 184 }
nuclear@0 185
nuclear@0 186 fstat(fileno(fp), &st);
nuclear@0 187
nuclear@0 188 src = new char[st.st_size + 1];
nuclear@0 189
nuclear@0 190 fread(src, 1, st.st_size, fp);
nuclear@0 191 src[st.st_size] = 0;
nuclear@0 192 fclose(fp);
nuclear@0 193
nuclear@0 194
nuclear@0 195 if(!(prog = clCreateProgramWithSource(ctx, 1, (const char**)&src, 0, 0))) {
nuclear@0 196 fprintf(stderr, "error creating program object: %s\n", fname);
nuclear@0 197 delete [] src;
nuclear@0 198 return false;
nuclear@0 199 }
nuclear@0 200 delete [] src;
nuclear@0 201 return true;
nuclear@0 202 }
nuclear@0 203
nuclear@1 204 bool CLProgram::set_argi(int idx, int val)
nuclear@1 205 {
nuclear@1 206 if((int)args.size() <= idx) {
nuclear@1 207 args.resize(idx + 1);
nuclear@1 208 }
nuclear@1 209
nuclear@1 210 CLArg *arg = &args[idx];
nuclear@1 211 arg->type = ARGTYPE_INT;
nuclear@1 212 arg->v.ival = val;
nuclear@1 213 return true;
nuclear@1 214 }
nuclear@1 215
nuclear@1 216 bool CLProgram::set_argf(int idx, float val)
nuclear@1 217 {
nuclear@1 218 if((int)args.size() <= idx) {
nuclear@1 219 args.resize(idx + 1);
nuclear@1 220 }
nuclear@1 221
nuclear@1 222 CLArg *arg = &args[idx];
nuclear@1 223 arg->type = ARGTYPE_FLOAT;
nuclear@1 224 arg->v.fval = val;
nuclear@1 225 return true;
nuclear@1 226 }
nuclear@1 227
nuclear@1 228 bool CLProgram::set_arg_buffer(int idx, int rdwr, size_t sz, void *ptr)
nuclear@0 229 {
nuclear@0 230 CLMemBuffer *buf;
nuclear@0 231
nuclear@0 232 if(!(buf = create_mem_buffer(rdwr, sz, ptr))) {
nuclear@0 233 return false;
nuclear@0 234 }
nuclear@0 235
nuclear@1 236 if((int)args.size() <= idx) {
nuclear@1 237 args.resize(idx + 1);
nuclear@0 238 }
nuclear@1 239 args[idx].type = ARGTYPE_MEM_BUF;
nuclear@1 240 args[idx].v.mbuf = buf;
nuclear@0 241 return true;
nuclear@0 242 }
nuclear@0 243
nuclear@0 244 CLMemBuffer *CLProgram::get_arg_buffer(int arg)
nuclear@0 245 {
nuclear@1 246 if(arg < 0 || arg >= (int)args.size() || args[arg].type != ARGTYPE_MEM_BUF) {
nuclear@0 247 return 0;
nuclear@0 248 }
nuclear@1 249 return args[arg].v.mbuf;
nuclear@0 250 }
nuclear@0 251
nuclear@0 252 bool CLProgram::build()
nuclear@0 253 {
nuclear@2 254 int err;
nuclear@0 255
nuclear@2 256 if((err = clBuildProgram(prog, 0, 0, 0, 0, 0)) != 0) {
nuclear@2 257 size_t sz;
nuclear@2 258 clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, 0, 0, &sz);
nuclear@0 259
nuclear@2 260 char *errlog = (char*)alloca(sz + 1);
nuclear@2 261 clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, sz, errlog, 0);
nuclear@8 262 fprintf(stderr, "failed to build program: %s\n%s\n", clstrerror(err), errlog);
nuclear@2 263
nuclear@0 264 clReleaseProgram(prog);
nuclear@0 265 prog = 0;
nuclear@0 266 return false;
nuclear@0 267 }
nuclear@0 268
nuclear@0 269
nuclear@0 270 if(!(kernel = clCreateKernel(prog, kname.c_str(), 0))) {
nuclear@0 271 fprintf(stderr, "failed to create kernel: %s\n", kname.c_str());
nuclear@0 272 clReleaseProgram(prog);
nuclear@0 273 prog = 0;
nuclear@0 274 return false;
nuclear@0 275 }
nuclear@0 276
nuclear@1 277 for(size_t i=0; i<args.size(); i++) {
nuclear@1 278 int err;
nuclear@0 279
nuclear@1 280 if(args[i].type == ARGTYPE_NONE) {
nuclear@1 281 break;
nuclear@1 282 }
nuclear@1 283
nuclear@1 284 switch(args[i].type) {
nuclear@1 285 case ARGTYPE_INT:
nuclear@1 286 if((err = clSetKernelArg(kernel, i, sizeof(int), &args[i].v.ival)) != 0) {
nuclear@8 287 fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
nuclear@1 288 goto fail;
nuclear@1 289 }
nuclear@1 290 break;
nuclear@1 291
nuclear@1 292 case ARGTYPE_FLOAT:
nuclear@1 293 if((err = clSetKernelArg(kernel, i, sizeof(float), &args[i].v.fval)) != 0) {
nuclear@8 294 fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
nuclear@1 295 goto fail;
nuclear@1 296 }
nuclear@1 297 break;
nuclear@1 298
nuclear@1 299 case ARGTYPE_MEM_BUF:
nuclear@1 300 {
nuclear@1 301 CLMemBuffer *mbuf = args[i].v.mbuf;
nuclear@1 302
nuclear@1 303 if((err = clSetKernelArg(kernel, i, sizeof mbuf->mem, &mbuf->mem)) != 0) {
nuclear@8 304 fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
nuclear@1 305 goto fail;
nuclear@1 306 }
nuclear@1 307 }
nuclear@1 308 break;
nuclear@1 309
nuclear@1 310 default:
nuclear@1 311 break;
nuclear@0 312 }
nuclear@0 313 }
nuclear@0 314
nuclear@0 315 built = true;
nuclear@0 316 return true;
nuclear@1 317
nuclear@1 318 fail:
nuclear@1 319 clReleaseProgram(prog);
nuclear@1 320 clReleaseKernel(kernel);
nuclear@1 321 prog = 0;
nuclear@1 322 kernel = 0;
nuclear@1 323 return false;
nuclear@0 324 }
nuclear@0 325
nuclear@0 326 bool CLProgram::run() const
nuclear@0 327 {
nuclear@0 328 return run(1, 1);
nuclear@0 329 }
nuclear@0 330
nuclear@0 331 bool CLProgram::run(int dim, ...) const
nuclear@0 332 {
nuclear@0 333 if(!built) {
nuclear@0 334 if(!((CLProgram*)this)->build()) {
nuclear@0 335 return false;
nuclear@0 336 }
nuclear@0 337 }
nuclear@0 338
nuclear@0 339 va_list ap;
nuclear@0 340 size_t *global_size = (size_t*)alloca(dim * sizeof *global_size);
nuclear@0 341
nuclear@0 342 va_start(ap, dim);
nuclear@0 343 for(int i=0; i<dim; i++) {
nuclear@0 344 global_size[i] = va_arg(ap, int);
nuclear@0 345 }
nuclear@0 346 va_end(ap);
nuclear@0 347
nuclear@0 348 int err;
nuclear@0 349 if((err = clEnqueueNDRangeKernel(cmdq, kernel, dim, 0, global_size, 0, 0, 0, 0)) != 0) {
nuclear@8 350 fprintf(stderr, "error executing kernel: %s\n", clstrerror(err));
nuclear@0 351 return false;
nuclear@0 352 }
nuclear@0 353 return true;
nuclear@0 354 }
nuclear@0 355
nuclear@0 356 static int select_device(struct device_info *dev_inf, int (*devcmp)(struct device_info*, struct device_info*))
nuclear@0 357 {
nuclear@8 358 unsigned int i, j, num_dev, num_plat, sel, ret;
nuclear@0 359 cl_device_id dev[32];
nuclear@8 360 cl_platform_id plat[32];
nuclear@0 361
nuclear@0 362 dev_inf->work_item_sizes = 0;
nuclear@0 363
nuclear@8 364 if((ret = clGetPlatformIDs(32, plat, &num_plat)) != 0) {
nuclear@8 365 fprintf(stderr, "clGetPlatformIDs failed: %s\n", clstrerror(ret));
nuclear@8 366 return -1;
nuclear@8 367 }
nuclear@8 368 if(!num_plat) {
nuclear@8 369 fprintf(stderr, "OpenCL not available!\n");
nuclear@8 370 return -1;
nuclear@8 371 }
nuclear@0 372
nuclear@8 373 for(i=0; i<num_plat; i++) {
nuclear@8 374 char buf[512];
nuclear@8 375
nuclear@8 376 clGetPlatformInfo(plat[i], CL_PLATFORM_NAME, sizeof buf, buf, 0);
nuclear@8 377 printf("[%d]: %s", i, buf);
nuclear@8 378 clGetPlatformInfo(plat[i], CL_PLATFORM_VENDOR, sizeof buf, buf, 0);
nuclear@8 379 printf(", %s", buf);
nuclear@8 380 clGetPlatformInfo(plat[i], CL_PLATFORM_VERSION, sizeof buf, buf, 0);
nuclear@8 381 printf(" (%s)\n", buf);
nuclear@8 382 }
nuclear@8 383
nuclear@8 384 if((ret = clGetDeviceIDs(plat[0], CL_DEVICE_TYPE_ALL, 32, dev, &num_dev)) != 0) {
nuclear@8 385 fprintf(stderr, "clGetDeviceIDs failed: %s\n", clstrerror(ret));
nuclear@8 386 return -1;
nuclear@8 387 }
nuclear@0 388 printf("found %d cl devices.\n", num_dev);
nuclear@0 389
nuclear@0 390 for(i=0; i<num_dev; i++) {
nuclear@0 391 struct device_info di;
nuclear@0 392
nuclear@0 393 if(get_dev_info(dev[i], &di) == -1) {
nuclear@0 394 free(dev_inf->work_item_sizes);
nuclear@0 395 return -1;
nuclear@0 396 }
nuclear@0 397
nuclear@0 398 printf("--> device %u (%s)\n", i, devtypestr(di.type));
nuclear@0 399 printf("max compute units: %u\n", di.units);
nuclear@0 400 printf("max clock frequency: %u\n", di.clock);
nuclear@0 401 printf("max work item dimensions: %u\n", di.dim);
nuclear@0 402
nuclear@0 403 printf("max work item sizes: ");
nuclear@0 404 for(j=0; j<di.dim; j++) {
nuclear@0 405 printf("%u", (unsigned int)di.work_item_sizes[j]);
nuclear@0 406 if(di.dim - j > 1) {
nuclear@0 407 printf(", ");
nuclear@0 408 }
nuclear@0 409 }
nuclear@0 410 putchar('\n');
nuclear@0 411
nuclear@0 412 printf("max work group size: %u\n", (unsigned int)di.work_group_size);
nuclear@0 413 printf("max object allocation size: ");
nuclear@0 414 print_memsize(stdout, di.mem_size);
nuclear@0 415 putchar('\n');
nuclear@0 416
nuclear@0 417 if(devcmp(&di, dev_inf) > 0) {
nuclear@0 418 free(dev_inf->work_item_sizes);
nuclear@0 419 memcpy(dev_inf, &di, sizeof di);
nuclear@0 420 sel = i;
nuclear@0 421 }
nuclear@0 422 }
nuclear@0 423
nuclear@0 424 if(num_dev) {
nuclear@0 425 printf("\nusing device: %d\n", sel);
nuclear@0 426 return 0;
nuclear@0 427 }
nuclear@0 428
nuclear@0 429 return -1;
nuclear@0 430 }
nuclear@0 431
nuclear@0 432 static int get_dev_info(cl_device_id dev, struct device_info *di)
nuclear@0 433 {
nuclear@0 434 di->id = dev;
nuclear@0 435
nuclear@0 436
nuclear@0 437 clGetDeviceInfo(dev, CL_DEVICE_TYPE, sizeof di->type, &di->type, 0);
nuclear@0 438 clGetDeviceInfo(dev, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof di->units, &di->units, 0);
nuclear@0 439 clGetDeviceInfo(dev, CL_DEVICE_MAX_CLOCK_FREQUENCY, sizeof di->clock, &di->clock, 0);
nuclear@0 440 clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, sizeof di->dim, &di->dim, 0);
nuclear@0 441
nuclear@0 442 di->work_item_sizes = new size_t[di->dim];
nuclear@0 443
nuclear@0 444 clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_ITEM_SIZES, di->dim * sizeof *di->work_item_sizes, di->work_item_sizes, 0);
nuclear@0 445 clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof di->work_group_size, &di->work_group_size, 0);
nuclear@0 446 clGetDeviceInfo(dev, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof di->mem_size, &di->mem_size, 0);
nuclear@0 447
nuclear@0 448 return 0;
nuclear@0 449 }
nuclear@0 450
nuclear@0 451 static int devcmp(struct device_info *a, struct device_info *b)
nuclear@0 452 {
nuclear@0 453 unsigned int aval = a->units * a->clock;
nuclear@0 454 unsigned int bval = b->units * b->clock;
nuclear@0 455
nuclear@0 456 return aval - bval;
nuclear@0 457 }
nuclear@0 458
nuclear@0 459 static const char *devtypestr(cl_device_type type)
nuclear@0 460 {
nuclear@0 461 switch(type) {
nuclear@0 462 case CL_DEVICE_TYPE_CPU:
nuclear@0 463 return "cpu";
nuclear@0 464 case CL_DEVICE_TYPE_GPU:
nuclear@0 465 return "gpu";
nuclear@0 466 case CL_DEVICE_TYPE_ACCELERATOR:
nuclear@0 467 return "accelerator";
nuclear@0 468 default:
nuclear@0 469 break;
nuclear@0 470 }
nuclear@0 471 return "unknown";
nuclear@0 472 }
nuclear@0 473
nuclear@0 474 static void print_memsize(FILE *out, unsigned long bytes)
nuclear@0 475 {
nuclear@0 476 int i;
nuclear@0 477 unsigned long memsz = bytes;
nuclear@0 478 const char *suffix[] = {"bytes", "kb", "mb", "gb", "tb", "pb", 0};
nuclear@0 479
nuclear@0 480 for(i=0; suffix[i]; i++) {
nuclear@0 481 if(memsz < 1024) {
nuclear@0 482 fprintf(out, "%lu %s", memsz, suffix[i]);
nuclear@0 483 if(i > 0) {
nuclear@0 484 fprintf(out, " (%lu bytes)", bytes);
nuclear@0 485 }
nuclear@0 486 return;
nuclear@0 487 }
nuclear@0 488
nuclear@0 489 memsz /= 1024;
nuclear@0 490 }
nuclear@0 491 }
nuclear@8 492
nuclear@8 493 static const char *clstrerror(int err)
nuclear@8 494 {
nuclear@8 495 if(err > 0) {
nuclear@8 496 return "<invalid error code>";
nuclear@8 497 }
nuclear@8 498 if(err <= -(int)(sizeof ocl_errstr / sizeof *ocl_errstr)) {
nuclear@8 499 return "<unknown error>";
nuclear@8 500 }
nuclear@8 501 return ocl_errstr[-err];
nuclear@8 502 }