clray

view src/ocl.cc @ 13:407935b73af3

bollocks
author John Tsiombikas <nuclear@member.fsf.org>
date Wed, 04 Aug 2010 04:51:06 +0100
parents 85fd61f374d9
children 29f9330cfa4b
line source
1 #define OCL_CC_
3 #include <stdio.h>
4 #include <stdlib.h>
5 #include <string.h>
6 #include <stdarg.h>
7 #include <errno.h>
8 #ifndef _MSC_VER
9 #include <alloca.h>
10 #else
11 #include <malloc.h>
12 #endif
13 #include <sys/stat.h>
14 #include "ocl.h"
15 #include "ocl_errstr.h"
18 class InitCL {
19 public:
20 InitCL();
21 };
23 struct device_info {
24 cl_device_id id;
25 cl_device_type type;
26 unsigned int units;
27 unsigned int clock;
29 unsigned int dim;
30 size_t *work_item_sizes;
31 size_t work_group_size;
33 unsigned long mem_size;
34 };
36 static bool init_opencl(void);
37 static int select_device(struct device_info *di, int (*devcmp)(struct device_info*, struct device_info*));
38 static int get_dev_info(cl_device_id dev, struct device_info *di);
39 static int devcmp(struct device_info *a, struct device_info *b);
40 static const char *devtypestr(cl_device_type type);
41 static void print_memsize(FILE *out, unsigned long memsz);
42 static const char *clstrerror(int err);
45 static InitCL initcl;
46 static cl_context ctx;
47 static cl_command_queue cmdq;
48 static device_info devinf;
50 InitCL::InitCL()
51 {
52 if(!init_opencl()) {
53 exit(0);
54 }
55 }
57 static bool init_opencl(void)
58 {
59 if(select_device(&devinf, devcmp) == -1) {
60 return false;
61 }
64 if(!(ctx = clCreateContext(0, 1, &devinf.id, 0, 0, 0))) {
65 fprintf(stderr, "failed to create opencl context\n");
66 return false;
67 }
69 if(!(cmdq = clCreateCommandQueue(ctx, devinf.id, 0, 0))) {
70 fprintf(stderr, "failed to create command queue\n");
71 return false;
72 }
73 return true;
74 }
77 CLMemBuffer *create_mem_buffer(int rdwr, size_t sz, void *buf)
78 {
79 int err;
80 cl_mem mem;
81 cl_mem_flags flags = rdwr | CL_MEM_ALLOC_HOST_PTR;
83 if(buf) {
84 flags |= CL_MEM_COPY_HOST_PTR;
85 }
88 if(!(mem = clCreateBuffer(ctx, flags, sz, buf, &err))) {
89 fprintf(stderr, "failed to create memory buffer: %s\n", clstrerror(err));
90 return 0;
91 }
93 CLMemBuffer *mbuf = new CLMemBuffer;
94 mbuf->mem = mem;
95 mbuf->size = sz;
96 mbuf->ptr = 0;
97 return mbuf;
98 }
100 void destroy_mem_buffer(CLMemBuffer *mbuf)
101 {
102 if(mbuf) {
103 clReleaseMemObject(mbuf->mem);
104 delete mbuf;
105 }
106 }
108 void *map_mem_buffer(CLMemBuffer *mbuf, int rdwr)
109 {
110 if(!mbuf) return 0;
112 #ifndef NDEBUG
113 if(mbuf->ptr) {
114 fprintf(stderr, "WARNING: map_mem_buffer called on already mapped buffer\n");
115 }
116 #endif
118 int err;
119 mbuf->ptr = clEnqueueMapBuffer(cmdq, mbuf->mem, 1, rdwr, 0, mbuf->size, 0, 0, 0, &err);
120 if(!mbuf->ptr) {
121 fprintf(stderr, "failed to map buffer: %s\n", clstrerror(err));
122 return 0;
123 }
124 return mbuf->ptr;
125 }
127 void unmap_mem_buffer(CLMemBuffer *mbuf)
128 {
129 if(!mbuf || !mbuf->ptr) return;
130 clEnqueueUnmapMemObject(cmdq, mbuf->mem, mbuf->ptr, 0, 0, 0);
131 mbuf->ptr = 0;
132 }
134 bool write_mem_buffer(CLMemBuffer *mbuf, size_t sz, void *src)
135 {
136 if(!mbuf) return false;
138 int err;
139 if((err = clEnqueueWriteBuffer(cmdq, mbuf->mem, 1, 0, sz, src, 0, 0, 0)) != 0) {
140 fprintf(stderr, "failed to write buffer: %s\n", clstrerror(err));
141 return false;
142 }
143 return true;
144 }
146 bool read_mem_buffer(CLMemBuffer *mbuf, size_t sz, void *dest)
147 {
148 if(!mbuf) return false;
150 int err;
151 if((err = clEnqueueReadBuffer(cmdq, mbuf->mem, 1, 0, sz, dest, 0, 0, 0)) != 0) {
152 fprintf(stderr, "failed to read buffer: %s\n", clstrerror(err));
153 return false;
154 }
155 return true;
156 }
159 CLProgram::CLProgram(const char *kname)
160 {
161 prog = 0;
162 kernel = 0;
163 this->kname = kname;
164 args.resize(16);
165 built = false;
166 }
168 CLProgram::~CLProgram()
169 {
170 if(prog) {
172 clReleaseProgram(prog);
173 }
174 if(kernel) {
176 clReleaseKernel(kernel);
177 }
178 for(size_t i=0; i<args.size(); i++) {
179 if(args[i].type == ARGTYPE_MEM_BUF) {
180 destroy_mem_buffer(args[i].v.mbuf);
181 }
182 }
183 }
185 bool CLProgram::load(const char *fname)
186 {
187 FILE *fp;
188 char *src;
189 struct stat st;
191 printf("loading opencl program (%s)\n", fname);
193 if(!(fp = fopen(fname, "rb"))) {
194 fprintf(stderr, "failed to open %s: %s\n", fname, strerror(errno));
195 return false;
196 }
198 fstat(fileno(fp), &st);
200 src = new char[st.st_size + 1];
202 fread(src, 1, st.st_size, fp);
203 src[st.st_size] = 0;
204 fclose(fp);
207 if(!(prog = clCreateProgramWithSource(ctx, 1, (const char**)&src, 0, 0))) {
208 fprintf(stderr, "error creating program object: %s\n", fname);
209 delete [] src;
210 return false;
211 }
212 delete [] src;
213 return true;
214 }
216 bool CLProgram::set_argi(int idx, int val)
217 {
218 if((int)args.size() <= idx) {
219 args.resize(idx + 1);
220 }
222 CLArg *arg = &args[idx];
223 arg->type = ARGTYPE_INT;
224 arg->v.ival = val;
225 return true;
226 }
228 bool CLProgram::set_argf(int idx, float val)
229 {
230 if((int)args.size() <= idx) {
231 args.resize(idx + 1);
232 }
234 CLArg *arg = &args[idx];
235 arg->type = ARGTYPE_FLOAT;
236 arg->v.fval = val;
237 return true;
238 }
240 bool CLProgram::set_arg_buffer(int idx, int rdwr, size_t sz, void *ptr)
241 {
242 printf("create argument %d buffer: %d bytes\n", idx, (int)sz);
243 CLMemBuffer *buf;
245 if(!(buf = create_mem_buffer(rdwr, sz, ptr))) {
246 return false;
247 }
249 if((int)args.size() <= idx) {
250 args.resize(idx + 1);
251 }
252 args[idx].type = ARGTYPE_MEM_BUF;
253 args[idx].v.mbuf = buf;
254 return true;
255 }
257 CLMemBuffer *CLProgram::get_arg_buffer(int arg)
258 {
259 if(arg < 0 || arg >= (int)args.size() || args[arg].type != ARGTYPE_MEM_BUF) {
260 return 0;
261 }
262 return args[arg].v.mbuf;
263 }
265 bool CLProgram::build()
266 {
267 int err;
269 if((err = clBuildProgram(prog, 0, 0, 0, 0, 0)) != 0) {
270 size_t sz;
271 clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, 0, 0, &sz);
273 char *errlog = (char*)alloca(sz + 1);
274 clGetProgramBuildInfo(prog, devinf.id, CL_PROGRAM_BUILD_LOG, sz, errlog, 0);
275 fprintf(stderr, "failed to build program: %s\n%s\n", clstrerror(err), errlog);
277 clReleaseProgram(prog);
278 prog = 0;
279 return false;
280 }
283 if(!(kernel = clCreateKernel(prog, kname.c_str(), 0))) {
284 fprintf(stderr, "failed to create kernel: %s\n", kname.c_str());
285 clReleaseProgram(prog);
286 prog = 0;
287 return false;
288 }
290 for(size_t i=0; i<args.size(); i++) {
291 int err;
293 if(args[i].type == ARGTYPE_NONE) {
294 break;
295 }
297 switch(args[i].type) {
298 case ARGTYPE_INT:
299 if((err = clSetKernelArg(kernel, i, sizeof(int), &args[i].v.ival)) != 0) {
300 fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
301 goto fail;
302 }
303 break;
305 case ARGTYPE_FLOAT:
306 if((err = clSetKernelArg(kernel, i, sizeof(float), &args[i].v.fval)) != 0) {
307 fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
308 goto fail;
309 }
310 break;
312 case ARGTYPE_MEM_BUF:
313 {
314 CLMemBuffer *mbuf = args[i].v.mbuf;
316 if((err = clSetKernelArg(kernel, i, sizeof mbuf->mem, &mbuf->mem)) != 0) {
317 fprintf(stderr, "failed to bind kernel argument %d: %s\n", (int)i, clstrerror(err));
318 goto fail;
319 }
320 }
321 break;
323 default:
324 break;
325 }
326 }
328 built = true;
329 return true;
331 fail:
332 clReleaseProgram(prog);
333 clReleaseKernel(kernel);
334 prog = 0;
335 kernel = 0;
336 return false;
337 }
339 bool CLProgram::run() const
340 {
341 return run(1, 1);
342 }
344 bool CLProgram::run(int dim, ...) const
345 {
346 if(!built) {
347 if(!((CLProgram*)this)->build()) {
348 return false;
349 }
350 }
352 va_list ap;
353 size_t *global_size = (size_t*)alloca(dim * sizeof *global_size);
355 va_start(ap, dim);
356 for(int i=0; i<dim; i++) {
357 global_size[i] = va_arg(ap, int);
358 }
359 va_end(ap);
361 int err;
362 if((err = clEnqueueNDRangeKernel(cmdq, kernel, dim, 0, global_size, 0, 0, 0, 0)) != 0) {
363 fprintf(stderr, "error executing kernel: %s\n", clstrerror(err));
364 return false;
365 }
366 return true;
367 }
369 static int select_device(struct device_info *dev_inf, int (*devcmp)(struct device_info*, struct device_info*))
370 {
371 unsigned int i, j, num_dev, num_plat, sel, ret;
372 cl_device_id dev[32];
373 cl_platform_id plat[32];
375 dev_inf->work_item_sizes = 0;
377 if((ret = clGetPlatformIDs(32, plat, &num_plat)) != 0) {
378 fprintf(stderr, "clGetPlatformIDs failed: %s\n", clstrerror(ret));
379 return -1;
380 }
381 if(!num_plat) {
382 fprintf(stderr, "OpenCL not available!\n");
383 return -1;
384 }
386 for(i=0; i<num_plat; i++) {
387 char buf[512];
389 clGetPlatformInfo(plat[i], CL_PLATFORM_NAME, sizeof buf, buf, 0);
390 printf("[%d]: %s", i, buf);
391 clGetPlatformInfo(plat[i], CL_PLATFORM_VENDOR, sizeof buf, buf, 0);
392 printf(", %s", buf);
393 clGetPlatformInfo(plat[i], CL_PLATFORM_VERSION, sizeof buf, buf, 0);
394 printf(" (%s)\n", buf);
395 }
397 if((ret = clGetDeviceIDs(plat[0], CL_DEVICE_TYPE_ALL, 32, dev, &num_dev)) != 0) {
398 fprintf(stderr, "clGetDeviceIDs failed: %s\n", clstrerror(ret));
399 return -1;
400 }
401 printf("found %d cl devices.\n", num_dev);
403 for(i=0; i<num_dev; i++) {
404 struct device_info di;
406 if(get_dev_info(dev[i], &di) == -1) {
407 free(dev_inf->work_item_sizes);
408 return -1;
409 }
411 printf("--> device %u (%s)\n", i, devtypestr(di.type));
412 printf("max compute units: %u\n", di.units);
413 printf("max clock frequency: %u\n", di.clock);
414 printf("max work item dimensions: %u\n", di.dim);
416 printf("max work item sizes: ");
417 for(j=0; j<di.dim; j++) {
418 printf("%u", (unsigned int)di.work_item_sizes[j]);
419 if(di.dim - j > 1) {
420 printf(", ");
421 }
422 }
423 putchar('\n');
425 printf("max work group size: %u\n", (unsigned int)di.work_group_size);
426 printf("max object allocation size: ");
427 print_memsize(stdout, di.mem_size);
428 putchar('\n');
430 if(devcmp(&di, dev_inf) > 0) {
431 free(dev_inf->work_item_sizes);
432 memcpy(dev_inf, &di, sizeof di);
433 sel = i;
434 }
435 }
437 if(num_dev) {
438 printf("\nusing device: %d\n", sel);
439 return 0;
440 }
442 return -1;
443 }
445 static int get_dev_info(cl_device_id dev, struct device_info *di)
446 {
447 di->id = dev;
450 clGetDeviceInfo(dev, CL_DEVICE_TYPE, sizeof di->type, &di->type, 0);
451 clGetDeviceInfo(dev, CL_DEVICE_MAX_COMPUTE_UNITS, sizeof di->units, &di->units, 0);
452 clGetDeviceInfo(dev, CL_DEVICE_MAX_CLOCK_FREQUENCY, sizeof di->clock, &di->clock, 0);
453 clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS, sizeof di->dim, &di->dim, 0);
455 di->work_item_sizes = new size_t[di->dim];
457 clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_ITEM_SIZES, di->dim * sizeof *di->work_item_sizes, di->work_item_sizes, 0);
458 clGetDeviceInfo(dev, CL_DEVICE_MAX_WORK_GROUP_SIZE, sizeof di->work_group_size, &di->work_group_size, 0);
459 clGetDeviceInfo(dev, CL_DEVICE_MAX_MEM_ALLOC_SIZE, sizeof di->mem_size, &di->mem_size, 0);
461 return 0;
462 }
464 static int devcmp(struct device_info *a, struct device_info *b)
465 {
466 unsigned int aval = a->units * a->clock;
467 unsigned int bval = b->units * b->clock;
469 return aval - bval;
470 }
472 static const char *devtypestr(cl_device_type type)
473 {
474 switch(type) {
475 case CL_DEVICE_TYPE_CPU:
476 return "cpu";
477 case CL_DEVICE_TYPE_GPU:
478 return "gpu";
479 case CL_DEVICE_TYPE_ACCELERATOR:
480 return "accelerator";
481 default:
482 break;
483 }
484 return "unknown";
485 }
487 static void print_memsize(FILE *out, unsigned long bytes)
488 {
489 int i;
490 unsigned long memsz = bytes;
491 const char *suffix[] = {"bytes", "kb", "mb", "gb", "tb", "pb", 0};
493 for(i=0; suffix[i]; i++) {
494 if(memsz < 1024) {
495 fprintf(out, "%lu %s", memsz, suffix[i]);
496 if(i > 0) {
497 fprintf(out, " (%lu bytes)", bytes);
498 }
499 return;
500 }
502 memsz /= 1024;
503 }
504 }
506 static const char *clstrerror(int err)
507 {
508 if(err > 0) {
509 return "<invalid error code>";
510 }
511 if(err <= -(int)(sizeof ocl_errstr / sizeof *ocl_errstr)) {
512 return "<unknown error>";
513 }
514 return ocl_errstr[-err];
515 }