Merge branch 'master' into blender2.8
[blender.git] / intern / cycles / device / device_cuda.cpp
1 /*
2  * Copyright 2011-2013 Blender Foundation
3  *
4  * Licensed under the Apache License, Version 2.0 (the "License");
5  * you may not use this file except in compliance with the License.
6  * You may obtain a copy of the License at
7  *
8  * http://www.apache.org/licenses/LICENSE-2.0
9  *
10  * Unless required by applicable law or agreed to in writing, software
11  * distributed under the License is distributed on an "AS IS" BASIS,
12  * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13  * See the License for the specific language governing permissions and
14  * limitations under the License.
15  */
16
17 #include <climits>
18 #include <limits.h>
19 #include <stdio.h>
20 #include <stdlib.h>
21 #include <string.h>
22
23 #include "device/device.h"
24 #include "device/device_denoising.h"
25 #include "device/device_intern.h"
26 #include "device/device_split_kernel.h"
27
28 #include "render/buffers.h"
29
30 #include "kernel/filter/filter_defines.h"
31
32 #ifdef WITH_CUDA_DYNLOAD
33 #  include "cuew.h"
34 #else
35 #  include "util/util_opengl.h"
36 #  include <cuda.h>
37 #  include <cudaGL.h>
38 #endif
39 #include "util/util_debug.h"
40 #include "util/util_foreach.h"
41 #include "util/util_logging.h"
42 #include "util/util_map.h"
43 #include "util/util_md5.h"
44 #include "util/util_opengl.h"
45 #include "util/util_path.h"
46 #include "util/util_string.h"
47 #include "util/util_system.h"
48 #include "util/util_types.h"
49 #include "util/util_time.h"
50
51 #include "kernel/split/kernel_split_data_types.h"
52
53 CCL_NAMESPACE_BEGIN
54
55 #ifndef WITH_CUDA_DYNLOAD
56
57 /* Transparently implement some functions, so majority of the file does not need
58  * to worry about difference between dynamically loaded and linked CUDA at all.
59  */
60
61 namespace {
62
63 const char *cuewErrorString(CUresult result)
64 {
65         /* We can only give error code here without major code duplication, that
66          * should be enough since dynamic loading is only being disabled by folks
67          * who knows what they're doing anyway.
68          *
69          * NOTE: Avoid call from several threads.
70          */
71         static string error;
72         error = string_printf("%d", result);
73         return error.c_str();
74 }
75
76 const char *cuewCompilerPath(void)
77 {
78         return CYCLES_CUDA_NVCC_EXECUTABLE;
79 }
80
81 int cuewCompilerVersion(void)
82 {
83         return (CUDA_VERSION / 100) + (CUDA_VERSION % 100 / 10);
84 }
85
86 }  /* namespace */
87 #endif  /* WITH_CUDA_DYNLOAD */
88
89 class CUDADevice;
90
91 class CUDASplitKernel : public DeviceSplitKernel {
92         CUDADevice *device;
93 public:
94         explicit CUDASplitKernel(CUDADevice *device);
95
96         virtual uint64_t state_buffer_size(device_memory& kg, device_memory& data, size_t num_threads);
97
98         virtual bool enqueue_split_kernel_data_init(const KernelDimensions& dim,
99                                                     RenderTile& rtile,
100                                                     int num_global_elements,
101                                                     device_memory& kernel_globals,
102                                                     device_memory& kernel_data_,
103                                                     device_memory& split_data,
104                                                     device_memory& ray_state,
105                                                     device_memory& queue_index,
106                                                     device_memory& use_queues_flag,
107                                                     device_memory& work_pool_wgs);
108
109         virtual SplitKernelFunction* get_split_kernel_function(const string& kernel_name,
110                                                                const DeviceRequestedFeatures&);
111         virtual int2 split_kernel_local_size();
112         virtual int2 split_kernel_global_size(device_memory& kg, device_memory& data, DeviceTask *task);
113 };
114
115 /* Utility to push/pop CUDA context. */
116 class CUDAContextScope {
117 public:
118         CUDAContextScope(CUDADevice *device);
119         ~CUDAContextScope();
120
121 private:
122         CUDADevice *device;
123 };
124
125 class CUDADevice : public Device
126 {
127 public:
128         DedicatedTaskPool task_pool;
129         CUdevice cuDevice;
130         CUcontext cuContext;
131         CUmodule cuModule, cuFilterModule;
132         size_t device_texture_headroom;
133         size_t device_working_headroom;
134         bool move_texture_to_host;
135         size_t map_host_used;
136         size_t map_host_limit;
137         int can_map_host;
138         int cuDevId;
139         int cuDevArchitecture;
140         bool first_error;
141         CUDASplitKernel *split_kernel;
142
143         struct CUDAMem {
144                 CUDAMem()
145                 : texobject(0), array(0), map_host_pointer(0), free_map_host(false) {}
146
147                 CUtexObject texobject;
148                 CUarray array;
149                 void *map_host_pointer;
150                 bool free_map_host;
151         };
152         typedef map<device_memory*, CUDAMem> CUDAMemMap;
153         CUDAMemMap cuda_mem_map;
154
155         struct PixelMem {
156                 GLuint cuPBO;
157                 CUgraphicsResource cuPBOresource;
158                 GLuint cuTexId;
159                 int w, h;
160         };
161         map<device_ptr, PixelMem> pixel_mem_map;
162
163         /* Bindless Textures */
164         device_vector<TextureInfo> texture_info;
165         bool need_texture_info;
166
167         CUdeviceptr cuda_device_ptr(device_ptr mem)
168         {
169                 return (CUdeviceptr)mem;
170         }
171
172         static bool have_precompiled_kernels()
173         {
174                 string cubins_path = path_get("lib");
175                 return path_exists(cubins_path);
176         }
177
178         virtual bool show_samples() const
179         {
180                 /* The CUDADevice only processes one tile at a time, so showing samples is fine. */
181                 return true;
182         }
183
184 /*#ifdef NDEBUG
185 #define cuda_abort()
186 #else
187 #define cuda_abort() abort()
188 #endif*/
189         void cuda_error_documentation()
190         {
191                 if(first_error) {
192                         fprintf(stderr, "\nRefer to the Cycles GPU rendering documentation for possible solutions:\n");
193                         fprintf(stderr, "https://docs.blender.org/manual/en/dev/render/cycles/gpu_rendering.html\n\n");
194                         first_error = false;
195                 }
196         }
197
198 #define cuda_assert(stmt) \
199         { \
200                 CUresult result = stmt; \
201                 \
202                 if(result != CUDA_SUCCESS) { \
203                         string message = string_printf("CUDA error: %s in %s, line %d", cuewErrorString(result), #stmt, __LINE__); \
204                         if(error_msg == "") \
205                                 error_msg = message; \
206                         fprintf(stderr, "%s\n", message.c_str()); \
207                         /*cuda_abort();*/ \
208                         cuda_error_documentation(); \
209                 } \
210         } (void)0
211
212         bool cuda_error_(CUresult result, const string& stmt)
213         {
214                 if(result == CUDA_SUCCESS)
215                         return false;
216
217                 string message = string_printf("CUDA error at %s: %s", stmt.c_str(), cuewErrorString(result));
218                 if(error_msg == "")
219                         error_msg = message;
220                 fprintf(stderr, "%s\n", message.c_str());
221                 cuda_error_documentation();
222                 return true;
223         }
224
225 #define cuda_error(stmt) cuda_error_(stmt, #stmt)
226
227         void cuda_error_message(const string& message)
228         {
229                 if(error_msg == "")
230                         error_msg = message;
231                 fprintf(stderr, "%s\n", message.c_str());
232                 cuda_error_documentation();
233         }
234
235         CUDADevice(DeviceInfo& info, Stats &stats, bool background_)
236         : Device(info, stats, background_),
237           texture_info(this, "__texture_info", MEM_TEXTURE)
238         {
239                 first_error = true;
240                 background = background_;
241
242                 cuDevId = info.num;
243                 cuDevice = 0;
244                 cuContext = 0;
245
246                 cuModule = 0;
247                 cuFilterModule = 0;
248
249                 split_kernel = NULL;
250
251                 need_texture_info = false;
252
253                 device_texture_headroom = 0;
254                 device_working_headroom = 0;
255                 move_texture_to_host = false;
256                 map_host_limit = 0;
257                 map_host_used = 0;
258                 can_map_host = 0;
259
260                 /* Intialize CUDA. */
261                 if(cuda_error(cuInit(0)))
262                         return;
263
264                 /* Setup device and context. */
265                 if(cuda_error(cuDeviceGet(&cuDevice, cuDevId)))
266                         return;
267
268                 /* CU_CTX_MAP_HOST for mapping host memory when out of device memory.
269                  * CU_CTX_LMEM_RESIZE_TO_MAX for reserving local memory ahead of render,
270                  * so we can predict which memory to map to host. */
271                 cuda_assert(cuDeviceGetAttribute(&can_map_host, CU_DEVICE_ATTRIBUTE_CAN_MAP_HOST_MEMORY, cuDevice));
272
273                 unsigned int ctx_flags = CU_CTX_LMEM_RESIZE_TO_MAX;
274                 if(can_map_host) {
275                         ctx_flags |= CU_CTX_MAP_HOST;
276                         init_host_memory();
277                 }
278
279                 /* Create context. */
280                 CUresult result;
281
282                 if(background) {
283                         result = cuCtxCreate(&cuContext, ctx_flags, cuDevice);
284                 }
285                 else {
286                         result = cuGLCtxCreate(&cuContext, ctx_flags, cuDevice);
287
288                         if(result != CUDA_SUCCESS) {
289                                 result = cuCtxCreate(&cuContext, ctx_flags, cuDevice);
290                                 background = true;
291                         }
292                 }
293
294                 if(cuda_error_(result, "cuCtxCreate"))
295                         return;
296
297                 int major, minor;
298                 cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevId);
299                 cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevId);
300                 cuDevArchitecture = major*100 + minor*10;
301
302                 /* Pop context set by cuCtxCreate. */
303                 cuCtxPopCurrent(NULL);
304         }
305
306         ~CUDADevice()
307         {
308                 task_pool.stop();
309
310                 delete split_kernel;
311
312                 texture_info.free();
313
314                 cuda_assert(cuCtxDestroy(cuContext));
315         }
316
317         bool support_device(const DeviceRequestedFeatures& /*requested_features*/)
318         {
319                 int major, minor;
320                 cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevId);
321                 cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevId);
322
323                 /* We only support sm_30 and above */
324                 if(major < 3) {
325                         cuda_error_message(string_printf("CUDA device supported only with compute capability 3.0 or up, found %d.%d.", major, minor));
326                         return false;
327                 }
328
329                 return true;
330         }
331
332         bool use_adaptive_compilation()
333         {
334                 return DebugFlags().cuda.adaptive_compile;
335         }
336
337         bool use_split_kernel()
338         {
339                 return DebugFlags().cuda.split_kernel;
340         }
341
342         /* Common NVCC flags which stays the same regardless of shading model,
343          * kernel sources md5 and only depends on compiler or compilation settings.
344          */
345         string compile_kernel_get_common_cflags(
346                 const DeviceRequestedFeatures& requested_features,
347                 bool filter=false, bool split=false)
348         {
349                 const int machine = system_cpu_bits();
350                 const string source_path = path_get("source");
351                 const string include_path = source_path;
352                 string cflags = string_printf("-m%d "
353                                               "--ptxas-options=\"-v\" "
354                                               "--use_fast_math "
355                                               "-DNVCC "
356                                                "-I\"%s\"",
357                                               machine,
358                                               include_path.c_str());
359                 if(!filter && use_adaptive_compilation()) {
360                         cflags += " " + requested_features.get_build_options();
361                 }
362                 const char *extra_cflags = getenv("CYCLES_CUDA_EXTRA_CFLAGS");
363                 if(extra_cflags) {
364                         cflags += string(" ") + string(extra_cflags);
365                 }
366 #ifdef WITH_CYCLES_DEBUG
367                 cflags += " -D__KERNEL_DEBUG__";
368 #endif
369
370                 if(split) {
371                         cflags += " -D__SPLIT__";
372                 }
373
374                 return cflags;
375         }
376
377         bool compile_check_compiler() {
378                 const char *nvcc = cuewCompilerPath();
379                 if(nvcc == NULL) {
380                         cuda_error_message("CUDA nvcc compiler not found. "
381                                            "Install CUDA toolkit in default location.");
382                         return false;
383                 }
384                 const int cuda_version = cuewCompilerVersion();
385                 VLOG(1) << "Found nvcc " << nvcc
386                         << ", CUDA version " << cuda_version
387                         << ".";
388                 const int major = cuda_version / 10, minor = cuda_version % 10;
389                 if(cuda_version == 0) {
390                         cuda_error_message("CUDA nvcc compiler version could not be parsed.");
391                         return false;
392                 }
393                 if(cuda_version < 80) {
394                         printf("Unsupported CUDA version %d.%d detected, "
395                                "you need CUDA 8.0 or newer.\n",
396                                major, minor);
397                         return false;
398                 }
399                 else if(cuda_version != 80) {
400                         printf("CUDA version %d.%d detected, build may succeed but only "
401                                "CUDA 8.0 is officially supported.\n",
402                                major, minor);
403                 }
404                 return true;
405         }
406
407         string compile_kernel(const DeviceRequestedFeatures& requested_features,
408                               bool filter=false, bool split=false)
409         {
410                 const char *name, *source;
411                 if(filter) {
412                         name = "filter";
413                         source = "filter.cu";
414                 }
415                 else if(split) {
416                         name = "kernel_split";
417                         source = "kernel_split.cu";
418                 }
419                 else {
420                         name = "kernel";
421                         source = "kernel.cu";
422                 }
423                 /* Compute cubin name. */
424                 int major, minor;
425                 cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, cuDevId);
426                 cuDeviceGetAttribute(&minor, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MINOR, cuDevId);
427
428                 /* Attempt to use kernel provided with Blender. */
429                 if(!use_adaptive_compilation()) {
430                         const string cubin = path_get(string_printf("lib/%s_sm_%d%d.cubin",
431                                                                     name, major, minor));
432                         VLOG(1) << "Testing for pre-compiled kernel " << cubin << ".";
433                         if(path_exists(cubin)) {
434                                 VLOG(1) << "Using precompiled kernel.";
435                                 return cubin;
436                         }
437                 }
438
439                 const string common_cflags =
440                         compile_kernel_get_common_cflags(requested_features, filter, split);
441
442                 /* Try to use locally compiled kernel. */
443                 const string source_path = path_get("source");
444                 const string kernel_md5 = path_files_md5_hash(source_path);
445
446                 /* We include cflags into md5 so changing cuda toolkit or changing other
447                  * compiler command line arguments makes sure cubin gets re-built.
448                  */
449                 const string cubin_md5 = util_md5_string(kernel_md5 + common_cflags);
450
451                 const string cubin_file = string_printf("cycles_%s_sm%d%d_%s.cubin",
452                                                         name, major, minor,
453                                                         cubin_md5.c_str());
454                 const string cubin = path_cache_get(path_join("kernels", cubin_file));
455                 VLOG(1) << "Testing for locally compiled kernel " << cubin << ".";
456                 if(path_exists(cubin)) {
457                         VLOG(1) << "Using locally compiled kernel.";
458                         return cubin;
459                 }
460
461 #ifdef _WIN32
462                 if(have_precompiled_kernels()) {
463                         if(major < 3) {
464                                 cuda_error_message(string_printf(
465                                         "CUDA device requires compute capability 3.0 or up, "
466                                         "found %d.%d. Your GPU is not supported.",
467                                         major, minor));
468                         }
469                         else {
470                                 cuda_error_message(string_printf(
471                                         "CUDA binary kernel for this graphics card compute "
472                                         "capability (%d.%d) not found.",
473                                         major, minor));
474                         }
475                         return "";
476                 }
477 #endif
478
479                 /* Compile. */
480                 if(!compile_check_compiler()) {
481                         return "";
482                 }
483                 const char *nvcc = cuewCompilerPath();
484                 const string kernel = path_join(
485                         path_join(source_path, "kernel"),
486                         path_join("kernels",
487                                   path_join("cuda", source)));
488                 double starttime = time_dt();
489                 printf("Compiling CUDA kernel ...\n");
490
491                 path_create_directories(cubin);
492
493                 string command = string_printf("\"%s\" "
494                                                "-arch=sm_%d%d "
495                                                "--cubin \"%s\" "
496                                                "-o \"%s\" "
497                                                "%s ",
498                                                nvcc,
499                                                major, minor,
500                                                kernel.c_str(),
501                                                cubin.c_str(),
502                                                common_cflags.c_str());
503
504                 printf("%s\n", command.c_str());
505
506                 if(system(command.c_str()) == -1) {
507                         cuda_error_message("Failed to execute compilation command, "
508                                            "see console for details.");
509                         return "";
510                 }
511
512                 /* Verify if compilation succeeded */
513                 if(!path_exists(cubin)) {
514                         cuda_error_message("CUDA kernel compilation failed, "
515                                            "see console for details.");
516                         return "";
517                 }
518
519                 printf("Kernel compilation finished in %.2lfs.\n", time_dt() - starttime);
520
521                 return cubin;
522         }
523
524         bool load_kernels(const DeviceRequestedFeatures& requested_features)
525         {
526                 /* TODO(sergey): Support kernels re-load for CUDA devices.
527                  *
528                  * Currently re-loading kernel will invalidate memory pointers,
529                  * causing problems in cuCtxSynchronize.
530                  */
531                 if(cuFilterModule && cuModule) {
532                         VLOG(1) << "Skipping kernel reload, not currently supported.";
533                         return true;
534                 }
535
536                 /* check if cuda init succeeded */
537                 if(cuContext == 0)
538                         return false;
539
540                 /* check if GPU is supported */
541                 if(!support_device(requested_features))
542                         return false;
543
544                 /* get kernel */
545                 string cubin = compile_kernel(requested_features, false, use_split_kernel());
546                 if(cubin == "")
547                         return false;
548
549                 string filter_cubin = compile_kernel(requested_features, true, false);
550                 if(filter_cubin == "")
551                         return false;
552
553                 /* open module */
554                 CUDAContextScope scope(this);
555
556                 string cubin_data;
557                 CUresult result;
558
559                 if(path_read_text(cubin, cubin_data))
560                         result = cuModuleLoadData(&cuModule, cubin_data.c_str());
561                 else
562                         result = CUDA_ERROR_FILE_NOT_FOUND;
563
564                 if(cuda_error_(result, "cuModuleLoad"))
565                         cuda_error_message(string_printf("Failed loading CUDA kernel %s.", cubin.c_str()));
566
567                 if(path_read_text(filter_cubin, cubin_data))
568                         result = cuModuleLoadData(&cuFilterModule, cubin_data.c_str());
569                 else
570                         result = CUDA_ERROR_FILE_NOT_FOUND;
571
572                 if(cuda_error_(result, "cuModuleLoad"))
573                         cuda_error_message(string_printf("Failed loading CUDA kernel %s.", filter_cubin.c_str()));
574
575                 if(result == CUDA_SUCCESS) {
576                         reserve_local_memory(requested_features);
577                 }
578
579                 return (result == CUDA_SUCCESS);
580         }
581
582         void reserve_local_memory(const DeviceRequestedFeatures& requested_features)
583         {
584                 if(use_split_kernel()) {
585                         /* Split kernel mostly uses global memory and adaptive compilation,
586                          * difficult to predict how much is needed currently. */
587                         return;
588                 }
589
590                 /* Together with CU_CTX_LMEM_RESIZE_TO_MAX, this reserves local memory
591                  * needed for kernel launches, so that we can reliably figure out when
592                  * to allocate scene data in mapped host memory. */
593                 CUDAContextScope scope(this);
594
595                 size_t total = 0, free_before = 0, free_after = 0;
596                 cuMemGetInfo(&free_before, &total);
597
598                 /* Get kernel function. */
599                 CUfunction cuPathTrace;
600
601                 if(requested_features.use_integrator_branched) {
602                         cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_branched_path_trace"));
603                 }
604                 else {
605                         cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_path_trace"));
606                 }
607
608                 cuda_assert(cuFuncSetCacheConfig(cuPathTrace, CU_FUNC_CACHE_PREFER_L1));
609
610                 int min_blocks, num_threads_per_block;
611                 cuda_assert(cuOccupancyMaxPotentialBlockSize(&min_blocks, &num_threads_per_block, cuPathTrace, NULL, 0, 0));
612
613                 /* Launch kernel, using just 1 block appears sufficient to reserve
614                  * memory for all multiprocessors. It would be good to do this in
615                  * parallel for the multi GPU case still to make it faster. */
616                 CUdeviceptr d_work_tiles = 0;
617                 uint total_work_size = 0;
618
619                 void *args[] = {&d_work_tiles,
620                                 &total_work_size};
621
622                 cuda_assert(cuLaunchKernel(cuPathTrace,
623                                            1, 1, 1,
624                                            num_threads_per_block, 1, 1,
625                                            0, 0, args, 0));
626
627                 cuda_assert(cuCtxSynchronize());
628
629                 cuMemGetInfo(&free_after, &total);
630                 VLOG(1) << "Local memory reserved "
631                         << string_human_readable_number(free_before - free_after) << " bytes. ("
632                         << string_human_readable_size(free_before - free_after) << ")";
633
634 #if 0
635                 /* For testing mapped host memory, fill up device memory. */
636                 const size_t keep_mb = 1024;
637
638                 while(free_after > keep_mb * 1024 * 1024LL) {
639                         CUdeviceptr tmp;
640                         cuda_assert(cuMemAlloc(&tmp, 10 * 1024 * 1024LL));
641                         cuMemGetInfo(&free_after, &total);
642                 }
643 #endif
644         }
645
646         void init_host_memory()
647         {
648                 /* Limit amount of host mapped memory, because allocating too much can
649                  * cause system instability. Leave at least half or 4 GB of system
650                  * memory free, whichever is smaller. */
651                 size_t default_limit = 4 * 1024 * 1024 * 1024LL;
652                 size_t system_ram = system_physical_ram();
653
654                 if(system_ram > 0) {
655                         if(system_ram / 2 > default_limit) {
656                                 map_host_limit = system_ram - default_limit;
657                         }
658                         else {
659                                 map_host_limit = system_ram / 2;
660                         }
661                 }
662                 else {
663                         VLOG(1) << "Mapped host memory disabled, failed to get system RAM";
664                         map_host_limit = 0;
665                 }
666
667                 /* Amount of device memory to keep is free after texture memory
668                  * and working memory allocations respectively. We set the working
669                  * memory limit headroom lower so that some space is left after all
670                  * texture memory allocations. */
671                 device_working_headroom = 32 * 1024 * 1024LL; // 32MB
672                 device_texture_headroom = 128 * 1024 * 1024LL; // 128MB
673
674                 VLOG(1) << "Mapped host memory limit set to "
675                         << string_human_readable_number(map_host_limit) << " bytes. ("
676                         << string_human_readable_size(map_host_limit) << ")";
677         }
678
679         void load_texture_info()
680         {
681                 if(need_texture_info) {
682                         texture_info.copy_to_device();
683                         need_texture_info = false;
684                 }
685         }
686
687         void move_textures_to_host(size_t size, bool for_texture)
688         {
689                 /* Signal to reallocate textures in host memory only. */
690                 move_texture_to_host = true;
691
692                 while(size > 0) {
693                         /* Find suitable memory allocation to move. */
694                         device_memory *max_mem = NULL;
695                         size_t max_size = 0;
696                         bool max_is_image = false;
697
698                         foreach(CUDAMemMap::value_type& pair, cuda_mem_map) {
699                                 device_memory& mem = *pair.first;
700                                 CUDAMem *cmem = &pair.second;
701
702                                 bool is_texture = (mem.type == MEM_TEXTURE) && (&mem != &texture_info);
703                                 bool is_image = is_texture && (mem.data_height > 1);
704
705                                 /* Can't move this type of memory. */
706                                 if(!is_texture || cmem->array) {
707                                         continue;
708                                 }
709
710                                 /* Already in host memory. */
711                                 if(cmem->map_host_pointer) {
712                                         continue;
713                                 }
714
715                                 /* For other textures, only move image textures. */
716                                 if(for_texture && !is_image) {
717                                         continue;
718                                 }
719
720                                 /* Try to move largest allocation, prefer moving images. */
721                                 if(is_image > max_is_image ||
722                                    (is_image == max_is_image && mem.device_size > max_size)) {
723                                         max_is_image = is_image;
724                                         max_size = mem.device_size;
725                                         max_mem = &mem;
726                                 }
727                         }
728
729                         /* Move to host memory. This part is mutex protected since
730                          * multiple CUDA devices could be moving the memory. The
731                          * first one will do it, and the rest will adopt the pointer. */
732                         if(max_mem) {
733                                 VLOG(1) << "Move memory from device to host: " << max_mem->name;
734
735                                 static thread_mutex move_mutex;
736                                 thread_scoped_lock lock(move_mutex);
737
738                                 /* Preserve the original device pointer, in case of multi device
739                                  * we can't change it because the pointer mapping would break. */
740                                 device_ptr prev_pointer = max_mem->device_pointer;
741                                 size_t prev_size = max_mem->device_size;
742
743                                 tex_free(*max_mem);
744                                 tex_alloc(*max_mem);
745                                 size = (max_size >= size)? 0: size - max_size;
746
747                                 max_mem->device_pointer = prev_pointer;
748                                 max_mem->device_size = prev_size;
749                         }
750                         else {
751                                 break;
752                         }
753                 }
754
755                 /* Update texture info array with new pointers. */
756                 load_texture_info();
757
758                 move_texture_to_host = false;
759         }
760
761         CUDAMem *generic_alloc(device_memory& mem, size_t pitch_padding = 0)
762         {
763                 CUDAContextScope scope(this);
764
765                 CUdeviceptr device_pointer = 0;
766                 size_t size = mem.memory_size() + pitch_padding;
767
768                 CUresult mem_alloc_result = CUDA_ERROR_OUT_OF_MEMORY;
769                 const char *status = "";
770
771                 /* First try allocating in device memory, respecting headroom. We make
772                  * an exception for texture info. It is small and frequently accessed,
773                  * so treat it as working memory.
774                  *
775                  * If there is not enough room for working memory, we will try to move
776                  * textures to host memory, assuming the performance impact would have
777                  * been worse for working memory. */
778                 bool is_texture = (mem.type == MEM_TEXTURE) && (&mem != &texture_info);
779                 bool is_image = is_texture && (mem.data_height > 1);
780
781                 size_t headroom = (is_texture)? device_texture_headroom:
782                                                 device_working_headroom;
783
784                 size_t total = 0, free = 0;
785                 cuMemGetInfo(&free, &total);
786
787                 /* Move textures to host memory if needed. */
788                 if(!move_texture_to_host && !is_image && (size + headroom) >= free) {
789                         move_textures_to_host(size + headroom - free, is_texture);
790                         cuMemGetInfo(&free, &total);
791                 }
792
793                 /* Allocate in device memory. */
794                 if(!move_texture_to_host && (size + headroom) < free) {
795                         mem_alloc_result = cuMemAlloc(&device_pointer, size);
796                         if(mem_alloc_result == CUDA_SUCCESS) {
797                                 status = " in device memory";
798                         }
799                 }
800
801                 /* Fall back to mapped host memory if needed and possible. */
802                 void *map_host_pointer = 0;
803                 bool free_map_host = false;
804
805                 if(mem_alloc_result != CUDA_SUCCESS && can_map_host &&
806                    map_host_used + size < map_host_limit) {
807                         if(mem.shared_pointer) {
808                                 /* Another device already allocated host memory. */
809                                 mem_alloc_result = CUDA_SUCCESS;
810                                 map_host_pointer = mem.shared_pointer;
811                         }
812                         else {
813                                 /* Allocate host memory ourselves. */
814                                 mem_alloc_result = cuMemHostAlloc(&map_host_pointer, size,
815                                                                   CU_MEMHOSTALLOC_DEVICEMAP |
816                                                                   CU_MEMHOSTALLOC_WRITECOMBINED);
817                                 mem.shared_pointer = map_host_pointer;
818                                 free_map_host = true;
819                         }
820
821                         if(mem_alloc_result == CUDA_SUCCESS) {
822                                 cuda_assert(cuMemHostGetDevicePointer_v2(&device_pointer, mem.shared_pointer, 0));
823                                 map_host_used += size;
824                                 status = " in host memory";
825
826                                 /* Replace host pointer with our host allocation. Only works if
827                                  * CUDA memory layout is the same and has no pitch padding. Also
828                                  * does not work if we move textures to host during a render,
829                                  * since other devices might be using the memory. */
830                                 if(!move_texture_to_host && pitch_padding == 0 &&
831                                    mem.host_pointer && mem.host_pointer != mem.shared_pointer) {
832                                         memcpy(mem.shared_pointer, mem.host_pointer, size);
833                                         mem.host_free();
834                                         mem.host_pointer = mem.shared_pointer;
835                                 }
836                         }
837                         else {
838                                 status = " failed, out of host memory";
839                         }
840                 }
841                 else if(mem_alloc_result != CUDA_SUCCESS) {
842                         status = " failed, out of device and host memory";
843                 }
844
845                 if(mem_alloc_result != CUDA_SUCCESS) {
846                         cuda_assert(mem_alloc_result);
847                 }
848
849                 if(mem.name) {
850                         VLOG(1) << "Buffer allocate: " << mem.name << ", "
851                                         << string_human_readable_number(mem.memory_size()) << " bytes. ("
852                                         << string_human_readable_size(mem.memory_size()) << ")"
853                                         << status;
854                 }
855
856                 mem.device_pointer = (device_ptr)device_pointer;
857                 mem.device_size = size;
858                 stats.mem_alloc(size);
859
860                 if(!mem.device_pointer) {
861                         return NULL;
862                 }
863
864                 /* Insert into map of allocations. */
865                 CUDAMem *cmem = &cuda_mem_map[&mem];
866                 cmem->map_host_pointer = map_host_pointer;
867                 cmem->free_map_host = free_map_host;
868                 return cmem;
869         }
870
871         void generic_copy_to(device_memory& mem)
872         {
873                 if(mem.host_pointer && mem.device_pointer) {
874                         CUDAContextScope scope(this);
875
876                         if(mem.host_pointer != mem.shared_pointer) {
877                                 cuda_assert(cuMemcpyHtoD(cuda_device_ptr(mem.device_pointer),
878                                                          mem.host_pointer,
879                                                          mem.memory_size()));
880                         }
881                 }
882         }
883
884         void generic_free(device_memory& mem)
885         {
886                 if(mem.device_pointer) {
887                         CUDAContextScope scope(this);
888                         const CUDAMem& cmem = cuda_mem_map[&mem];
889
890                         if(cmem.map_host_pointer) {
891                                 /* Free host memory. */
892                                 if(cmem.free_map_host) {
893                                         cuMemFreeHost(cmem.map_host_pointer);
894                                         if(mem.host_pointer == mem.shared_pointer) {
895                                                 mem.host_pointer = 0;
896                                         }
897                                         mem.shared_pointer = 0;
898                                 }
899
900                                 map_host_used -= mem.device_size;
901                         }
902                         else {
903                                 /* Free device memory. */
904                                 cuMemFree(mem.device_pointer);
905                         }
906
907                         stats.mem_free(mem.device_size);
908                         mem.device_pointer = 0;
909                         mem.device_size = 0;
910
911                         cuda_mem_map.erase(cuda_mem_map.find(&mem));
912                 }
913         }
914
915         void mem_alloc(device_memory& mem)
916         {
917                 if(mem.type == MEM_PIXELS && !background) {
918                         pixels_alloc(mem);
919                 }
920                 else if(mem.type == MEM_TEXTURE) {
921                         assert(!"mem_alloc not supported for textures.");
922                 }
923                 else {
924                         generic_alloc(mem);
925                 }
926         }
927
928         void mem_copy_to(device_memory& mem)
929         {
930                 if(mem.type == MEM_PIXELS) {
931                         assert(!"mem_copy_to not supported for pixels.");
932                 }
933                 else if(mem.type == MEM_TEXTURE) {
934                         tex_free(mem);
935                         tex_alloc(mem);
936                 }
937                 else {
938                         if(!mem.device_pointer) {
939                                 generic_alloc(mem);
940                         }
941
942                         generic_copy_to(mem);
943                 }
944         }
945
946         void mem_copy_from(device_memory& mem, int y, int w, int h, int elem)
947         {
948                 if(mem.type == MEM_PIXELS && !background) {
949                         pixels_copy_from(mem, y, w, h);
950                 }
951                 else if(mem.type == MEM_TEXTURE) {
952                         assert(!"mem_copy_from not supported for textures.");
953                 }
954                 else {
955                         CUDAContextScope scope(this);
956                         size_t offset = elem*y*w;
957                         size_t size = elem*w*h;
958
959                         if(mem.host_pointer && mem.device_pointer) {
960                                 cuda_assert(cuMemcpyDtoH((uchar*)mem.host_pointer + offset,
961                                                                                  (CUdeviceptr)(mem.device_pointer + offset), size));
962                         }
963                         else if(mem.host_pointer) {
964                                 memset((char*)mem.host_pointer + offset, 0, size);
965                         }
966                 }
967         }
968
969         void mem_zero(device_memory& mem)
970         {
971                 if(!mem.device_pointer) {
972                         mem_alloc(mem);
973                 }
974
975                 if(mem.host_pointer) {
976                         memset(mem.host_pointer, 0, mem.memory_size());
977                 }
978
979                 if(mem.device_pointer &&
980                    (!mem.host_pointer || mem.host_pointer != mem.shared_pointer)) {
981                         CUDAContextScope scope(this);
982                         cuda_assert(cuMemsetD8(cuda_device_ptr(mem.device_pointer), 0, mem.memory_size()));
983                 }
984         }
985
986         void mem_free(device_memory& mem)
987         {
988                 if(mem.type == MEM_PIXELS && !background) {
989                         pixels_free(mem);
990                 }
991                 else if(mem.type == MEM_TEXTURE) {
992                         tex_free(mem);
993                 }
994                 else {
995                         generic_free(mem);
996                 }
997         }
998
999         virtual device_ptr mem_alloc_sub_ptr(device_memory& mem, int offset, int /*size*/)
1000         {
1001                 return (device_ptr) (((char*) mem.device_pointer) + mem.memory_elements_size(offset));
1002         }
1003
1004         void const_copy_to(const char *name, void *host, size_t size)
1005         {
1006                 CUDAContextScope scope(this);
1007                 CUdeviceptr mem;
1008                 size_t bytes;
1009
1010                 cuda_assert(cuModuleGetGlobal(&mem, &bytes, cuModule, name));
1011                 //assert(bytes == size);
1012                 cuda_assert(cuMemcpyHtoD(mem, host, size));
1013         }
1014
1015         void tex_alloc(device_memory& mem)
1016         {
1017                 CUDAContextScope scope(this);
1018
1019                 /* General variables for both architectures */
1020                 string bind_name = mem.name;
1021                 size_t dsize = datatype_size(mem.data_type);
1022                 size_t size = mem.memory_size();
1023
1024                 CUaddress_mode address_mode = CU_TR_ADDRESS_MODE_WRAP;
1025                 switch(mem.extension) {
1026                         case EXTENSION_REPEAT:
1027                                 address_mode = CU_TR_ADDRESS_MODE_WRAP;
1028                                 break;
1029                         case EXTENSION_EXTEND:
1030                                 address_mode = CU_TR_ADDRESS_MODE_CLAMP;
1031                                 break;
1032                         case EXTENSION_CLIP:
1033                                 address_mode = CU_TR_ADDRESS_MODE_BORDER;
1034                                 break;
1035                         default:
1036                                 assert(0);
1037                                 break;
1038                 }
1039
1040                 CUfilter_mode filter_mode;
1041                 if(mem.interpolation == INTERPOLATION_CLOSEST) {
1042                         filter_mode = CU_TR_FILTER_MODE_POINT;
1043                 }
1044                 else {
1045                         filter_mode = CU_TR_FILTER_MODE_LINEAR;
1046                 }
1047
1048                 /* Data Storage */
1049                 if(mem.interpolation == INTERPOLATION_NONE) {
1050                         generic_alloc(mem);
1051                         generic_copy_to(mem);
1052
1053                         CUdeviceptr cumem;
1054                         size_t cubytes;
1055
1056                         cuda_assert(cuModuleGetGlobal(&cumem, &cubytes, cuModule, bind_name.c_str()));
1057
1058                         if(cubytes == 8) {
1059                                 /* 64 bit device pointer */
1060                                 uint64_t ptr = mem.device_pointer;
1061                                 cuda_assert(cuMemcpyHtoD(cumem, (void*)&ptr, cubytes));
1062                         }
1063                         else {
1064                                 /* 32 bit device pointer */
1065                                 uint32_t ptr = (uint32_t)mem.device_pointer;
1066                                 cuda_assert(cuMemcpyHtoD(cumem, (void*)&ptr, cubytes));
1067                         }
1068                         return;
1069                 }
1070
1071                 /* Image Texture Storage */
1072                 CUarray_format_enum format;
1073                 switch(mem.data_type) {
1074                         case TYPE_UCHAR: format = CU_AD_FORMAT_UNSIGNED_INT8; break;
1075                         case TYPE_UINT16: format = CU_AD_FORMAT_UNSIGNED_INT16; break;
1076                         case TYPE_UINT: format = CU_AD_FORMAT_UNSIGNED_INT32; break;
1077                         case TYPE_INT: format = CU_AD_FORMAT_SIGNED_INT32; break;
1078                         case TYPE_FLOAT: format = CU_AD_FORMAT_FLOAT; break;
1079                         case TYPE_HALF: format = CU_AD_FORMAT_HALF; break;
1080                         default: assert(0); return;
1081                 }
1082
1083                 CUDAMem *cmem = NULL;
1084                 CUarray array_3d = NULL;
1085                 size_t src_pitch = mem.data_width * dsize * mem.data_elements;
1086                 size_t dst_pitch = src_pitch;
1087
1088                 if(mem.data_depth > 1) {
1089                         /* 3D texture using array, there is no API for linear memory. */
1090                         CUDA_ARRAY3D_DESCRIPTOR desc;
1091
1092                         desc.Width = mem.data_width;
1093                         desc.Height = mem.data_height;
1094                         desc.Depth = mem.data_depth;
1095                         desc.Format = format;
1096                         desc.NumChannels = mem.data_elements;
1097                         desc.Flags = 0;
1098
1099                         VLOG(1) << "Array 3D allocate: " << mem.name << ", "
1100                                 << string_human_readable_number(mem.memory_size()) << " bytes. ("
1101                                 << string_human_readable_size(mem.memory_size()) << ")";
1102
1103                         cuda_assert(cuArray3DCreate(&array_3d, &desc));
1104
1105                         if(!array_3d) {
1106                                 return;
1107                         }
1108
1109                         CUDA_MEMCPY3D param;
1110                         memset(&param, 0, sizeof(param));
1111                         param.dstMemoryType = CU_MEMORYTYPE_ARRAY;
1112                         param.dstArray = array_3d;
1113                         param.srcMemoryType = CU_MEMORYTYPE_HOST;
1114                         param.srcHost = mem.host_pointer;
1115                         param.srcPitch = src_pitch;
1116                         param.WidthInBytes = param.srcPitch;
1117                         param.Height = mem.data_height;
1118                         param.Depth = mem.data_depth;
1119
1120                         cuda_assert(cuMemcpy3D(&param));
1121
1122                         mem.device_pointer = (device_ptr)array_3d;
1123                         mem.device_size = size;
1124                         stats.mem_alloc(size);
1125
1126                         cmem = &cuda_mem_map[&mem];
1127                         cmem->texobject = 0;
1128                         cmem->array = array_3d;
1129                 }
1130                 else if(mem.data_height > 0) {
1131                         /* 2D texture, using pitch aligned linear memory. */
1132                         int alignment = 0;
1133                         cuda_assert(cuDeviceGetAttribute(&alignment, CU_DEVICE_ATTRIBUTE_TEXTURE_PITCH_ALIGNMENT, cuDevice));
1134                         dst_pitch = align_up(src_pitch, alignment);
1135                         size_t dst_size = dst_pitch * mem.data_height;
1136
1137                         cmem = generic_alloc(mem, dst_size - mem.memory_size());
1138                         if(!cmem) {
1139                                 return;
1140                         }
1141
1142                         CUDA_MEMCPY2D param;
1143                         memset(&param, 0, sizeof(param));
1144                         param.dstMemoryType = CU_MEMORYTYPE_DEVICE;
1145                         param.dstDevice = mem.device_pointer;
1146                         param.dstPitch = dst_pitch;
1147                         param.srcMemoryType = CU_MEMORYTYPE_HOST;
1148                         param.srcHost = mem.host_pointer;
1149                         param.srcPitch = src_pitch;
1150                         param.WidthInBytes = param.srcPitch;
1151                         param.Height = mem.data_height;
1152
1153                         cuda_assert(cuMemcpy2DUnaligned(&param));
1154                 }
1155                 else {
1156                         /* 1D texture, using linear memory. */
1157                         cmem = generic_alloc(mem);
1158                         if(!cmem) {
1159                                 return;
1160                         }
1161
1162                         cuda_assert(cuMemcpyHtoD(mem.device_pointer, mem.host_pointer, size));
1163                 }
1164
1165                 /* Kepler+, bindless textures. */
1166                 int flat_slot = 0;
1167                 if(string_startswith(mem.name, "__tex_image")) {
1168                         int pos =  string(mem.name).rfind("_");
1169                         flat_slot = atoi(mem.name + pos + 1);
1170                 }
1171                 else {
1172                         assert(0);
1173                 }
1174
1175                 CUDA_RESOURCE_DESC resDesc;
1176                 memset(&resDesc, 0, sizeof(resDesc));
1177
1178                 if(array_3d) {
1179                         resDesc.resType = CU_RESOURCE_TYPE_ARRAY;
1180                         resDesc.res.array.hArray = array_3d;
1181                         resDesc.flags = 0;
1182                 }
1183                 else if(mem.data_height > 0) {
1184                         resDesc.resType = CU_RESOURCE_TYPE_PITCH2D;
1185                         resDesc.res.pitch2D.devPtr = mem.device_pointer;
1186                         resDesc.res.pitch2D.format = format;
1187                         resDesc.res.pitch2D.numChannels = mem.data_elements;
1188                         resDesc.res.pitch2D.height = mem.data_height;
1189                         resDesc.res.pitch2D.width = mem.data_width;
1190                         resDesc.res.pitch2D.pitchInBytes = dst_pitch;
1191                 }
1192                 else {
1193                         resDesc.resType = CU_RESOURCE_TYPE_LINEAR;
1194                         resDesc.res.linear.devPtr = mem.device_pointer;
1195                         resDesc.res.linear.format = format;
1196                         resDesc.res.linear.numChannels = mem.data_elements;
1197                         resDesc.res.linear.sizeInBytes = mem.device_size;
1198                 }
1199
1200                 CUDA_TEXTURE_DESC texDesc;
1201                 memset(&texDesc, 0, sizeof(texDesc));
1202                 texDesc.addressMode[0] = address_mode;
1203                 texDesc.addressMode[1] = address_mode;
1204                 texDesc.addressMode[2] = address_mode;
1205                 texDesc.filterMode = filter_mode;
1206                 texDesc.flags = CU_TRSF_NORMALIZED_COORDINATES;
1207
1208                 cuda_assert(cuTexObjectCreate(&cmem->texobject, &resDesc, &texDesc, NULL));
1209
1210                 /* Resize once */
1211                 if(flat_slot >= texture_info.size()) {
1212                         /* Allocate some slots in advance, to reduce amount
1213                          * of re-allocations. */
1214                         texture_info.resize(flat_slot + 128);
1215                 }
1216
1217                 /* Set Mapping and tag that we need to (re-)upload to device */
1218                 TextureInfo& info = texture_info[flat_slot];
1219                 info.data = (uint64_t)cmem->texobject;
1220                 info.cl_buffer = 0;
1221                 info.interpolation = mem.interpolation;
1222                 info.extension = mem.extension;
1223                 info.width = mem.data_width;
1224                 info.height = mem.data_height;
1225                 info.depth = mem.data_depth;
1226                 need_texture_info = true;
1227         }
1228
1229         void tex_free(device_memory& mem)
1230         {
1231                 if(mem.device_pointer) {
1232                         CUDAContextScope scope(this);
1233                         const CUDAMem& cmem = cuda_mem_map[&mem];
1234
1235                         if(cmem.texobject) {
1236                                 /* Free bindless texture. */
1237                                 cuTexObjectDestroy(cmem.texobject);
1238                         }
1239
1240                         if(cmem.array) {
1241                                 /* Free array. */
1242                                 cuArrayDestroy(cmem.array);
1243                                 stats.mem_free(mem.device_size);
1244                                 mem.device_pointer = 0;
1245                                 mem.device_size = 0;
1246
1247                                 cuda_mem_map.erase(cuda_mem_map.find(&mem));
1248                         }
1249                         else {
1250                                 generic_free(mem);
1251                         }
1252                 }
1253         }
1254
1255 #define CUDA_GET_BLOCKSIZE(func, w, h)                                                                          \
1256                         int threads_per_block;                                                                              \
1257                         cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, func)); \
1258                         int threads = (int)sqrt((float)threads_per_block);                                                  \
1259                         int xblocks = ((w) + threads - 1)/threads;                                                          \
1260                         int yblocks = ((h) + threads - 1)/threads;
1261
1262 #define CUDA_LAUNCH_KERNEL(func, args)                      \
1263                         cuda_assert(cuLaunchKernel(func,                \
1264                                                    xblocks, yblocks, 1, \
1265                                                    threads, threads, 1, \
1266                                                    0, 0, args, 0));
1267
1268 /* Similar as above, but for 1-dimensional blocks. */
1269 #define CUDA_GET_BLOCKSIZE_1D(func, w, h)                                                                       \
1270                         int threads_per_block;                                                                              \
1271                         cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, func)); \
1272                         int xblocks = ((w) + threads_per_block - 1)/threads_per_block;                                      \
1273                         int yblocks = h;
1274
1275 #define CUDA_LAUNCH_KERNEL_1D(func, args)                       \
1276                         cuda_assert(cuLaunchKernel(func,                    \
1277                                                    xblocks, yblocks, 1,     \
1278                                                    threads_per_block, 1, 1, \
1279                                                    0, 0, args, 0));
1280
1281         bool denoising_non_local_means(device_ptr image_ptr, device_ptr guide_ptr, device_ptr variance_ptr, device_ptr out_ptr,
1282                                        DenoisingTask *task)
1283         {
1284                 if(have_error())
1285                         return false;
1286
1287                 CUDAContextScope scope(this);
1288
1289                 int stride = task->buffer.stride;
1290                 int w = task->buffer.width;
1291                 int h = task->buffer.h;
1292                 int r = task->nlm_state.r;
1293                 int f = task->nlm_state.f;
1294                 float a = task->nlm_state.a;
1295                 float k_2 = task->nlm_state.k_2;
1296
1297                 int shift_stride = stride*h;
1298                 int num_shifts = (2*r+1)*(2*r+1);
1299                 int mem_size = sizeof(float)*shift_stride*num_shifts;
1300                 int channel_offset = 0;
1301
1302                 device_only_memory<uchar> temporary_mem(this, "Denoising temporary_mem");
1303                 temporary_mem.alloc_to_device(2*mem_size);
1304
1305                 if(have_error())
1306                         return false;
1307
1308                 CUdeviceptr difference     = cuda_device_ptr(temporary_mem.device_pointer);
1309                 CUdeviceptr blurDifference = difference + mem_size;
1310
1311                 CUdeviceptr weightAccum = task->nlm_state.temporary_3_ptr;
1312                 cuda_assert(cuMemsetD8(weightAccum, 0, sizeof(float)*shift_stride));
1313                 cuda_assert(cuMemsetD8(out_ptr, 0, sizeof(float)*shift_stride));
1314
1315                 {
1316                         CUfunction cuNLMCalcDifference, cuNLMBlur, cuNLMCalcWeight, cuNLMUpdateOutput;
1317                         cuda_assert(cuModuleGetFunction(&cuNLMCalcDifference, cuFilterModule, "kernel_cuda_filter_nlm_calc_difference"));
1318                         cuda_assert(cuModuleGetFunction(&cuNLMBlur,           cuFilterModule, "kernel_cuda_filter_nlm_blur"));
1319                         cuda_assert(cuModuleGetFunction(&cuNLMCalcWeight,     cuFilterModule, "kernel_cuda_filter_nlm_calc_weight"));
1320                         cuda_assert(cuModuleGetFunction(&cuNLMUpdateOutput,   cuFilterModule, "kernel_cuda_filter_nlm_update_output"));
1321
1322                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcDifference, CU_FUNC_CACHE_PREFER_L1));
1323                         cuda_assert(cuFuncSetCacheConfig(cuNLMBlur,           CU_FUNC_CACHE_PREFER_L1));
1324                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcWeight,     CU_FUNC_CACHE_PREFER_L1));
1325                         cuda_assert(cuFuncSetCacheConfig(cuNLMUpdateOutput,   CU_FUNC_CACHE_PREFER_L1));
1326
1327                         CUDA_GET_BLOCKSIZE_1D(cuNLMCalcDifference, w*h, num_shifts);
1328
1329                         void *calc_difference_args[] = {&guide_ptr, &variance_ptr, &difference, &w, &h, &stride, &shift_stride, &r, &channel_offset, &a, &k_2};
1330                         void *blur_args[]            = {&difference, &blurDifference, &w, &h, &stride, &shift_stride, &r, &f};
1331                         void *calc_weight_args[]     = {&blurDifference, &difference, &w, &h, &stride, &shift_stride, &r, &f};
1332                         void *update_output_args[]   = {&blurDifference, &image_ptr, &out_ptr, &weightAccum, &w, &h, &stride, &shift_stride, &r, &f};
1333
1334                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcDifference, calc_difference_args);
1335                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1336                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcWeight, calc_weight_args);
1337                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1338                         CUDA_LAUNCH_KERNEL_1D(cuNLMUpdateOutput, update_output_args);
1339                 }
1340
1341                 temporary_mem.free();
1342
1343                 {
1344                         CUfunction cuNLMNormalize;
1345                         cuda_assert(cuModuleGetFunction(&cuNLMNormalize, cuFilterModule, "kernel_cuda_filter_nlm_normalize"));
1346                         cuda_assert(cuFuncSetCacheConfig(cuNLMNormalize, CU_FUNC_CACHE_PREFER_L1));
1347                         void *normalize_args[] = {&out_ptr, &weightAccum, &w, &h, &stride};
1348                         CUDA_GET_BLOCKSIZE(cuNLMNormalize, w, h);
1349                         CUDA_LAUNCH_KERNEL(cuNLMNormalize, normalize_args);
1350                         cuda_assert(cuCtxSynchronize());
1351                 }
1352
1353                 return !have_error();
1354         }
1355
1356         bool denoising_construct_transform(DenoisingTask *task)
1357         {
1358                 if(have_error())
1359                         return false;
1360
1361                 CUDAContextScope scope(this);
1362
1363                 CUfunction cuFilterConstructTransform;
1364                 cuda_assert(cuModuleGetFunction(&cuFilterConstructTransform, cuFilterModule, "kernel_cuda_filter_construct_transform"));
1365                 cuda_assert(cuFuncSetCacheConfig(cuFilterConstructTransform, CU_FUNC_CACHE_PREFER_SHARED));
1366                 CUDA_GET_BLOCKSIZE(cuFilterConstructTransform,
1367                                    task->storage.w,
1368                                    task->storage.h);
1369
1370                 void *args[] = {&task->buffer.mem.device_pointer,
1371                                 &task->storage.transform.device_pointer,
1372                                 &task->storage.rank.device_pointer,
1373                                 &task->filter_area,
1374                                 &task->rect,
1375                                 &task->radius,
1376                                 &task->pca_threshold,
1377                                 &task->buffer.pass_stride};
1378                 CUDA_LAUNCH_KERNEL(cuFilterConstructTransform, args);
1379                 cuda_assert(cuCtxSynchronize());
1380
1381                 return !have_error();
1382         }
1383
1384         bool denoising_reconstruct(device_ptr color_ptr,
1385                                    device_ptr color_variance_ptr,
1386                                    device_ptr output_ptr,
1387                                    DenoisingTask *task)
1388         {
1389                 if(have_error())
1390                         return false;
1391
1392                 CUDAContextScope scope(this);
1393
1394                 mem_zero(task->storage.XtWX);
1395                 mem_zero(task->storage.XtWY);
1396
1397                 int r = task->radius;
1398                 int f = 4;
1399                 float a = 1.0f;
1400                 float k_2 = task->nlm_k_2;
1401
1402                 int w = task->reconstruction_state.source_w;
1403                 int h = task->reconstruction_state.source_h;
1404                 int stride = task->buffer.stride;
1405
1406                 int shift_stride = stride*h;
1407                 int num_shifts = (2*r+1)*(2*r+1);
1408                 int mem_size = sizeof(float)*shift_stride*num_shifts;
1409
1410                 device_only_memory<uchar> temporary_mem(this, "Denoising temporary_mem");
1411                 temporary_mem.alloc_to_device(2*mem_size);
1412
1413                 if(have_error())
1414                         return false;
1415
1416                 CUdeviceptr difference     = cuda_device_ptr(temporary_mem.device_pointer);
1417                 CUdeviceptr blurDifference = difference + mem_size;
1418
1419                 {
1420                         CUfunction cuNLMCalcDifference, cuNLMBlur, cuNLMCalcWeight, cuNLMConstructGramian;
1421                         cuda_assert(cuModuleGetFunction(&cuNLMCalcDifference,   cuFilterModule, "kernel_cuda_filter_nlm_calc_difference"));
1422                         cuda_assert(cuModuleGetFunction(&cuNLMBlur,             cuFilterModule, "kernel_cuda_filter_nlm_blur"));
1423                         cuda_assert(cuModuleGetFunction(&cuNLMCalcWeight,       cuFilterModule, "kernel_cuda_filter_nlm_calc_weight"));
1424                         cuda_assert(cuModuleGetFunction(&cuNLMConstructGramian, cuFilterModule, "kernel_cuda_filter_nlm_construct_gramian"));
1425
1426                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcDifference,   CU_FUNC_CACHE_PREFER_L1));
1427                         cuda_assert(cuFuncSetCacheConfig(cuNLMBlur,             CU_FUNC_CACHE_PREFER_L1));
1428                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcWeight,       CU_FUNC_CACHE_PREFER_L1));
1429                         cuda_assert(cuFuncSetCacheConfig(cuNLMConstructGramian, CU_FUNC_CACHE_PREFER_SHARED));
1430
1431                         CUDA_GET_BLOCKSIZE_1D(cuNLMCalcDifference,
1432                                              task->reconstruction_state.source_w * task->reconstruction_state.source_h,
1433                                              num_shifts);
1434
1435                         void *calc_difference_args[] = {&color_ptr, &color_variance_ptr, &difference, &w, &h, &stride, &shift_stride, &r, &task->buffer.pass_stride, &a, &k_2};
1436                         void *blur_args[]            = {&difference, &blurDifference, &w, &h, &stride, &shift_stride, &r, &f};
1437                         void *calc_weight_args[]     = {&blurDifference, &difference, &w, &h, &stride, &shift_stride, &r, &f};
1438                         void *construct_gramian_args[] = {&blurDifference,
1439                                                           &task->buffer.mem.device_pointer,
1440                                                           &task->storage.transform.device_pointer,
1441                                                           &task->storage.rank.device_pointer,
1442                                                           &task->storage.XtWX.device_pointer,
1443                                                           &task->storage.XtWY.device_pointer,
1444                                                           &task->reconstruction_state.filter_window,
1445                                                           &w, &h, &stride,
1446                                                           &shift_stride, &r,
1447                                                           &f,
1448                                                       &task->buffer.pass_stride};
1449
1450                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcDifference, calc_difference_args);
1451                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1452                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcWeight, calc_weight_args);
1453                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1454                         CUDA_LAUNCH_KERNEL_1D(cuNLMConstructGramian, construct_gramian_args);
1455                 }
1456
1457                 temporary_mem.free();
1458
1459                 {
1460                         CUfunction cuFinalize;
1461                         cuda_assert(cuModuleGetFunction(&cuFinalize, cuFilterModule, "kernel_cuda_filter_finalize"));
1462                         cuda_assert(cuFuncSetCacheConfig(cuFinalize, CU_FUNC_CACHE_PREFER_L1));
1463                         void *finalize_args[] = {&output_ptr,
1464                                                          &task->storage.rank.device_pointer,
1465                                                          &task->storage.XtWX.device_pointer,
1466                                                          &task->storage.XtWY.device_pointer,
1467                                                          &task->filter_area,
1468                                                          &task->reconstruction_state.buffer_params.x,
1469                                                          &task->render_buffer.samples};
1470                         CUDA_GET_BLOCKSIZE(cuFinalize,
1471                                            task->reconstruction_state.source_w,
1472                                            task->reconstruction_state.source_h);
1473                         CUDA_LAUNCH_KERNEL(cuFinalize, finalize_args);
1474                 }
1475
1476                 cuda_assert(cuCtxSynchronize());
1477
1478                 return !have_error();
1479         }
1480
1481         bool denoising_combine_halves(device_ptr a_ptr, device_ptr b_ptr,
1482                                       device_ptr mean_ptr, device_ptr variance_ptr,
1483                                       int r, int4 rect, DenoisingTask *task)
1484         {
1485                 if(have_error())
1486                         return false;
1487
1488                 CUDAContextScope scope(this);
1489
1490                 CUfunction cuFilterCombineHalves;
1491                 cuda_assert(cuModuleGetFunction(&cuFilterCombineHalves, cuFilterModule, "kernel_cuda_filter_combine_halves"));
1492                 cuda_assert(cuFuncSetCacheConfig(cuFilterCombineHalves, CU_FUNC_CACHE_PREFER_L1));
1493                 CUDA_GET_BLOCKSIZE(cuFilterCombineHalves,
1494                                    task->rect.z-task->rect.x,
1495                                    task->rect.w-task->rect.y);
1496
1497                 void *args[] = {&mean_ptr,
1498                                 &variance_ptr,
1499                                 &a_ptr,
1500                                 &b_ptr,
1501                                 &rect,
1502                                 &r};
1503                 CUDA_LAUNCH_KERNEL(cuFilterCombineHalves, args);
1504                 cuda_assert(cuCtxSynchronize());
1505
1506                 return !have_error();
1507         }
1508
1509         bool denoising_divide_shadow(device_ptr a_ptr, device_ptr b_ptr,
1510                                      device_ptr sample_variance_ptr, device_ptr sv_variance_ptr,
1511                                      device_ptr buffer_variance_ptr, DenoisingTask *task)
1512         {
1513                 if(have_error())
1514                         return false;
1515
1516                 CUDAContextScope scope(this);
1517
1518                 CUfunction cuFilterDivideShadow;
1519                 cuda_assert(cuModuleGetFunction(&cuFilterDivideShadow, cuFilterModule, "kernel_cuda_filter_divide_shadow"));
1520                 cuda_assert(cuFuncSetCacheConfig(cuFilterDivideShadow, CU_FUNC_CACHE_PREFER_L1));
1521                 CUDA_GET_BLOCKSIZE(cuFilterDivideShadow,
1522                                    task->rect.z-task->rect.x,
1523                                    task->rect.w-task->rect.y);
1524
1525                 void *args[] = {&task->render_buffer.samples,
1526                                 &task->tile_info_mem.device_pointer,
1527                                 &a_ptr,
1528                                 &b_ptr,
1529                                 &sample_variance_ptr,
1530                                 &sv_variance_ptr,
1531                                 &buffer_variance_ptr,
1532                                 &task->rect,
1533                                 &task->render_buffer.pass_stride,
1534                                 &task->render_buffer.offset};
1535                 CUDA_LAUNCH_KERNEL(cuFilterDivideShadow, args);
1536                 cuda_assert(cuCtxSynchronize());
1537
1538                 return !have_error();
1539         }
1540
1541         bool denoising_get_feature(int mean_offset,
1542                                    int variance_offset,
1543                                    device_ptr mean_ptr,
1544                                    device_ptr variance_ptr,
1545                                    DenoisingTask *task)
1546         {
1547                 if(have_error())
1548                         return false;
1549
1550                 CUDAContextScope scope(this);
1551
1552                 CUfunction cuFilterGetFeature;
1553                 cuda_assert(cuModuleGetFunction(&cuFilterGetFeature, cuFilterModule, "kernel_cuda_filter_get_feature"));
1554                 cuda_assert(cuFuncSetCacheConfig(cuFilterGetFeature, CU_FUNC_CACHE_PREFER_L1));
1555                 CUDA_GET_BLOCKSIZE(cuFilterGetFeature,
1556                                    task->rect.z-task->rect.x,
1557                                    task->rect.w-task->rect.y);
1558
1559                 void *args[] = {&task->render_buffer.samples,
1560                                 &task->tile_info_mem.device_pointer,
1561                                 &mean_offset,
1562                                 &variance_offset,
1563                                 &mean_ptr,
1564                                 &variance_ptr,
1565                                 &task->rect,
1566                                 &task->render_buffer.pass_stride,
1567                                 &task->render_buffer.offset};
1568                 CUDA_LAUNCH_KERNEL(cuFilterGetFeature, args);
1569                 cuda_assert(cuCtxSynchronize());
1570
1571                 return !have_error();
1572         }
1573
1574         bool denoising_detect_outliers(device_ptr image_ptr,
1575                                        device_ptr variance_ptr,
1576                                        device_ptr depth_ptr,
1577                                        device_ptr output_ptr,
1578                                        DenoisingTask *task)
1579         {
1580                 if(have_error())
1581                         return false;
1582
1583                 CUDAContextScope scope(this);
1584
1585                 CUfunction cuFilterDetectOutliers;
1586                 cuda_assert(cuModuleGetFunction(&cuFilterDetectOutliers, cuFilterModule, "kernel_cuda_filter_detect_outliers"));
1587                 cuda_assert(cuFuncSetCacheConfig(cuFilterDetectOutliers, CU_FUNC_CACHE_PREFER_L1));
1588                 CUDA_GET_BLOCKSIZE(cuFilterDetectOutliers,
1589                                    task->rect.z-task->rect.x,
1590                                    task->rect.w-task->rect.y);
1591
1592                 void *args[] = {&image_ptr,
1593                                 &variance_ptr,
1594                                 &depth_ptr,
1595                                 &output_ptr,
1596                                 &task->rect,
1597                                 &task->buffer.pass_stride};
1598
1599                 CUDA_LAUNCH_KERNEL(cuFilterDetectOutliers, args);
1600                 cuda_assert(cuCtxSynchronize());
1601
1602                 return !have_error();
1603         }
1604
1605         void denoise(RenderTile &rtile, DenoisingTask& denoising)
1606         {
1607                 denoising.functions.construct_transform = function_bind(&CUDADevice::denoising_construct_transform, this, &denoising);
1608                 denoising.functions.reconstruct = function_bind(&CUDADevice::denoising_reconstruct, this, _1, _2, _3, &denoising);
1609                 denoising.functions.divide_shadow = function_bind(&CUDADevice::denoising_divide_shadow, this, _1, _2, _3, _4, _5, &denoising);
1610                 denoising.functions.non_local_means = function_bind(&CUDADevice::denoising_non_local_means, this, _1, _2, _3, _4, &denoising);
1611                 denoising.functions.combine_halves = function_bind(&CUDADevice::denoising_combine_halves, this, _1, _2, _3, _4, _5, _6, &denoising);
1612                 denoising.functions.get_feature = function_bind(&CUDADevice::denoising_get_feature, this, _1, _2, _3, _4, &denoising);
1613                 denoising.functions.detect_outliers = function_bind(&CUDADevice::denoising_detect_outliers, this, _1, _2, _3, _4, &denoising);
1614
1615                 denoising.filter_area = make_int4(rtile.x, rtile.y, rtile.w, rtile.h);
1616                 denoising.render_buffer.samples = rtile.sample;
1617
1618                 denoising.run_denoising(&rtile);
1619         }
1620
1621         void path_trace(DeviceTask& task, RenderTile& rtile, device_vector<WorkTile>& work_tiles)
1622         {
1623                 scoped_timer timer(&rtile.buffers->render_time);
1624
1625                 if(have_error())
1626                         return;
1627
1628                 CUDAContextScope scope(this);
1629                 CUfunction cuPathTrace;
1630
1631                 /* Get kernel function. */
1632                 if(task.integrator_branched) {
1633                         cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_branched_path_trace"));
1634                 }
1635                 else {
1636                         cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_path_trace"));
1637                 }
1638
1639                 if(have_error()) {
1640                         return;
1641                 }
1642
1643                 cuda_assert(cuFuncSetCacheConfig(cuPathTrace, CU_FUNC_CACHE_PREFER_L1));
1644
1645                 /* Allocate work tile. */
1646                 work_tiles.alloc(1);
1647
1648                 WorkTile *wtile = work_tiles.data();
1649                 wtile->x = rtile.x;
1650                 wtile->y = rtile.y;
1651                 wtile->w = rtile.w;
1652                 wtile->h = rtile.h;
1653                 wtile->offset = rtile.offset;
1654                 wtile->stride = rtile.stride;
1655                 wtile->buffer = (float*)cuda_device_ptr(rtile.buffer);
1656
1657                 /* Prepare work size. More step samples render faster, but for now we
1658                  * remain conservative for GPUs connected to a display to avoid driver
1659                  * timeouts and display freezing. */
1660                 int min_blocks, num_threads_per_block;
1661                 cuda_assert(cuOccupancyMaxPotentialBlockSize(&min_blocks, &num_threads_per_block, cuPathTrace, NULL, 0, 0));
1662                 if(!info.display_device) {
1663                         min_blocks *= 8;
1664                 }
1665
1666                 uint step_samples = divide_up(min_blocks * num_threads_per_block, wtile->w * wtile->h);
1667
1668                 /* Render all samples. */
1669                 int start_sample = rtile.start_sample;
1670                 int end_sample = rtile.start_sample + rtile.num_samples;
1671
1672                 for(int sample = start_sample; sample < end_sample; sample += step_samples) {
1673                         /* Setup and copy work tile to device. */
1674                         wtile->start_sample = sample;
1675                         wtile->num_samples = min(step_samples, end_sample - sample);;
1676                         work_tiles.copy_to_device();
1677
1678                         CUdeviceptr d_work_tiles = cuda_device_ptr(work_tiles.device_pointer);
1679                         uint total_work_size = wtile->w * wtile->h * wtile->num_samples;
1680                         uint num_blocks = divide_up(total_work_size, num_threads_per_block);
1681
1682                         /* Launch kernel. */
1683                         void *args[] = {&d_work_tiles,
1684                                         &total_work_size};
1685
1686                         cuda_assert(cuLaunchKernel(cuPathTrace,
1687                                                    num_blocks, 1, 1,
1688                                                    num_threads_per_block, 1, 1,
1689                                                    0, 0, args, 0));
1690
1691                         cuda_assert(cuCtxSynchronize());
1692
1693                         /* Update progress. */
1694                         rtile.sample = sample + wtile->num_samples;
1695                         task.update_progress(&rtile, rtile.w*rtile.h*wtile->num_samples);
1696
1697                         if(task.get_cancel()) {
1698                                 if(task.need_finish_queue == false)
1699                                         break;
1700                         }
1701                 }
1702         }
1703
1704         void film_convert(DeviceTask& task, device_ptr buffer, device_ptr rgba_byte, device_ptr rgba_half)
1705         {
1706                 if(have_error())
1707                         return;
1708
1709                 CUDAContextScope scope(this);
1710
1711                 CUfunction cuFilmConvert;
1712                 CUdeviceptr d_rgba = map_pixels((rgba_byte)? rgba_byte: rgba_half);
1713                 CUdeviceptr d_buffer = cuda_device_ptr(buffer);
1714
1715                 /* get kernel function */
1716                 if(rgba_half) {
1717                         cuda_assert(cuModuleGetFunction(&cuFilmConvert, cuModule, "kernel_cuda_convert_to_half_float"));
1718                 }
1719                 else {
1720                         cuda_assert(cuModuleGetFunction(&cuFilmConvert, cuModule, "kernel_cuda_convert_to_byte"));
1721                 }
1722
1723
1724                 float sample_scale = 1.0f/(task.sample + 1);
1725
1726                 /* pass in parameters */
1727                 void *args[] = {&d_rgba,
1728                                 &d_buffer,
1729                                 &sample_scale,
1730                                 &task.x,
1731                                 &task.y,
1732                                 &task.w,
1733                                 &task.h,
1734                                 &task.offset,
1735                                 &task.stride};
1736
1737                 /* launch kernel */
1738                 int threads_per_block;
1739                 cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, cuFilmConvert));
1740
1741                 int xthreads = (int)sqrt(threads_per_block);
1742                 int ythreads = (int)sqrt(threads_per_block);
1743                 int xblocks = (task.w + xthreads - 1)/xthreads;
1744                 int yblocks = (task.h + ythreads - 1)/ythreads;
1745
1746                 cuda_assert(cuFuncSetCacheConfig(cuFilmConvert, CU_FUNC_CACHE_PREFER_L1));
1747
1748                 cuda_assert(cuLaunchKernel(cuFilmConvert,
1749                                            xblocks , yblocks, 1, /* blocks */
1750                                            xthreads, ythreads, 1, /* threads */
1751                                            0, 0, args, 0));
1752
1753                 unmap_pixels((rgba_byte)? rgba_byte: rgba_half);
1754
1755                 cuda_assert(cuCtxSynchronize());
1756         }
1757
1758         void shader(DeviceTask& task)
1759         {
1760                 if(have_error())
1761                         return;
1762
1763                 CUDAContextScope scope(this);
1764
1765                 CUfunction cuShader;
1766                 CUdeviceptr d_input = cuda_device_ptr(task.shader_input);
1767                 CUdeviceptr d_output = cuda_device_ptr(task.shader_output);
1768
1769                 /* get kernel function */
1770                 if(task.shader_eval_type >= SHADER_EVAL_BAKE) {
1771                         cuda_assert(cuModuleGetFunction(&cuShader, cuModule, "kernel_cuda_bake"));
1772                 }
1773                 else if(task.shader_eval_type == SHADER_EVAL_DISPLACE) {
1774                         cuda_assert(cuModuleGetFunction(&cuShader, cuModule, "kernel_cuda_displace"));
1775                 }
1776                 else {
1777                         cuda_assert(cuModuleGetFunction(&cuShader, cuModule, "kernel_cuda_background"));
1778                 }
1779
1780                 /* do tasks in smaller chunks, so we can cancel it */
1781                 const int shader_chunk_size = 65536;
1782                 const int start = task.shader_x;
1783                 const int end = task.shader_x + task.shader_w;
1784                 int offset = task.offset;
1785
1786                 bool canceled = false;
1787                 for(int sample = 0; sample < task.num_samples && !canceled; sample++) {
1788                         for(int shader_x = start; shader_x < end; shader_x += shader_chunk_size) {
1789                                 int shader_w = min(shader_chunk_size, end - shader_x);
1790
1791                                 /* pass in parameters */
1792                                 void *args[8];
1793                                 int arg = 0;
1794                                 args[arg++] = &d_input;
1795                                 args[arg++] = &d_output;
1796                                 args[arg++] = &task.shader_eval_type;
1797                                 if(task.shader_eval_type >= SHADER_EVAL_BAKE) {
1798                                         args[arg++] = &task.shader_filter;
1799                                 }
1800                                 args[arg++] = &shader_x;
1801                                 args[arg++] = &shader_w;
1802                                 args[arg++] = &offset;
1803                                 args[arg++] = &sample;
1804
1805                                 /* launch kernel */
1806                                 int threads_per_block;
1807                                 cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, cuShader));
1808
1809                                 int xblocks = (shader_w + threads_per_block - 1)/threads_per_block;
1810
1811                                 cuda_assert(cuFuncSetCacheConfig(cuShader, CU_FUNC_CACHE_PREFER_L1));
1812                                 cuda_assert(cuLaunchKernel(cuShader,
1813                                                            xblocks , 1, 1, /* blocks */
1814                                                            threads_per_block, 1, 1, /* threads */
1815                                                            0, 0, args, 0));
1816
1817                                 cuda_assert(cuCtxSynchronize());
1818
1819                                 if(task.get_cancel()) {
1820                                         canceled = true;
1821                                         break;
1822                                 }
1823                         }
1824
1825                         task.update_progress(NULL);
1826                 }
1827         }
1828
1829         CUdeviceptr map_pixels(device_ptr mem)
1830         {
1831                 if(!background) {
1832                         PixelMem pmem = pixel_mem_map[mem];
1833                         CUdeviceptr buffer;
1834
1835                         size_t bytes;
1836                         cuda_assert(cuGraphicsMapResources(1, &pmem.cuPBOresource, 0));
1837                         cuda_assert(cuGraphicsResourceGetMappedPointer(&buffer, &bytes, pmem.cuPBOresource));
1838
1839                         return buffer;
1840                 }
1841
1842                 return cuda_device_ptr(mem);
1843         }
1844
1845         void unmap_pixels(device_ptr mem)
1846         {
1847                 if(!background) {
1848                         PixelMem pmem = pixel_mem_map[mem];
1849
1850                         cuda_assert(cuGraphicsUnmapResources(1, &pmem.cuPBOresource, 0));
1851                 }
1852         }
1853
1854         void pixels_alloc(device_memory& mem)
1855         {
1856                 PixelMem pmem;
1857
1858                 pmem.w = mem.data_width;
1859                 pmem.h = mem.data_height;
1860
1861                 CUDAContextScope scope(this);
1862
1863                 glGenBuffers(1, &pmem.cuPBO);
1864                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO);
1865                 if(mem.data_type == TYPE_HALF)
1866                         glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(GLhalf)*4, NULL, GL_DYNAMIC_DRAW);
1867                 else
1868                         glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(uint8_t)*4, NULL, GL_DYNAMIC_DRAW);
1869
1870                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
1871
1872                 glGenTextures(1, &pmem.cuTexId);
1873                 glBindTexture(GL_TEXTURE_2D, pmem.cuTexId);
1874                 if(mem.data_type == TYPE_HALF)
1875                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F, pmem.w, pmem.h, 0, GL_RGBA, GL_HALF_FLOAT, NULL);
1876                 else
1877                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, pmem.w, pmem.h, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1878                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1879                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1880                 glBindTexture(GL_TEXTURE_2D, 0);
1881
1882                 CUresult result = cuGraphicsGLRegisterBuffer(&pmem.cuPBOresource, pmem.cuPBO, CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE);
1883
1884                 if(result == CUDA_SUCCESS) {
1885                         mem.device_pointer = pmem.cuTexId;
1886                         pixel_mem_map[mem.device_pointer] = pmem;
1887
1888                         mem.device_size = mem.memory_size();
1889                         stats.mem_alloc(mem.device_size);
1890
1891                         return;
1892                 }
1893                 else {
1894                         /* failed to register buffer, fallback to no interop */
1895                         glDeleteBuffers(1, &pmem.cuPBO);
1896                         glDeleteTextures(1, &pmem.cuTexId);
1897
1898                         background = true;
1899                 }
1900         }
1901
1902         void pixels_copy_from(device_memory& mem, int y, int w, int h)
1903         {
1904                 PixelMem pmem = pixel_mem_map[mem.device_pointer];
1905
1906                 CUDAContextScope scope(this);
1907
1908                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO);
1909                 uchar *pixels = (uchar*)glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_ONLY);
1910                 size_t offset = sizeof(uchar)*4*y*w;
1911                 memcpy((uchar*)mem.host_pointer + offset, pixels + offset, sizeof(uchar)*4*w*h);
1912                 glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
1913                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
1914         }
1915
1916         void pixels_free(device_memory& mem)
1917         {
1918                 if(mem.device_pointer) {
1919                         PixelMem pmem = pixel_mem_map[mem.device_pointer];
1920
1921                         CUDAContextScope scope(this);
1922
1923                         cuda_assert(cuGraphicsUnregisterResource(pmem.cuPBOresource));
1924                         glDeleteBuffers(1, &pmem.cuPBO);
1925                         glDeleteTextures(1, &pmem.cuTexId);
1926
1927                         pixel_mem_map.erase(pixel_mem_map.find(mem.device_pointer));
1928                         mem.device_pointer = 0;
1929
1930                         stats.mem_free(mem.device_size);
1931                         mem.device_size = 0;
1932                 }
1933         }
1934
1935         void draw_pixels(
1936             device_memory& mem, int y,
1937             int w, int h, int width, int height,
1938             int dx, int dy, int dw, int dh, bool transparent,
1939                 const DeviceDrawParams &draw_params)
1940         {
1941                 assert(mem.type == MEM_PIXELS);
1942
1943                 if(!background) {
1944                         const bool use_fallback_shader = (draw_params.bind_display_space_shader_cb == NULL);
1945                         PixelMem pmem = pixel_mem_map[mem.device_pointer];
1946                         float *vpointer;
1947
1948                         CUDAContextScope scope(this);
1949
1950                         /* for multi devices, this assumes the inefficient method that we allocate
1951                          * all pixels on the device even though we only render to a subset */
1952                         size_t offset = 4*y*w;
1953
1954                         if(mem.data_type == TYPE_HALF)
1955                                 offset *= sizeof(GLhalf);
1956                         else
1957                                 offset *= sizeof(uint8_t);
1958
1959                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO);
1960                         glBindTexture(GL_TEXTURE_2D, pmem.cuTexId);
1961                         if(mem.data_type == TYPE_HALF) {
1962                                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, w, h, GL_RGBA, GL_HALF_FLOAT, (void*)offset);
1963                         }
1964                         else {
1965                                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, w, h, GL_RGBA, GL_UNSIGNED_BYTE, (void*)offset);
1966                         }
1967                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
1968
1969                         if(transparent) {
1970                                 glEnable(GL_BLEND);
1971                                 glBlendFunc(GL_ONE, GL_ONE_MINUS_SRC_ALPHA);
1972                         }
1973
1974                         GLint shader_program;
1975                         if(use_fallback_shader) {
1976                                 if(!bind_fallback_display_space_shader(dw, dh)) {
1977                                         return;
1978                                 }
1979                                 shader_program = fallback_shader_program;
1980                         }
1981                         else {
1982                                 draw_params.bind_display_space_shader_cb();
1983                                 glGetIntegerv(GL_CURRENT_PROGRAM, &shader_program);
1984                         }
1985
1986                         if(!vertex_buffer) {
1987                                 glGenBuffers(1, &vertex_buffer);
1988                         }
1989
1990                         glBindBuffer(GL_ARRAY_BUFFER, vertex_buffer);
1991                         /* invalidate old contents - avoids stalling if buffer is still waiting in queue to be rendered */
1992                         glBufferData(GL_ARRAY_BUFFER, 16 * sizeof(float), NULL, GL_STREAM_DRAW);
1993
1994                         vpointer = (float *)glMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY);
1995
1996                         if(vpointer) {
1997                                 /* texture coordinate - vertex pair */
1998                                 vpointer[0] = 0.0f;
1999                                 vpointer[1] = 0.0f;
2000                                 vpointer[2] = dx;
2001                                 vpointer[3] = dy;
2002
2003                                 vpointer[4] = (float)w/(float)pmem.w;
2004                                 vpointer[5] = 0.0f;
2005                                 vpointer[6] = (float)width + dx;
2006                                 vpointer[7] = dy;
2007
2008                                 vpointer[8] = (float)w/(float)pmem.w;
2009                                 vpointer[9] = (float)h/(float)pmem.h;
2010                                 vpointer[10] = (float)width + dx;
2011                                 vpointer[11] = (float)height + dy;
2012
2013                                 vpointer[12] = 0.0f;
2014                                 vpointer[13] = (float)h/(float)pmem.h;
2015                                 vpointer[14] = dx;
2016                                 vpointer[15] = (float)height + dy;
2017
2018                                 glUnmapBuffer(GL_ARRAY_BUFFER);
2019                         }
2020
2021                         GLuint vertex_array_object;
2022                         GLuint position_attribute, texcoord_attribute;
2023
2024                         glGenVertexArrays(1, &vertex_array_object);
2025                         glBindVertexArray(vertex_array_object);
2026
2027                         texcoord_attribute = glGetAttribLocation(shader_program, "texCoord");
2028                         position_attribute = glGetAttribLocation(shader_program, "pos");
2029
2030                         glEnableVertexAttribArray(texcoord_attribute);
2031                         glEnableVertexAttribArray(position_attribute);
2032
2033                         glVertexAttribPointer(texcoord_attribute, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (const GLvoid *)0);
2034                         glVertexAttribPointer(position_attribute, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (const GLvoid *)(sizeof(float) * 2));
2035
2036                         glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
2037
2038                         if(use_fallback_shader) {
2039                                 glUseProgram(0);
2040                         }
2041                         else {
2042                                 draw_params.unbind_display_space_shader_cb();
2043                         }
2044
2045                         if(transparent) {
2046                                 glDisable(GL_BLEND);
2047                         }
2048
2049                         glBindTexture(GL_TEXTURE_2D, 0);
2050
2051                         return;
2052                 }
2053
2054                 Device::draw_pixels(mem, y, w, h, width, height, dx, dy, dw, dh, transparent, draw_params);
2055         }
2056
2057         void thread_run(DeviceTask *task)
2058         {
2059                 CUDAContextScope scope(this);
2060
2061                 if(task->type == DeviceTask::RENDER) {
2062                         DeviceRequestedFeatures requested_features;
2063                         if(use_split_kernel()) {
2064                                 if(split_kernel == NULL) {
2065                                         split_kernel = new CUDASplitKernel(this);
2066                                         split_kernel->load_kernels(requested_features);
2067                                 }
2068                         }
2069
2070                         device_vector<WorkTile> work_tiles(this, "work_tiles", MEM_READ_ONLY);
2071
2072                         /* keep rendering tiles until done */
2073                         RenderTile tile;
2074                         DenoisingTask denoising(this, *task);
2075
2076                         while(task->acquire_tile(this, tile)) {
2077                                 if(tile.task == RenderTile::PATH_TRACE) {
2078                                         if(use_split_kernel()) {
2079                                                 device_only_memory<uchar> void_buffer(this, "void_buffer");
2080                                                 split_kernel->path_trace(task, tile, void_buffer, void_buffer);
2081                                         }
2082                                         else {
2083                                                 path_trace(*task, tile, work_tiles);
2084                                         }
2085                                 }
2086                                 else if(tile.task == RenderTile::DENOISE) {
2087                                         tile.sample = tile.start_sample + tile.num_samples;
2088
2089                                         denoise(tile, denoising);
2090
2091                                         task->update_progress(&tile, tile.w*tile.h);
2092                                 }
2093
2094                                 task->release_tile(tile);
2095
2096                                 if(task->get_cancel()) {
2097                                         if(task->need_finish_queue == false)
2098                                                 break;
2099                                 }
2100                         }
2101
2102                         work_tiles.free();
2103                 }
2104                 else if(task->type == DeviceTask::SHADER) {
2105                         shader(*task);
2106
2107                         cuda_assert(cuCtxSynchronize());
2108                 }
2109         }
2110
2111         class CUDADeviceTask : public DeviceTask {
2112         public:
2113                 CUDADeviceTask(CUDADevice *device, DeviceTask& task)
2114                 : DeviceTask(task)
2115                 {
2116                         run = function_bind(&CUDADevice::thread_run, device, this);
2117                 }
2118         };
2119
2120         int get_split_task_count(DeviceTask& /*task*/)
2121         {
2122                 return 1;
2123         }
2124
2125         void task_add(DeviceTask& task)
2126         {
2127                 CUDAContextScope scope(this);
2128
2129                 /* Load texture info. */
2130                 load_texture_info();
2131
2132                 /* Synchronize all memory copies before executing task. */
2133                 cuda_assert(cuCtxSynchronize());
2134
2135                 if(task.type == DeviceTask::FILM_CONVERT) {
2136                         /* must be done in main thread due to opengl access */
2137                         film_convert(task, task.buffer, task.rgba_byte, task.rgba_half);
2138                 }
2139                 else {
2140                         task_pool.push(new CUDADeviceTask(this, task));
2141                 }
2142         }
2143
2144         void task_wait()
2145         {
2146                 task_pool.wait();
2147         }
2148
2149         void task_cancel()
2150         {
2151                 task_pool.cancel();
2152         }
2153
2154         friend class CUDASplitKernelFunction;
2155         friend class CUDASplitKernel;
2156         friend class CUDAContextScope;
2157 };
2158
2159 /* redefine the cuda_assert macro so it can be used outside of the CUDADevice class
2160  * now that the definition of that class is complete
2161  */
2162 #undef cuda_assert
2163 #define cuda_assert(stmt) \
2164         { \
2165                 CUresult result = stmt; \
2166                 \
2167                 if(result != CUDA_SUCCESS) { \
2168                         string message = string_printf("CUDA error: %s in %s", cuewErrorString(result), #stmt); \
2169                         if(device->error_msg == "") \
2170                                 device->error_msg = message; \
2171                         fprintf(stderr, "%s\n", message.c_str()); \
2172                         /*cuda_abort();*/ \
2173                         device->cuda_error_documentation(); \
2174                 } \
2175         } (void)0
2176
2177
2178 /* CUDA context scope. */
2179
2180 CUDAContextScope::CUDAContextScope(CUDADevice *device)
2181 : device(device)
2182 {
2183         cuda_assert(cuCtxPushCurrent(device->cuContext));
2184 }
2185
2186 CUDAContextScope::~CUDAContextScope()
2187 {
2188         cuda_assert(cuCtxPopCurrent(NULL));
2189 }
2190
2191 /* split kernel */
2192
2193 class CUDASplitKernelFunction : public SplitKernelFunction{
2194         CUDADevice* device;
2195         CUfunction func;
2196 public:
2197         CUDASplitKernelFunction(CUDADevice *device, CUfunction func) : device(device), func(func) {}
2198
2199         /* enqueue the kernel, returns false if there is an error */
2200         bool enqueue(const KernelDimensions &dim, device_memory &/*kg*/, device_memory &/*data*/)
2201         {
2202                 return enqueue(dim, NULL);
2203         }
2204
2205         /* enqueue the kernel, returns false if there is an error */
2206         bool enqueue(const KernelDimensions &dim, void *args[])
2207         {
2208                 if(device->have_error())
2209                         return false;
2210
2211                 CUDAContextScope scope(device);
2212
2213                 /* we ignore dim.local_size for now, as this is faster */
2214                 int threads_per_block;
2215                 cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, func));
2216
2217                 int xblocks = (dim.global_size[0]*dim.global_size[1] + threads_per_block - 1)/threads_per_block;
2218
2219                 cuda_assert(cuFuncSetCacheConfig(func, CU_FUNC_CACHE_PREFER_L1));
2220
2221                 cuda_assert(cuLaunchKernel(func,
2222                                            xblocks, 1, 1, /* blocks */
2223                                            threads_per_block, 1, 1, /* threads */
2224                                            0, 0, args, 0));
2225
2226                 return !device->have_error();
2227         }
2228 };
2229
2230 CUDASplitKernel::CUDASplitKernel(CUDADevice *device) : DeviceSplitKernel(device), device(device)
2231 {
2232 }
2233
2234 uint64_t CUDASplitKernel::state_buffer_size(device_memory& /*kg*/, device_memory& /*data*/, size_t num_threads)
2235 {
2236         CUDAContextScope scope(device);
2237
2238         device_vector<uint64_t> size_buffer(device, "size_buffer", MEM_READ_WRITE);
2239         size_buffer.alloc(1);
2240         size_buffer.zero_to_device();
2241
2242         uint threads = num_threads;
2243         CUdeviceptr d_size = device->cuda_device_ptr(size_buffer.device_pointer);
2244
2245         struct args_t {
2246                 uint* num_threads;
2247                 CUdeviceptr* size;
2248         };
2249
2250         args_t args = {
2251                 &threads,
2252                 &d_size
2253         };
2254
2255         CUfunction state_buffer_size;
2256         cuda_assert(cuModuleGetFunction(&state_buffer_size, device->cuModule, "kernel_cuda_state_buffer_size"));
2257
2258         cuda_assert(cuLaunchKernel(state_buffer_size,
2259                                    1, 1, 1,
2260                                    1, 1, 1,
2261                                    0, 0, (void**)&args, 0));
2262
2263         size_buffer.copy_from_device(0, 1, 1);
2264         size_t size = size_buffer[0];
2265         size_buffer.free();
2266
2267         return size;
2268 }
2269
2270 bool CUDASplitKernel::enqueue_split_kernel_data_init(const KernelDimensions& dim,
2271                                     RenderTile& rtile,
2272                                     int num_global_elements,
2273                                     device_memory& /*kernel_globals*/,
2274                                     device_memory& /*kernel_data*/,
2275                                     device_memory& split_data,
2276                                     device_memory& ray_state,
2277                                     device_memory& queue_index,
2278                                     device_memory& use_queues_flag,
2279                                     device_memory& work_pool_wgs)
2280 {
2281         CUDAContextScope scope(device);
2282
2283         CUdeviceptr d_split_data = device->cuda_device_ptr(split_data.device_pointer);
2284         CUdeviceptr d_ray_state = device->cuda_device_ptr(ray_state.device_pointer);
2285         CUdeviceptr d_queue_index = device->cuda_device_ptr(queue_index.device_pointer);
2286         CUdeviceptr d_use_queues_flag = device->cuda_device_ptr(use_queues_flag.device_pointer);
2287         CUdeviceptr d_work_pool_wgs = device->cuda_device_ptr(work_pool_wgs.device_pointer);
2288
2289         CUdeviceptr d_buffer = device->cuda_device_ptr(rtile.buffer);
2290
2291         int end_sample = rtile.start_sample + rtile.num_samples;
2292         int queue_size = dim.global_size[0] * dim.global_size[1];
2293
2294         struct args_t {
2295                 CUdeviceptr* split_data_buffer;
2296                 int* num_elements;
2297                 CUdeviceptr* ray_state;
2298                 int* start_sample;
2299                 int* end_sample;
2300                 int* sx;
2301                 int* sy;
2302                 int* sw;
2303                 int* sh;
2304                 int* offset;
2305                 int* stride;
2306                 CUdeviceptr* queue_index;
2307                 int* queuesize;
2308                 CUdeviceptr* use_queues_flag;
2309                 CUdeviceptr* work_pool_wgs;
2310                 int* num_samples;
2311                 CUdeviceptr* buffer;
2312         };
2313
2314         args_t args = {
2315                 &d_split_data,
2316                 &num_global_elements,
2317                 &d_ray_state,
2318                 &rtile.start_sample,
2319                 &end_sample,
2320                 &rtile.x,
2321                 &rtile.y,
2322                 &rtile.w,
2323                 &rtile.h,
2324                 &rtile.offset,
2325                 &rtile.stride,
2326                 &d_queue_index,
2327                 &queue_size,
2328                 &d_use_queues_flag,
2329                 &d_work_pool_wgs,
2330                 &rtile.num_samples,
2331                 &d_buffer
2332         };
2333
2334         CUfunction data_init;
2335         cuda_assert(cuModuleGetFunction(&data_init, device->cuModule, "kernel_cuda_path_trace_data_init"));
2336         if(device->have_error()) {
2337                 return false;
2338         }
2339
2340         CUDASplitKernelFunction(device, data_init).enqueue(dim, (void**)&args);
2341
2342         return !device->have_error();
2343 }
2344
2345 SplitKernelFunction* CUDASplitKernel::get_split_kernel_function(const string& kernel_name,
2346                                                                 const DeviceRequestedFeatures&)
2347 {
2348         CUDAContextScope scope(device);
2349         CUfunction func;
2350
2351         cuda_assert(cuModuleGetFunction(&func, device->cuModule, (string("kernel_cuda_") + kernel_name).data()));
2352         if(device->have_error()) {
2353                 device->cuda_error_message(string_printf("kernel \"kernel_cuda_%s\" not found in module", kernel_name.data()));
2354                 return NULL;
2355         }
2356
2357         return new CUDASplitKernelFunction(device, func);
2358 }
2359
2360 int2 CUDASplitKernel::split_kernel_local_size()
2361 {
2362         return make_int2(32, 1);
2363 }
2364
2365 int2 CUDASplitKernel::split_kernel_global_size(device_memory& kg, device_memory& data, DeviceTask * /*task*/)
2366 {
2367         CUDAContextScope scope(device);
2368         size_t free;
2369         size_t total;
2370
2371         cuda_assert(cuMemGetInfo(&free, &total));
2372
2373         VLOG(1) << "Maximum device allocation size: "
2374                 << string_human_readable_number(free) << " bytes. ("
2375                 << string_human_readable_size(free) << ").";
2376
2377         size_t num_elements = max_elements_for_max_buffer_size(kg, data, free / 2);
2378         size_t side = round_down((int)sqrt(num_elements), 32);
2379         int2 global_size = make_int2(side, round_down(num_elements / side, 16));
2380         VLOG(1) << "Global size: " << global_size << ".";
2381         return global_size;
2382 }
2383
2384 bool device_cuda_init(void)
2385 {
2386 #ifdef WITH_CUDA_DYNLOAD
2387         static bool initialized = false;
2388         static bool result = false;
2389
2390         if(initialized)
2391                 return result;
2392
2393         initialized = true;
2394         int cuew_result = cuewInit(CUEW_INIT_CUDA);
2395         if(cuew_result == CUEW_SUCCESS) {
2396                 VLOG(1) << "CUEW initialization succeeded";
2397                 if(CUDADevice::have_precompiled_kernels()) {
2398                         VLOG(1) << "Found precompiled kernels";
2399                         result = true;
2400                 }
2401 #ifndef _WIN32
2402                 else if(cuewCompilerPath() != NULL) {
2403                         VLOG(1) << "Found CUDA compiler " << cuewCompilerPath();
2404                         result = true;
2405                 }
2406                 else {
2407                         VLOG(1) << "Neither precompiled kernels nor CUDA compiler was found,"
2408                                 << " unable to use CUDA";
2409                 }
2410 #endif
2411         }
2412         else {
2413                 VLOG(1) << "CUEW initialization failed: "
2414                         << ((cuew_result == CUEW_ERROR_ATEXIT_FAILED)
2415                             ? "Error setting up atexit() handler"
2416                             : "Error opening the library");
2417         }
2418
2419         return result;
2420 #else  /* WITH_CUDA_DYNLOAD */
2421         return true;
2422 #endif /* WITH_CUDA_DYNLOAD */
2423 }
2424
2425 Device *device_cuda_create(DeviceInfo& info, Stats &stats, bool background)
2426 {
2427         return new CUDADevice(info, stats, background);
2428 }
2429
2430 static CUresult device_cuda_safe_init()
2431 {
2432 #ifdef _WIN32
2433         __try {
2434                 return cuInit(0);
2435         }
2436         __except(EXCEPTION_EXECUTE_HANDLER) {
2437                 /* Ignore crashes inside the CUDA driver and hope we can
2438                  * survive even with corrupted CUDA installs. */
2439                 fprintf(stderr, "Cycles CUDA: driver crashed, continuing without CUDA.\n");
2440         }
2441
2442         return CUDA_ERROR_NO_DEVICE;
2443 #else
2444         return cuInit(0);
2445 #endif
2446 }
2447
2448 void device_cuda_info(vector<DeviceInfo>& devices)
2449 {
2450         CUresult result = device_cuda_safe_init();
2451         if(result != CUDA_SUCCESS) {
2452                 if(result != CUDA_ERROR_NO_DEVICE)
2453                         fprintf(stderr, "CUDA cuInit: %s\n", cuewErrorString(result));
2454                 return;
2455         }
2456
2457         int count = 0;
2458         result = cuDeviceGetCount(&count);
2459         if(result != CUDA_SUCCESS) {
2460                 fprintf(stderr, "CUDA cuDeviceGetCount: %s\n", cuewErrorString(result));
2461                 return;
2462         }
2463
2464         vector<DeviceInfo> display_devices;
2465
2466         for(int num = 0; num < count; num++) {
2467                 char name[256];
2468
2469                 result = cuDeviceGetName(name, 256, num);
2470                 if(result != CUDA_SUCCESS) {
2471                         fprintf(stderr, "CUDA cuDeviceGetName: %s\n", cuewErrorString(result));
2472                         continue;
2473                 }
2474
2475                 int major;
2476                 cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, num);
2477                 if(major < 3) {
2478                         VLOG(1) << "Ignoring device \"" << name
2479                                 << "\", this graphics card is no longer supported.";
2480                         continue;
2481                 }
2482
2483                 DeviceInfo info;
2484
2485                 info.type = DEVICE_CUDA;
2486                 info.description = string(name);
2487                 info.num = num;
2488
2489                 info.advanced_shading = (major >= 3);
2490                 info.has_half_images = (major >= 3);
2491                 info.has_volume_decoupled = false;
2492                 info.bvh_layout_mask = BVH_LAYOUT_BVH2;
2493
2494                 int pci_location[3] = {0, 0, 0};
2495                 cuDeviceGetAttribute(&pci_location[0], CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, num);
2496                 cuDeviceGetAttribute(&pci_location[1], CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, num);
2497                 cuDeviceGetAttribute(&pci_location[2], CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, num);
2498                 info.id = string_printf("CUDA_%s_%04x:%02x:%02x",
2499                                         name,
2500                                         (unsigned int)pci_location[0],
2501                                         (unsigned int)pci_location[1],
2502                                         (unsigned int)pci_location[2]);
2503
2504                 /* If device has a kernel timeout and no compute preemption, we assume
2505                  * it is connected to a display and will freeze the display while doing
2506                  * computations. */
2507                 int timeout_attr = 0, preempt_attr = 0;
2508                 cuDeviceGetAttribute(&timeout_attr, CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, num);
2509                 cuDeviceGetAttribute(&preempt_attr, CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED, num);
2510
2511                 if(timeout_attr && !preempt_attr) {
2512                         VLOG(1) << "Device is recognized as display.";
2513                         info.description += " (Display)";
2514                         info.display_device = true;
2515                         display_devices.push_back(info);
2516                 }
2517                 else {
2518                         devices.push_back(info);
2519                 }
2520                 VLOG(1) << "Added device \"" << name << "\" with id \"" << info.id << "\".";
2521         }
2522
2523         if(!display_devices.empty())
2524                 devices.insert(devices.end(), display_devices.begin(), display_devices.end());
2525 }
2526
2527 string device_cuda_capabilities(void)
2528 {
2529         CUresult result = device_cuda_safe_init();
2530         if(result != CUDA_SUCCESS) {
2531                 if(result != CUDA_ERROR_NO_DEVICE) {
2532                         return string("Error initializing CUDA: ") + cuewErrorString(result);
2533                 }
2534                 return "No CUDA device found\n";
2535         }
2536
2537         int count;
2538         result = cuDeviceGetCount(&count);
2539         if(result != CUDA_SUCCESS) {
2540                 return string("Error getting devices: ") + cuewErrorString(result);
2541         }
2542
2543         string capabilities = "";
2544         for(int num = 0; num < count; num++) {
2545                 char name[256];
2546                 if(cuDeviceGetName(name, 256, num) != CUDA_SUCCESS) {
2547                         continue;
2548                 }
2549                 capabilities += string("\t") + name + "\n";
2550                 int value;
2551 #define GET_ATTR(attr) \
2552                 { \
2553                         if(cuDeviceGetAttribute(&value, \
2554                                                 CU_DEVICE_ATTRIBUTE_##attr, \
2555                                                 num) == CUDA_SUCCESS) \
2556                         { \
2557                                 capabilities += string_printf("\t\tCU_DEVICE_ATTRIBUTE_" #attr "\t\t\t%d\n", \
2558                                                               value); \
2559                         } \
2560                 } (void)0
2561                 /* TODO(sergey): Strip all attributes which are not useful for us
2562                  * or does not depend on the driver.
2563                  */
2564                 GET_ATTR(MAX_THREADS_PER_BLOCK);
2565                 GET_ATTR(MAX_BLOCK_DIM_X);
2566                 GET_ATTR(MAX_BLOCK_DIM_Y);
2567                 GET_ATTR(MAX_BLOCK_DIM_Z);
2568                 GET_ATTR(MAX_GRID_DIM_X);
2569                 GET_ATTR(MAX_GRID_DIM_Y);
2570                 GET_ATTR(MAX_GRID_DIM_Z);
2571                 GET_ATTR(MAX_SHARED_MEMORY_PER_BLOCK);
2572                 GET_ATTR(SHARED_MEMORY_PER_BLOCK);
2573                 GET_ATTR(TOTAL_CONSTANT_MEMORY);
2574                 GET_ATTR(WARP_SIZE);
2575                 GET_ATTR(MAX_PITCH);
2576                 GET_ATTR(MAX_REGISTERS_PER_BLOCK);
2577                 GET_ATTR(REGISTERS_PER_BLOCK);
2578                 GET_ATTR(CLOCK_RATE);
2579                 GET_ATTR(TEXTURE_ALIGNMENT);
2580                 GET_ATTR(GPU_OVERLAP);
2581                 GET_ATTR(MULTIPROCESSOR_COUNT);
2582                 GET_ATTR(KERNEL_EXEC_TIMEOUT);
2583                 GET_ATTR(INTEGRATED);
2584                 GET_ATTR(CAN_MAP_HOST_MEMORY);
2585                 GET_ATTR(COMPUTE_MODE);
2586                 GET_ATTR(MAXIMUM_TEXTURE1D_WIDTH);
2587                 GET_ATTR(MAXIMUM_TEXTURE2D_WIDTH);
2588                 GET_ATTR(MAXIMUM_TEXTURE2D_HEIGHT);
2589                 GET_ATTR(MAXIMUM_TEXTURE3D_WIDTH);
2590                 GET_ATTR(MAXIMUM_TEXTURE3D_HEIGHT);
2591                 GET_ATTR(MAXIMUM_TEXTURE3D_DEPTH);
2592                 GET_ATTR(MAXIMUM_TEXTURE2D_LAYERED_WIDTH);
2593                 GET_ATTR(MAXIMUM_TEXTURE2D_LAYERED_HEIGHT);
2594                 GET_ATTR(MAXIMUM_TEXTURE2D_LAYERED_LAYERS);
2595                 GET_ATTR(MAXIMUM_TEXTURE2D_ARRAY_WIDTH);
2596                 GET_ATTR(MAXIMUM_TEXTURE2D_ARRAY_HEIGHT);
2597                 GET_ATTR(MAXIMUM_TEXTURE2D_ARRAY_NUMSLICES);
2598                 GET_ATTR(SURFACE_ALIGNMENT);
2599                 GET_ATTR(CONCURRENT_KERNELS);
2600                 GET_ATTR(ECC_ENABLED);
2601                 GET_ATTR(TCC_DRIVER);
2602                 GET_ATTR(MEMORY_CLOCK_RATE);
2603                 GET_ATTR(GLOBAL_MEMORY_BUS_WIDTH);
2604                 GET_ATTR(L2_CACHE_SIZE);
2605                 GET_ATTR(MAX_THREADS_PER_MULTIPROCESSOR);
2606                 GET_ATTR(ASYNC_ENGINE_COUNT);
2607                 GET_ATTR(UNIFIED_ADDRESSING);
2608                 GET_ATTR(MAXIMUM_TEXTURE1D_LAYERED_WIDTH);
2609                 GET_ATTR(MAXIMUM_TEXTURE1D_LAYERED_LAYERS);
2610                 GET_ATTR(CAN_TEX2D_GATHER);
2611                 GET_ATTR(MAXIMUM_TEXTURE2D_GATHER_WIDTH);
2612                 GET_ATTR(MAXIMUM_TEXTURE2D_GATHER_HEIGHT);
2613                 GET_ATTR(MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE);
2614                 GET_ATTR(MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE);
2615                 GET_ATTR(MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE);
2616                 GET_ATTR(TEXTURE_PITCH_ALIGNMENT);
2617                 GET_ATTR(MAXIMUM_TEXTURECUBEMAP_WIDTH);
2618                 GET_ATTR(MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH);
2619                 GET_ATTR(MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS);
2620                 GET_ATTR(MAXIMUM_SURFACE1D_WIDTH);
2621                 GET_ATTR(MAXIMUM_SURFACE2D_WIDTH);
2622                 GET_ATTR(MAXIMUM_SURFACE2D_HEIGHT);
2623                 GET_ATTR(MAXIMUM_SURFACE3D_WIDTH);
2624                 GET_ATTR(MAXIMUM_SURFACE3D_HEIGHT);
2625                 GET_ATTR(MAXIMUM_SURFACE3D_DEPTH);
2626                 GET_ATTR(MAXIMUM_SURFACE1D_LAYERED_WIDTH);
2627                 GET_ATTR(MAXIMUM_SURFACE1D_LAYERED_LAYERS);
2628                 GET_ATTR(MAXIMUM_SURFACE2D_LAYERED_WIDTH);
2629                 GET_ATTR(MAXIMUM_SURFACE2D_LAYERED_HEIGHT);
2630                 GET_ATTR(MAXIMUM_SURFACE2D_LAYERED_LAYERS);
2631                 GET_ATTR(MAXIMUM_SURFACECUBEMAP_WIDTH);
2632                 GET_ATTR(MAXIMUM_SURFACECUBEMAP_LAYERED_WIDTH);
2633                 GET_ATTR(MAXIMUM_SURFACECUBEMAP_LAYERED_LAYERS);
2634                 GET_ATTR(MAXIMUM_TEXTURE1D_LINEAR_WIDTH);
2635                 GET_ATTR(MAXIMUM_TEXTURE2D_LINEAR_WIDTH);
2636                 GET_ATTR(MAXIMUM_TEXTURE2D_LINEAR_HEIGHT);
2637                 GET_ATTR(MAXIMUM_TEXTURE2D_LINEAR_PITCH);
2638                 GET_ATTR(MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH);
2639                 GET_ATTR(MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT);
2640                 GET_ATTR(COMPUTE_CAPABILITY_MAJOR);
2641                 GET_ATTR(COMPUTE_CAPABILITY_MINOR);
2642                 GET_ATTR(MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH);
2643                 GET_ATTR(STREAM_PRIORITIES_SUPPORTED);
2644                 GET_ATTR(GLOBAL_L1_CACHE_SUPPORTED);
2645                 GET_ATTR(LOCAL_L1_CACHE_SUPPORTED);
2646                 GET_ATTR(MAX_SHARED_MEMORY_PER_MULTIPROCESSOR);
2647                 GET_ATTR(MAX_REGISTERS_PER_MULTIPROCESSOR);
2648                 GET_ATTR(MANAGED_MEMORY);
2649                 GET_ATTR(MULTI_GPU_BOARD);
2650                 GET_ATTR(MULTI_GPU_BOARD_GROUP_ID);
2651 #undef GET_ATTR
2652                 capabilities += "\n";
2653         }
2654
2655         return capabilities;
2656 }
2657
2658 CCL_NAMESPACE_END