Merging r50522 through r50572 from trunk into soc-2011-tomato
[blender.git] / intern / cycles / device / device_opencl.cpp
1 /*
2  * Copyright 2011, Blender Foundation.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU General Public License
6  * as published by the Free Software Foundation; either version 2
7  * of the License, or (at your option) any later version.
8  *
9  * This program is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
12  * GNU General Public License for more details.
13  *
14  * You should have received a copy of the GNU General Public License
15  * along with this program; if not, write to the Free Software Foundation,
16  * Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA.
17  */
18
19 #ifdef WITH_OPENCL
20
21 #include <stdio.h>
22 #include <stdlib.h>
23 #include <string.h>
24
25 #include "device.h"
26 #include "device_intern.h"
27
28 #include "buffers.h"
29
30 #include "util_foreach.h"
31 #include "util_map.h"
32 #include "util_math.h"
33 #include "util_md5.h"
34 #include "util_opencl.h"
35 #include "util_opengl.h"
36 #include "util_path.h"
37 #include "util_time.h"
38
39 CCL_NAMESPACE_BEGIN
40
41 #define CL_MEM_PTR(p) ((cl_mem)(unsigned long)(p))
42
43 class OpenCLDevice : public Device
44 {
45 public:
46         TaskPool task_pool;
47         cl_context cxContext;
48         cl_command_queue cqCommandQueue;
49         cl_platform_id cpPlatform;
50         cl_device_id cdDevice;
51         cl_program cpProgram;
52         cl_kernel ckPathTraceKernel;
53         cl_kernel ckFilmConvertKernel;
54         cl_int ciErr;
55         map<string, device_vector<uchar>*> const_mem_map;
56         map<string, device_memory*> mem_map;
57         device_ptr null_mem;
58         bool device_initialized;
59         string platform_name;
60
61         const char *opencl_error_string(cl_int err)
62         {
63                 switch (err) {
64                         case CL_SUCCESS: return "Success!";
65                         case CL_DEVICE_NOT_FOUND: return "Device not found.";
66                         case CL_DEVICE_NOT_AVAILABLE: return "Device not available";
67                         case CL_COMPILER_NOT_AVAILABLE: return "Compiler not available";
68                         case CL_MEM_OBJECT_ALLOCATION_FAILURE: return "Memory object allocation failure";
69                         case CL_OUT_OF_RESOURCES: return "Out of resources";
70                         case CL_OUT_OF_HOST_MEMORY: return "Out of host memory";
71                         case CL_PROFILING_INFO_NOT_AVAILABLE: return "Profiling information not available";
72                         case CL_MEM_COPY_OVERLAP: return "Memory copy overlap";
73                         case CL_IMAGE_FORMAT_MISMATCH: return "Image format mismatch";
74                         case CL_IMAGE_FORMAT_NOT_SUPPORTED: return "Image format not supported";
75                         case CL_BUILD_PROGRAM_FAILURE: return "Program build failure";
76                         case CL_MAP_FAILURE: return "Map failure";
77                         case CL_INVALID_VALUE: return "Invalid value";
78                         case CL_INVALID_DEVICE_TYPE: return "Invalid device type";
79                         case CL_INVALID_PLATFORM: return "Invalid platform";
80                         case CL_INVALID_DEVICE: return "Invalid device";
81                         case CL_INVALID_CONTEXT: return "Invalid context";
82                         case CL_INVALID_QUEUE_PROPERTIES: return "Invalid queue properties";
83                         case CL_INVALID_COMMAND_QUEUE: return "Invalid command queue";
84                         case CL_INVALID_HOST_PTR: return "Invalid host pointer";
85                         case CL_INVALID_MEM_OBJECT: return "Invalid memory object";
86                         case CL_INVALID_IMAGE_FORMAT_DESCRIPTOR: return "Invalid image format descriptor";
87                         case CL_INVALID_IMAGE_SIZE: return "Invalid image size";
88                         case CL_INVALID_SAMPLER: return "Invalid sampler";
89                         case CL_INVALID_BINARY: return "Invalid binary";
90                         case CL_INVALID_BUILD_OPTIONS: return "Invalid build options";
91                         case CL_INVALID_PROGRAM: return "Invalid program";
92                         case CL_INVALID_PROGRAM_EXECUTABLE: return "Invalid program executable";
93                         case CL_INVALID_KERNEL_NAME: return "Invalid kernel name";
94                         case CL_INVALID_KERNEL_DEFINITION: return "Invalid kernel definition";
95                         case CL_INVALID_KERNEL: return "Invalid kernel";
96                         case CL_INVALID_ARG_INDEX: return "Invalid argument index";
97                         case CL_INVALID_ARG_VALUE: return "Invalid argument value";
98                         case CL_INVALID_ARG_SIZE: return "Invalid argument size";
99                         case CL_INVALID_KERNEL_ARGS: return "Invalid kernel arguments";
100                         case CL_INVALID_WORK_DIMENSION: return "Invalid work dimension";
101                         case CL_INVALID_WORK_GROUP_SIZE: return "Invalid work group size";
102                         case CL_INVALID_WORK_ITEM_SIZE: return "Invalid work item size";
103                         case CL_INVALID_GLOBAL_OFFSET: return "Invalid global offset";
104                         case CL_INVALID_EVENT_WAIT_LIST: return "Invalid event wait list";
105                         case CL_INVALID_EVENT: return "Invalid event";
106                         case CL_INVALID_OPERATION: return "Invalid operation";
107                         case CL_INVALID_GL_OBJECT: return "Invalid OpenGL object";
108                         case CL_INVALID_BUFFER_SIZE: return "Invalid buffer size";
109                         case CL_INVALID_MIP_LEVEL: return "Invalid mip-map level";
110                         default: return "Unknown";
111                 }
112         }
113
114         bool opencl_error(cl_int err)
115         {
116                 if(err != CL_SUCCESS) {
117                         string message = string_printf("OpenCL error (%d): %s", err, opencl_error_string(err));
118                         if(error_msg == "")
119                                 error_msg = message;
120                         fprintf(stderr, "%s\n", message.c_str());
121                         return true;
122                 }
123
124                 return false;
125         }
126
127         void opencl_error(const string& message)
128         {
129                 if(error_msg == "")
130                         error_msg = message;
131                 fprintf(stderr, "%s\n", message.c_str());
132         }
133
134         void opencl_assert(cl_int err)
135         {
136                 if(err != CL_SUCCESS) {
137                         string message = string_printf("OpenCL error (%d): %s", err, opencl_error_string(err));
138                         if(error_msg == "")
139                                 error_msg = message;
140                         fprintf(stderr, "%s\n", message.c_str());
141 #ifndef NDEBUG
142                         abort();
143 #endif
144                 }
145         }
146
147         OpenCLDevice(DeviceInfo& info, bool background_)
148         {
149                 background = background_;
150                 cpPlatform = NULL;
151                 cxContext = NULL;
152                 cqCommandQueue = NULL;
153                 cpProgram = NULL;
154                 ckPathTraceKernel = NULL;
155                 ckFilmConvertKernel = NULL;
156                 null_mem = 0;
157                 device_initialized = false;
158
159                 /* setup platform */
160                 cl_uint num_platforms;
161
162                 ciErr = clGetPlatformIDs(0, NULL, &num_platforms);
163                 if(opencl_error(ciErr))
164                         return;
165
166                 if(num_platforms == 0) {
167                         opencl_error("OpenCL: no platforms found.");
168                         return;
169                 }
170
171                 ciErr = clGetPlatformIDs(1, &cpPlatform, NULL);
172                 if(opencl_error(ciErr))
173                         return;
174
175                 char name[256];
176                 clGetPlatformInfo(cpPlatform, CL_PLATFORM_NAME, sizeof(name), &name, NULL);
177                 platform_name = name;
178
179                 /* get devices */
180                 vector<cl_device_id> device_ids;
181                 cl_uint num_devices;
182
183                 if(opencl_error(clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU|CL_DEVICE_TYPE_ACCELERATOR, 0, NULL, &num_devices)))
184                         return;
185
186                 if(info.num > num_devices) {
187                         if(num_devices == 0)
188                                 opencl_error("OpenCL: no devices found.");
189                         else
190                                 opencl_error("OpenCL: specified device not found.");
191                         return;
192                 }
193
194                 device_ids.resize(num_devices);
195                 
196                 if(opencl_error(clGetDeviceIDs(cpPlatform, CL_DEVICE_TYPE_GPU|CL_DEVICE_TYPE_ACCELERATOR, num_devices, &device_ids[0], NULL)))
197                         return;
198
199                 cdDevice = device_ids[info.num];
200
201                 /* create context */
202                 cxContext = clCreateContext(0, 1, &cdDevice, NULL, NULL, &ciErr);
203                 if(opencl_error(ciErr))
204                         return;
205
206                 cqCommandQueue = clCreateCommandQueue(cxContext, cdDevice, 0, &ciErr);
207                 if(opencl_error(ciErr))
208                         return;
209
210                 null_mem = (device_ptr)clCreateBuffer(cxContext, CL_MEM_READ_ONLY, 1, NULL, &ciErr);
211                 device_initialized = true;
212         }
213
214         bool opencl_version_check()
215         {
216                 char version[256];
217
218                 int major, minor, req_major = 1, req_minor = 0;
219
220                 clGetPlatformInfo(cpPlatform, CL_PLATFORM_VERSION, sizeof(version), &version, NULL);
221
222                 if(sscanf(version, "OpenCL %d.%d", &major, &minor) < 2) {
223                         opencl_error(string_printf("OpenCL: failed to parse platform version string (%s).", version));
224                         return false;
225                 }
226
227                 if(!((major == req_major && minor >= req_minor) || (major > req_major))) {
228                         opencl_error(string_printf("OpenCL: platform version 1.1 or later required, found %d.%d", major, minor));
229                         return false;
230                 }
231
232                 clGetDeviceInfo(cdDevice, CL_DEVICE_OPENCL_C_VERSION, sizeof(version), &version, NULL);
233
234                 if(sscanf(version, "OpenCL C %d.%d", &major, &minor) < 2) {
235                         opencl_error(string_printf("OpenCL: failed to parse OpenCL C version string (%s).", version));
236                         return false;
237                 }
238
239                 if(!((major == req_major && minor >= req_minor) || (major > req_major))) {
240                         opencl_error(string_printf("OpenCL: C version 1.1 or later required, found %d.%d", major, minor));
241                         return false;
242                 }
243
244                 /* we don't check CL_DEVICE_VERSION since for e.g. nvidia sm 1.3 cards this is
245                  * 1.0 even if the language features are there, just limited shared memory */
246
247                 return true;
248         }
249
250         bool load_binary(const string& kernel_path, const string& clbin)
251         {
252                 /* read binary into memory */
253                 vector<uint8_t> binary;
254
255                 if(!path_read_binary(clbin, binary)) {
256                         opencl_error(string_printf("OpenCL failed to read cached binary %s.", clbin.c_str()));
257                         return false;
258                 }
259
260                 /* create program */
261                 cl_int status;
262                 size_t size = binary.size();
263                 const uint8_t *bytes = &binary[0];
264
265                 cpProgram = clCreateProgramWithBinary(cxContext, 1, &cdDevice,
266                         &size, &bytes, &status, &ciErr);
267
268                 if(opencl_error(status) || opencl_error(ciErr)) {
269                         opencl_error(string_printf("OpenCL failed create program from cached binary %s.", clbin.c_str()));
270                         return false;
271                 }
272
273                 if(!build_kernel(kernel_path))
274                         return false;
275
276                 return true;
277         }
278
279         bool save_binary(const string& clbin)
280         {
281                 size_t size = 0;
282                 clGetProgramInfo(cpProgram, CL_PROGRAM_BINARY_SIZES, sizeof(size_t), &size, NULL);
283
284                 if(!size)
285                         return false;
286
287                 vector<uint8_t> binary(size);
288                 uint8_t *bytes = &binary[0];
289
290                 clGetProgramInfo(cpProgram, CL_PROGRAM_BINARIES, sizeof(uint8_t*), &bytes, NULL);
291
292                 if(!path_write_binary(clbin, binary)) {
293                         opencl_error(string_printf("OpenCL failed to write cached binary %s.", clbin.c_str()));
294                         return false;
295                 }
296
297                 return true;
298         }
299
300         string kernel_build_options()
301         {
302                 string build_options = " -cl-fast-relaxed-math ";
303                 
304                 if(platform_name == "NVIDIA CUDA")
305                         build_options += "-D__KERNEL_SHADING__ -D__KERNEL_OPENCL_NVIDIA__ -cl-nv-maxrregcount=24 -cl-nv-verbose ";
306
307                 else if(platform_name == "Apple")
308                         build_options += "-D__CL_NO_FLOAT3__ -D__KERNEL_OPENCL_APPLE__ ";
309
310                 else if(platform_name == "AMD Accelerated Parallel Processing")
311                         build_options += "-D__CL_NO_FLOAT3__ -D__KERNEL_OPENCL_AMD__ ";
312
313                 return build_options;
314         }
315
316         bool build_kernel(const string& kernel_path)
317         {
318                 string build_options = kernel_build_options();
319         
320                 ciErr = clBuildProgram(cpProgram, 0, NULL, build_options.c_str(), NULL, NULL);
321
322                 if(ciErr != CL_SUCCESS) {
323                         /* show build errors */
324                         char *build_log;
325                         size_t ret_val_size;
326
327                         clGetProgramBuildInfo(cpProgram, cdDevice, CL_PROGRAM_BUILD_LOG, 0, NULL, &ret_val_size);
328
329                         build_log = new char[ret_val_size+1];
330                         clGetProgramBuildInfo(cpProgram, cdDevice, CL_PROGRAM_BUILD_LOG, ret_val_size, build_log, NULL);
331
332                         build_log[ret_val_size] = '\0';
333                         opencl_error("OpenCL build failed: errors in console");
334                         fprintf(stderr, "%s\n", build_log);
335
336                         delete[] build_log;
337
338                         return false;
339                 }
340
341                 return true;
342         }
343
344         bool compile_kernel(const string& kernel_path, const string& kernel_md5)
345         {
346                 /* we compile kernels consisting of many files. unfortunately opencl
347                  * kernel caches do not seem to recognize changes in included files.
348                  * so we force recompile on changes by adding the md5 hash of all files */
349                 string source = "#include \"kernel.cl\" // " + kernel_md5 + "\n";
350                 source = path_source_replace_includes(source, kernel_path);
351
352                 size_t source_len = source.size();
353                 const char *source_str = source.c_str();
354
355                 cpProgram = clCreateProgramWithSource(cxContext, 1, &source_str, &source_len, &ciErr);
356
357                 if(opencl_error(ciErr))
358                         return false;
359
360                 double starttime = time_dt();
361                 printf("Compiling OpenCL kernel ...\n");
362
363                 if(!build_kernel(kernel_path))
364                         return false;
365
366                 printf("Kernel compilation finished in %.2lfs.\n", time_dt() - starttime);
367
368                 return true;
369         }
370
371         string device_md5_hash()
372         {
373                 MD5Hash md5;
374                 char version[256], driver[256], name[256], vendor[256];
375
376                 clGetPlatformInfo(cpPlatform, CL_PLATFORM_VENDOR, sizeof(vendor), &vendor, NULL);
377                 clGetDeviceInfo(cdDevice, CL_DEVICE_VERSION, sizeof(version), &version, NULL);
378                 clGetDeviceInfo(cdDevice, CL_DEVICE_NAME, sizeof(name), &name, NULL);
379                 clGetDeviceInfo(cdDevice, CL_DRIVER_VERSION, sizeof(driver), &driver, NULL);
380
381                 md5.append((uint8_t*)vendor, strlen(vendor));
382                 md5.append((uint8_t*)version, strlen(version));
383                 md5.append((uint8_t*)name, strlen(name));
384                 md5.append((uint8_t*)driver, strlen(driver));
385
386                 string options = kernel_build_options();
387                 md5.append((uint8_t*)options.c_str(), options.size());
388
389                 return md5.get_hex();
390         }
391
392         bool load_kernels(bool experimental)
393         {
394                 /* verify if device was initialized */
395                 if(!device_initialized) {
396                         fprintf(stderr, "OpenCL: failed to initialize device.\n");
397                         return false;
398                 }
399
400                 /* verify we have right opencl version */
401                 if(!opencl_version_check())
402                         return false;
403
404                 /* md5 hash to detect changes */
405                 string kernel_path = path_get("kernel");
406                 string kernel_md5 = path_files_md5_hash(kernel_path);
407                 string device_md5 = device_md5_hash();
408
409                 /* try to use cache binary */
410                 string clbin = string_printf("cycles_kernel_%s_%s.clbin", device_md5.c_str(), kernel_md5.c_str());
411                 clbin = path_user_get(path_join("cache", clbin));
412
413                 if(path_exists(clbin)) {
414                         /* if exists already, try use it */
415                         if(!load_binary(kernel_path, clbin))
416                                 return false;
417                 }
418                 else {
419                         /* compile kernel */
420                         if(!compile_kernel(kernel_path, kernel_md5))
421                                 return false;
422
423                         /* save binary for reuse */
424                         save_binary(clbin);
425                 }
426
427                 /* find kernels */
428                 ckPathTraceKernel = clCreateKernel(cpProgram, "kernel_ocl_path_trace", &ciErr);
429                 if(opencl_error(ciErr))
430                         return false;
431
432                 ckFilmConvertKernel = clCreateKernel(cpProgram, "kernel_ocl_tonemap", &ciErr);
433                 if(opencl_error(ciErr))
434                         return false;
435
436                 return true;
437         }
438
439         ~OpenCLDevice()
440         {
441                 task_pool.stop();
442
443                 if(null_mem)
444                         clReleaseMemObject(CL_MEM_PTR(null_mem));
445
446                 map<string, device_vector<uchar>*>::iterator mt;
447                 for(mt = const_mem_map.begin(); mt != const_mem_map.end(); mt++) {
448                         mem_free(*(mt->second));
449                         delete mt->second;
450                 }
451
452                 if(ckPathTraceKernel)
453                         clReleaseKernel(ckPathTraceKernel);  
454                 if(ckFilmConvertKernel)
455                         clReleaseKernel(ckFilmConvertKernel);  
456                 if(cpProgram)
457                         clReleaseProgram(cpProgram);
458                 if(cqCommandQueue)
459                         clReleaseCommandQueue(cqCommandQueue);
460                 if(cxContext)
461                         clReleaseContext(cxContext);
462         }
463
464         void mem_alloc(device_memory& mem, MemoryType type)
465         {
466                 size_t size = mem.memory_size();
467
468                 if(type == MEM_READ_ONLY)
469                         mem.device_pointer = (device_ptr)clCreateBuffer(cxContext, CL_MEM_READ_ONLY, size, NULL, &ciErr);
470                 else if(type == MEM_WRITE_ONLY)
471                         mem.device_pointer = (device_ptr)clCreateBuffer(cxContext, CL_MEM_WRITE_ONLY, size, NULL, &ciErr);
472                 else
473                         mem.device_pointer = (device_ptr)clCreateBuffer(cxContext, CL_MEM_READ_WRITE, size, NULL, &ciErr);
474
475                 opencl_assert(ciErr);
476         }
477
478         void mem_copy_to(device_memory& mem)
479         {
480                 /* this is blocking */
481                 size_t size = mem.memory_size();
482                 ciErr = clEnqueueWriteBuffer(cqCommandQueue, CL_MEM_PTR(mem.device_pointer), CL_TRUE, 0, size, (void*)mem.data_pointer, 0, NULL, NULL);
483                 opencl_assert(ciErr);
484         }
485
486         void mem_copy_from(device_memory& mem, int y, int w, int h, int elem)
487         {
488                 size_t offset = elem*y*w;
489                 size_t size = elem*w*h;
490
491                 ciErr = clEnqueueReadBuffer(cqCommandQueue, CL_MEM_PTR(mem.device_pointer), CL_TRUE, offset, size, (uchar*)mem.data_pointer + offset, 0, NULL, NULL);
492                 opencl_assert(ciErr);
493         }
494
495         void mem_zero(device_memory& mem)
496         {
497                 if(mem.device_pointer) {
498                         memset((void*)mem.data_pointer, 0, mem.memory_size());
499                         mem_copy_to(mem);
500                 }
501         }
502
503         void mem_free(device_memory& mem)
504         {
505                 if(mem.device_pointer) {
506                         ciErr = clReleaseMemObject(CL_MEM_PTR(mem.device_pointer));
507                         mem.device_pointer = 0;
508                         opencl_assert(ciErr);
509                 }
510         }
511
512         void const_copy_to(const char *name, void *host, size_t size)
513         {
514                 if(const_mem_map.find(name) == const_mem_map.end()) {
515                         device_vector<uchar> *data = new device_vector<uchar>();
516                         data->copy((uchar*)host, size);
517
518                         mem_alloc(*data, MEM_READ_ONLY);
519                         const_mem_map[name] = data;
520                 }
521                 else {
522                         device_vector<uchar> *data = const_mem_map[name];
523                         data->copy((uchar*)host, size);
524                 }
525
526                 mem_copy_to(*const_mem_map[name]);
527         }
528
529         void tex_alloc(const char *name, device_memory& mem, bool interpolation, bool periodic)
530         {
531                 mem_alloc(mem, MEM_READ_ONLY);
532                 mem_copy_to(mem);
533                 mem_map[name] = &mem;
534         }
535
536         void tex_free(device_memory& mem)
537         {
538                 if(mem.data_pointer)
539                         mem_free(mem);
540         }
541
542         size_t global_size_round_up(int group_size, int global_size)
543         {
544                 int r = global_size % group_size;
545                 return global_size + ((r == 0)? 0: group_size - r);
546         }
547
548         void path_trace(RenderTile& rtile, int sample)
549         {
550                 /* cast arguments to cl types */
551                 cl_mem d_data = CL_MEM_PTR(const_mem_map["__data"]->device_pointer);
552                 cl_mem d_buffer = CL_MEM_PTR(rtile.buffer);
553                 cl_mem d_rng_state = CL_MEM_PTR(rtile.rng_state);
554                 cl_int d_x = rtile.x;
555                 cl_int d_y = rtile.y;
556                 cl_int d_w = rtile.w;
557                 cl_int d_h = rtile.h;
558                 cl_int d_sample = sample;
559                 cl_int d_offset = rtile.offset;
560                 cl_int d_stride = rtile.stride;
561
562                 /* sample arguments */
563                 int narg = 0;
564                 ciErr = 0;
565
566                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_data), (void*)&d_data);
567                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_buffer), (void*)&d_buffer);
568                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_rng_state), (void*)&d_rng_state);
569
570 #define KERNEL_TEX(type, ttype, name) \
571         ciErr |= set_kernel_arg_mem(ckPathTraceKernel, &narg, #name);
572 #include "kernel_textures.h"
573
574                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_sample), (void*)&d_sample);
575                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_x), (void*)&d_x);
576                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_y), (void*)&d_y);
577                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_w), (void*)&d_w);
578                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_h), (void*)&d_h);
579                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_offset), (void*)&d_offset);
580                 ciErr |= clSetKernelArg(ckPathTraceKernel, narg++, sizeof(d_stride), (void*)&d_stride);
581
582                 opencl_assert(ciErr);
583
584                 size_t workgroup_size;
585
586                 clGetKernelWorkGroupInfo(ckPathTraceKernel, cdDevice,
587                         CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &workgroup_size, NULL);
588         
589                 workgroup_size = max(sqrt((double)workgroup_size), 1.0);
590
591                 size_t local_size[2] = {workgroup_size, workgroup_size};
592                 size_t global_size[2] = {global_size_round_up(local_size[0], d_w), global_size_round_up(local_size[1], d_h)};
593
594                 /* run kernel */
595                 ciErr = clEnqueueNDRangeKernel(cqCommandQueue, ckPathTraceKernel, 2, NULL, global_size, local_size, 0, NULL, NULL);
596                 opencl_assert(ciErr);
597                 opencl_assert(clFinish(cqCommandQueue));
598         }
599
600         cl_int set_kernel_arg_mem(cl_kernel kernel, int *narg, const char *name)
601         {
602                 cl_mem ptr;
603                 cl_int err = 0;
604
605                 if(mem_map.find(name) != mem_map.end()) {
606                         device_memory *mem = mem_map[name];
607                 
608                         ptr = CL_MEM_PTR(mem->device_pointer);
609                 }
610                 else {
611                         /* work around NULL not working, even though the spec says otherwise */
612                         ptr = CL_MEM_PTR(null_mem);
613                 }
614                 
615                 err |= clSetKernelArg(kernel, (*narg)++, sizeof(ptr), (void*)&ptr);
616                 opencl_assert(err);
617
618                 return err;
619         }
620
621         void tonemap(DeviceTask& task, device_ptr buffer, device_ptr rgba)
622         {
623                 /* cast arguments to cl types */
624                 cl_mem d_data = CL_MEM_PTR(const_mem_map["__data"]->device_pointer);
625                 cl_mem d_rgba = CL_MEM_PTR(rgba);
626                 cl_mem d_buffer = CL_MEM_PTR(buffer);
627                 cl_int d_x = task.x;
628                 cl_int d_y = task.y;
629                 cl_int d_w = task.w;
630                 cl_int d_h = task.h;
631                 cl_int d_sample = task.sample;
632                 cl_int d_resolution = task.resolution;
633                 cl_int d_offset = task.offset;
634                 cl_int d_stride = task.stride;
635
636                 /* sample arguments */
637                 int narg = 0;
638                 ciErr = 0;
639
640                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_data), (void*)&d_data);
641                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_rgba), (void*)&d_rgba);
642                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_buffer), (void*)&d_buffer);
643
644 #define KERNEL_TEX(type, ttype, name) \
645         ciErr |= set_kernel_arg_mem(ckFilmConvertKernel, &narg, #name);
646 #include "kernel_textures.h"
647
648                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_sample), (void*)&d_sample);
649                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_resolution), (void*)&d_resolution);
650                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_x), (void*)&d_x);
651                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_y), (void*)&d_y);
652                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_w), (void*)&d_w);
653                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_h), (void*)&d_h);
654                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_offset), (void*)&d_offset);
655                 ciErr |= clSetKernelArg(ckFilmConvertKernel, narg++, sizeof(d_stride), (void*)&d_stride);
656
657                 opencl_assert(ciErr);
658
659                 size_t workgroup_size;
660
661                 clGetKernelWorkGroupInfo(ckFilmConvertKernel, cdDevice,
662                         CL_KERNEL_WORK_GROUP_SIZE, sizeof(size_t), &workgroup_size, NULL);
663         
664                 workgroup_size = max(sqrt((double)workgroup_size), 1.0);
665
666                 size_t local_size[2] = {workgroup_size, workgroup_size};
667                 size_t global_size[2] = {global_size_round_up(local_size[0], d_w), global_size_round_up(local_size[1], d_h)};
668
669                 /* run kernel */
670                 ciErr = clEnqueueNDRangeKernel(cqCommandQueue, ckFilmConvertKernel, 2, NULL, global_size, local_size, 0, NULL, NULL);
671                 opencl_assert(ciErr);
672                 opencl_assert(clFinish(cqCommandQueue));
673         }
674
675         void thread_run(DeviceTask *task)
676         {
677                 if(task->type == DeviceTask::TONEMAP) {
678                         tonemap(*task, task->buffer, task->rgba);
679                 }
680                 else if(task->type == DeviceTask::PATH_TRACE) {
681                         RenderTile tile;
682                         
683                         /* keep rendering tiles until done */
684                         while(task->acquire_tile(this, tile)) {
685                                 int start_sample = tile.start_sample;
686                                 int end_sample = tile.start_sample + tile.num_samples;
687
688                                 for(int sample = start_sample; sample < end_sample; sample++) {
689                                         if (task->get_cancel())
690                                                 break;
691
692                                         path_trace(tile, sample);
693
694                                         tile.sample = sample + 1;
695
696                                         task->update_progress(tile);
697                                 }
698
699                                 task->release_tile(tile);
700                         }
701                 }
702         }
703
704         class OpenCLDeviceTask : public DeviceTask {
705         public:
706                 OpenCLDeviceTask(OpenCLDevice *device, DeviceTask& task)
707                 : DeviceTask(task)
708                 {
709                         run = function_bind(&OpenCLDevice::thread_run, device, this);
710                 }
711         };
712
713         void task_add(DeviceTask& task)
714         {
715                 task_pool.push(new OpenCLDeviceTask(this, task));
716         }
717
718         void task_wait()
719         {
720                 task_pool.wait_work();
721         }
722
723         void task_cancel()
724         {
725                 task_pool.cancel();
726         }
727
728         bool task_cancelled()
729         {
730                 return task_pool.cancelled();
731         }
732 };
733
734 Device *device_opencl_create(DeviceInfo& info, bool background)
735 {
736         return new OpenCLDevice(info, background);
737 }
738
739 void device_opencl_info(vector<DeviceInfo>& devices)
740 {
741         vector<cl_device_id> device_ids;
742         cl_uint num_devices = 0;
743         vector<cl_platform_id> platform_ids;
744         cl_uint num_platforms = 0;
745
746         /* get devices */
747         if(clGetPlatformIDs(0, NULL, &num_platforms) != CL_SUCCESS || num_platforms == 0)
748                 return;
749         
750         platform_ids.resize(num_platforms);
751
752         if(clGetPlatformIDs(num_platforms, &platform_ids[0], NULL) != CL_SUCCESS)
753                 return;
754
755         if(clGetDeviceIDs(platform_ids[0], CL_DEVICE_TYPE_GPU|CL_DEVICE_TYPE_ACCELERATOR, 0, NULL, &num_devices) != CL_SUCCESS || num_devices == 0)
756                 return;
757         
758         device_ids.resize(num_devices);
759
760         if(clGetDeviceIDs(platform_ids[0], CL_DEVICE_TYPE_GPU|CL_DEVICE_TYPE_ACCELERATOR, num_devices, &device_ids[0], NULL) != CL_SUCCESS)
761                 return;
762         
763         /* add devices */
764         for(int num = 0; num < num_devices; num++) {
765                 cl_device_id device_id = device_ids[num];
766                 char name[1024] = "\0";
767
768                 if(clGetDeviceInfo(device_id, CL_DEVICE_NAME, sizeof(name), &name, NULL) != CL_SUCCESS)
769                         continue;
770
771                 DeviceInfo info;
772
773                 info.type = DEVICE_OPENCL;
774                 info.description = string(name);
775                 info.id = string_printf("OPENCL_%d", num);
776                 info.num = num;
777                 /* we don't know if it's used for display, but assume it is */
778                 info.display_device = true;
779                 info.advanced_shading = false;
780                 info.pack_images = true;
781
782                 devices.push_back(info);
783         }
784 }
785
786 CCL_NAMESPACE_END
787
788 #endif /* WITH_OPENCL */
789