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 pass_stride = task->buffer.pass_stride;
1298                 int num_shifts = (2*r+1)*(2*r+1);
1299                 int channel_offset = 0;
1300
1301                 if(have_error())
1302                         return false;
1303
1304                 CUdeviceptr difference     = cuda_device_ptr(task->buffer.temporary_mem.device_pointer);
1305                 CUdeviceptr blurDifference = difference + sizeof(float)*pass_stride*num_shifts;
1306                 CUdeviceptr weightAccum = difference + 2*sizeof(float)*pass_stride*num_shifts;
1307
1308                 cuda_assert(cuMemsetD8(weightAccum, 0, sizeof(float)*pass_stride));
1309                 cuda_assert(cuMemsetD8(out_ptr, 0, sizeof(float)*pass_stride));
1310
1311                 {
1312                         CUfunction cuNLMCalcDifference, cuNLMBlur, cuNLMCalcWeight, cuNLMUpdateOutput;
1313                         cuda_assert(cuModuleGetFunction(&cuNLMCalcDifference, cuFilterModule, "kernel_cuda_filter_nlm_calc_difference"));
1314                         cuda_assert(cuModuleGetFunction(&cuNLMBlur,           cuFilterModule, "kernel_cuda_filter_nlm_blur"));
1315                         cuda_assert(cuModuleGetFunction(&cuNLMCalcWeight,     cuFilterModule, "kernel_cuda_filter_nlm_calc_weight"));
1316                         cuda_assert(cuModuleGetFunction(&cuNLMUpdateOutput,   cuFilterModule, "kernel_cuda_filter_nlm_update_output"));
1317
1318                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcDifference, CU_FUNC_CACHE_PREFER_L1));
1319                         cuda_assert(cuFuncSetCacheConfig(cuNLMBlur,           CU_FUNC_CACHE_PREFER_L1));
1320                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcWeight,     CU_FUNC_CACHE_PREFER_L1));
1321                         cuda_assert(cuFuncSetCacheConfig(cuNLMUpdateOutput,   CU_FUNC_CACHE_PREFER_L1));
1322
1323                         CUDA_GET_BLOCKSIZE_1D(cuNLMCalcDifference, w*h, num_shifts);
1324
1325                         void *calc_difference_args[] = {&guide_ptr, &variance_ptr, &difference, &w, &h, &stride, &pass_stride, &r, &channel_offset, &a, &k_2};
1326                         void *blur_args[]            = {&difference, &blurDifference, &w, &h, &stride, &pass_stride, &r, &f};
1327                         void *calc_weight_args[]     = {&blurDifference, &difference, &w, &h, &stride, &pass_stride, &r, &f};
1328                         void *update_output_args[]   = {&blurDifference, &image_ptr, &out_ptr, &weightAccum, &w, &h, &stride, &pass_stride, &r, &f};
1329
1330                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcDifference, calc_difference_args);
1331                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1332                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcWeight, calc_weight_args);
1333                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1334                         CUDA_LAUNCH_KERNEL_1D(cuNLMUpdateOutput, update_output_args);
1335                 }
1336
1337                 {
1338                         CUfunction cuNLMNormalize;
1339                         cuda_assert(cuModuleGetFunction(&cuNLMNormalize, cuFilterModule, "kernel_cuda_filter_nlm_normalize"));
1340                         cuda_assert(cuFuncSetCacheConfig(cuNLMNormalize, CU_FUNC_CACHE_PREFER_L1));
1341                         void *normalize_args[] = {&out_ptr, &weightAccum, &w, &h, &stride};
1342                         CUDA_GET_BLOCKSIZE(cuNLMNormalize, w, h);
1343                         CUDA_LAUNCH_KERNEL(cuNLMNormalize, normalize_args);
1344                         cuda_assert(cuCtxSynchronize());
1345                 }
1346
1347                 return !have_error();
1348         }
1349
1350         bool denoising_construct_transform(DenoisingTask *task)
1351         {
1352                 if(have_error())
1353                         return false;
1354
1355                 CUDAContextScope scope(this);
1356
1357                 CUfunction cuFilterConstructTransform;
1358                 cuda_assert(cuModuleGetFunction(&cuFilterConstructTransform, cuFilterModule, "kernel_cuda_filter_construct_transform"));
1359                 cuda_assert(cuFuncSetCacheConfig(cuFilterConstructTransform, CU_FUNC_CACHE_PREFER_SHARED));
1360                 CUDA_GET_BLOCKSIZE(cuFilterConstructTransform,
1361                                    task->storage.w,
1362                                    task->storage.h);
1363
1364                 void *args[] = {&task->buffer.mem.device_pointer,
1365                                 &task->storage.transform.device_pointer,
1366                                 &task->storage.rank.device_pointer,
1367                                 &task->filter_area,
1368                                 &task->rect,
1369                                 &task->radius,
1370                                 &task->pca_threshold,
1371                                 &task->buffer.pass_stride};
1372                 CUDA_LAUNCH_KERNEL(cuFilterConstructTransform, args);
1373                 cuda_assert(cuCtxSynchronize());
1374
1375                 return !have_error();
1376         }
1377
1378         bool denoising_reconstruct(device_ptr color_ptr,
1379                                    device_ptr color_variance_ptr,
1380                                    device_ptr output_ptr,
1381                                    DenoisingTask *task)
1382         {
1383                 if(have_error())
1384                         return false;
1385
1386                 CUDAContextScope scope(this);
1387
1388                 mem_zero(task->storage.XtWX);
1389                 mem_zero(task->storage.XtWY);
1390
1391                 int r = task->radius;
1392                 int f = 4;
1393                 float a = 1.0f;
1394                 float k_2 = task->nlm_k_2;
1395
1396                 int w = task->reconstruction_state.source_w;
1397                 int h = task->reconstruction_state.source_h;
1398                 int stride = task->buffer.stride;
1399
1400                 int pass_stride = task->buffer.pass_stride;
1401                 int num_shifts = (2*r+1)*(2*r+1);
1402
1403                 if(have_error())
1404                         return false;
1405
1406                 CUdeviceptr difference     = cuda_device_ptr(task->buffer.temporary_mem.device_pointer);
1407                 CUdeviceptr blurDifference = difference + sizeof(float)*pass_stride*num_shifts;
1408
1409                 {
1410                         CUfunction cuNLMCalcDifference, cuNLMBlur, cuNLMCalcWeight, cuNLMConstructGramian;
1411                         cuda_assert(cuModuleGetFunction(&cuNLMCalcDifference,   cuFilterModule, "kernel_cuda_filter_nlm_calc_difference"));
1412                         cuda_assert(cuModuleGetFunction(&cuNLMBlur,             cuFilterModule, "kernel_cuda_filter_nlm_blur"));
1413                         cuda_assert(cuModuleGetFunction(&cuNLMCalcWeight,       cuFilterModule, "kernel_cuda_filter_nlm_calc_weight"));
1414                         cuda_assert(cuModuleGetFunction(&cuNLMConstructGramian, cuFilterModule, "kernel_cuda_filter_nlm_construct_gramian"));
1415
1416                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcDifference,   CU_FUNC_CACHE_PREFER_L1));
1417                         cuda_assert(cuFuncSetCacheConfig(cuNLMBlur,             CU_FUNC_CACHE_PREFER_L1));
1418                         cuda_assert(cuFuncSetCacheConfig(cuNLMCalcWeight,       CU_FUNC_CACHE_PREFER_L1));
1419                         cuda_assert(cuFuncSetCacheConfig(cuNLMConstructGramian, CU_FUNC_CACHE_PREFER_SHARED));
1420
1421                         CUDA_GET_BLOCKSIZE_1D(cuNLMCalcDifference,
1422                                              task->reconstruction_state.source_w * task->reconstruction_state.source_h,
1423                                              num_shifts);
1424
1425                         void *calc_difference_args[] = {&color_ptr, &color_variance_ptr, &difference, &w, &h, &stride, &pass_stride, &r, &pass_stride, &a, &k_2};
1426                         void *blur_args[]            = {&difference, &blurDifference, &w, &h, &stride, &pass_stride, &r, &f};
1427                         void *calc_weight_args[]     = {&blurDifference, &difference, &w, &h, &stride, &pass_stride, &r, &f};
1428                         void *construct_gramian_args[] = {&blurDifference,
1429                                                           &task->buffer.mem.device_pointer,
1430                                                           &task->storage.transform.device_pointer,
1431                                                           &task->storage.rank.device_pointer,
1432                                                           &task->storage.XtWX.device_pointer,
1433                                                           &task->storage.XtWY.device_pointer,
1434                                                           &task->reconstruction_state.filter_window,
1435                                                           &w, &h, &stride,
1436                                                           &pass_stride, &r,
1437                                                           &f};
1438
1439                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcDifference, calc_difference_args);
1440                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1441                         CUDA_LAUNCH_KERNEL_1D(cuNLMCalcWeight, calc_weight_args);
1442                         CUDA_LAUNCH_KERNEL_1D(cuNLMBlur, blur_args);
1443                         CUDA_LAUNCH_KERNEL_1D(cuNLMConstructGramian, construct_gramian_args);
1444                 }
1445
1446                 {
1447                         CUfunction cuFinalize;
1448                         cuda_assert(cuModuleGetFunction(&cuFinalize, cuFilterModule, "kernel_cuda_filter_finalize"));
1449                         cuda_assert(cuFuncSetCacheConfig(cuFinalize, CU_FUNC_CACHE_PREFER_L1));
1450                         void *finalize_args[] = {&output_ptr,
1451                                                          &task->storage.rank.device_pointer,
1452                                                          &task->storage.XtWX.device_pointer,
1453                                                          &task->storage.XtWY.device_pointer,
1454                                                          &task->filter_area,
1455                                                          &task->reconstruction_state.buffer_params.x,
1456                                                          &task->render_buffer.samples};
1457                         CUDA_GET_BLOCKSIZE(cuFinalize,
1458                                            task->reconstruction_state.source_w,
1459                                            task->reconstruction_state.source_h);
1460                         CUDA_LAUNCH_KERNEL(cuFinalize, finalize_args);
1461                 }
1462
1463                 cuda_assert(cuCtxSynchronize());
1464
1465                 return !have_error();
1466         }
1467
1468         bool denoising_combine_halves(device_ptr a_ptr, device_ptr b_ptr,
1469                                       device_ptr mean_ptr, device_ptr variance_ptr,
1470                                       int r, int4 rect, DenoisingTask *task)
1471         {
1472                 if(have_error())
1473                         return false;
1474
1475                 CUDAContextScope scope(this);
1476
1477                 CUfunction cuFilterCombineHalves;
1478                 cuda_assert(cuModuleGetFunction(&cuFilterCombineHalves, cuFilterModule, "kernel_cuda_filter_combine_halves"));
1479                 cuda_assert(cuFuncSetCacheConfig(cuFilterCombineHalves, CU_FUNC_CACHE_PREFER_L1));
1480                 CUDA_GET_BLOCKSIZE(cuFilterCombineHalves,
1481                                    task->rect.z-task->rect.x,
1482                                    task->rect.w-task->rect.y);
1483
1484                 void *args[] = {&mean_ptr,
1485                                 &variance_ptr,
1486                                 &a_ptr,
1487                                 &b_ptr,
1488                                 &rect,
1489                                 &r};
1490                 CUDA_LAUNCH_KERNEL(cuFilterCombineHalves, args);
1491                 cuda_assert(cuCtxSynchronize());
1492
1493                 return !have_error();
1494         }
1495
1496         bool denoising_divide_shadow(device_ptr a_ptr, device_ptr b_ptr,
1497                                      device_ptr sample_variance_ptr, device_ptr sv_variance_ptr,
1498                                      device_ptr buffer_variance_ptr, DenoisingTask *task)
1499         {
1500                 if(have_error())
1501                         return false;
1502
1503                 CUDAContextScope scope(this);
1504
1505                 CUfunction cuFilterDivideShadow;
1506                 cuda_assert(cuModuleGetFunction(&cuFilterDivideShadow, cuFilterModule, "kernel_cuda_filter_divide_shadow"));
1507                 cuda_assert(cuFuncSetCacheConfig(cuFilterDivideShadow, CU_FUNC_CACHE_PREFER_L1));
1508                 CUDA_GET_BLOCKSIZE(cuFilterDivideShadow,
1509                                    task->rect.z-task->rect.x,
1510                                    task->rect.w-task->rect.y);
1511
1512                 void *args[] = {&task->render_buffer.samples,
1513                                 &task->tile_info_mem.device_pointer,
1514                                 &a_ptr,
1515                                 &b_ptr,
1516                                 &sample_variance_ptr,
1517                                 &sv_variance_ptr,
1518                                 &buffer_variance_ptr,
1519                                 &task->rect,
1520                                 &task->render_buffer.pass_stride,
1521                                 &task->render_buffer.offset};
1522                 CUDA_LAUNCH_KERNEL(cuFilterDivideShadow, args);
1523                 cuda_assert(cuCtxSynchronize());
1524
1525                 return !have_error();
1526         }
1527
1528         bool denoising_get_feature(int mean_offset,
1529                                    int variance_offset,
1530                                    device_ptr mean_ptr,
1531                                    device_ptr variance_ptr,
1532                                    DenoisingTask *task)
1533         {
1534                 if(have_error())
1535                         return false;
1536
1537                 CUDAContextScope scope(this);
1538
1539                 CUfunction cuFilterGetFeature;
1540                 cuda_assert(cuModuleGetFunction(&cuFilterGetFeature, cuFilterModule, "kernel_cuda_filter_get_feature"));
1541                 cuda_assert(cuFuncSetCacheConfig(cuFilterGetFeature, CU_FUNC_CACHE_PREFER_L1));
1542                 CUDA_GET_BLOCKSIZE(cuFilterGetFeature,
1543                                    task->rect.z-task->rect.x,
1544                                    task->rect.w-task->rect.y);
1545
1546                 void *args[] = {&task->render_buffer.samples,
1547                                 &task->tile_info_mem.device_pointer,
1548                                 &mean_offset,
1549                                 &variance_offset,
1550                                 &mean_ptr,
1551                                 &variance_ptr,
1552                                 &task->rect,
1553                                 &task->render_buffer.pass_stride,
1554                                 &task->render_buffer.offset};
1555                 CUDA_LAUNCH_KERNEL(cuFilterGetFeature, args);
1556                 cuda_assert(cuCtxSynchronize());
1557
1558                 return !have_error();
1559         }
1560
1561         bool denoising_detect_outliers(device_ptr image_ptr,
1562                                        device_ptr variance_ptr,
1563                                        device_ptr depth_ptr,
1564                                        device_ptr output_ptr,
1565                                        DenoisingTask *task)
1566         {
1567                 if(have_error())
1568                         return false;
1569
1570                 CUDAContextScope scope(this);
1571
1572                 CUfunction cuFilterDetectOutliers;
1573                 cuda_assert(cuModuleGetFunction(&cuFilterDetectOutliers, cuFilterModule, "kernel_cuda_filter_detect_outliers"));
1574                 cuda_assert(cuFuncSetCacheConfig(cuFilterDetectOutliers, CU_FUNC_CACHE_PREFER_L1));
1575                 CUDA_GET_BLOCKSIZE(cuFilterDetectOutliers,
1576                                    task->rect.z-task->rect.x,
1577                                    task->rect.w-task->rect.y);
1578
1579                 void *args[] = {&image_ptr,
1580                                 &variance_ptr,
1581                                 &depth_ptr,
1582                                 &output_ptr,
1583                                 &task->rect,
1584                                 &task->buffer.pass_stride};
1585
1586                 CUDA_LAUNCH_KERNEL(cuFilterDetectOutliers, args);
1587                 cuda_assert(cuCtxSynchronize());
1588
1589                 return !have_error();
1590         }
1591
1592         void denoise(RenderTile &rtile, DenoisingTask& denoising)
1593         {
1594                 denoising.functions.construct_transform = function_bind(&CUDADevice::denoising_construct_transform, this, &denoising);
1595                 denoising.functions.reconstruct = function_bind(&CUDADevice::denoising_reconstruct, this, _1, _2, _3, &denoising);
1596                 denoising.functions.divide_shadow = function_bind(&CUDADevice::denoising_divide_shadow, this, _1, _2, _3, _4, _5, &denoising);
1597                 denoising.functions.non_local_means = function_bind(&CUDADevice::denoising_non_local_means, this, _1, _2, _3, _4, &denoising);
1598                 denoising.functions.combine_halves = function_bind(&CUDADevice::denoising_combine_halves, this, _1, _2, _3, _4, _5, _6, &denoising);
1599                 denoising.functions.get_feature = function_bind(&CUDADevice::denoising_get_feature, this, _1, _2, _3, _4, &denoising);
1600                 denoising.functions.detect_outliers = function_bind(&CUDADevice::denoising_detect_outliers, this, _1, _2, _3, _4, &denoising);
1601
1602                 denoising.filter_area = make_int4(rtile.x, rtile.y, rtile.w, rtile.h);
1603                 denoising.render_buffer.samples = rtile.sample;
1604                 denoising.buffer.gpu_temporary_mem = true;
1605
1606                 denoising.run_denoising(&rtile);
1607         }
1608
1609         void path_trace(DeviceTask& task, RenderTile& rtile, device_vector<WorkTile>& work_tiles)
1610         {
1611                 scoped_timer timer(&rtile.buffers->render_time);
1612
1613                 if(have_error())
1614                         return;
1615
1616                 CUDAContextScope scope(this);
1617                 CUfunction cuPathTrace;
1618
1619                 /* Get kernel function. */
1620                 if(task.integrator_branched) {
1621                         cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_branched_path_trace"));
1622                 }
1623                 else {
1624                         cuda_assert(cuModuleGetFunction(&cuPathTrace, cuModule, "kernel_cuda_path_trace"));
1625                 }
1626
1627                 if(have_error()) {
1628                         return;
1629                 }
1630
1631                 cuda_assert(cuFuncSetCacheConfig(cuPathTrace, CU_FUNC_CACHE_PREFER_L1));
1632
1633                 /* Allocate work tile. */
1634                 work_tiles.alloc(1);
1635
1636                 WorkTile *wtile = work_tiles.data();
1637                 wtile->x = rtile.x;
1638                 wtile->y = rtile.y;
1639                 wtile->w = rtile.w;
1640                 wtile->h = rtile.h;
1641                 wtile->offset = rtile.offset;
1642                 wtile->stride = rtile.stride;
1643                 wtile->buffer = (float*)cuda_device_ptr(rtile.buffer);
1644
1645                 /* Prepare work size. More step samples render faster, but for now we
1646                  * remain conservative for GPUs connected to a display to avoid driver
1647                  * timeouts and display freezing. */
1648                 int min_blocks, num_threads_per_block;
1649                 cuda_assert(cuOccupancyMaxPotentialBlockSize(&min_blocks, &num_threads_per_block, cuPathTrace, NULL, 0, 0));
1650                 if(!info.display_device) {
1651                         min_blocks *= 8;
1652                 }
1653
1654                 uint step_samples = divide_up(min_blocks * num_threads_per_block, wtile->w * wtile->h);
1655
1656                 /* Render all samples. */
1657                 int start_sample = rtile.start_sample;
1658                 int end_sample = rtile.start_sample + rtile.num_samples;
1659
1660                 for(int sample = start_sample; sample < end_sample; sample += step_samples) {
1661                         /* Setup and copy work tile to device. */
1662                         wtile->start_sample = sample;
1663                         wtile->num_samples = min(step_samples, end_sample - sample);
1664                         work_tiles.copy_to_device();
1665
1666                         CUdeviceptr d_work_tiles = cuda_device_ptr(work_tiles.device_pointer);
1667                         uint total_work_size = wtile->w * wtile->h * wtile->num_samples;
1668                         uint num_blocks = divide_up(total_work_size, num_threads_per_block);
1669
1670                         /* Launch kernel. */
1671                         void *args[] = {&d_work_tiles,
1672                                         &total_work_size};
1673
1674                         cuda_assert(cuLaunchKernel(cuPathTrace,
1675                                                    num_blocks, 1, 1,
1676                                                    num_threads_per_block, 1, 1,
1677                                                    0, 0, args, 0));
1678
1679                         cuda_assert(cuCtxSynchronize());
1680
1681                         /* Update progress. */
1682                         rtile.sample = sample + wtile->num_samples;
1683                         task.update_progress(&rtile, rtile.w*rtile.h*wtile->num_samples);
1684
1685                         if(task.get_cancel()) {
1686                                 if(task.need_finish_queue == false)
1687                                         break;
1688                         }
1689                 }
1690         }
1691
1692         void film_convert(DeviceTask& task, device_ptr buffer, device_ptr rgba_byte, device_ptr rgba_half)
1693         {
1694                 if(have_error())
1695                         return;
1696
1697                 CUDAContextScope scope(this);
1698
1699                 CUfunction cuFilmConvert;
1700                 CUdeviceptr d_rgba = map_pixels((rgba_byte)? rgba_byte: rgba_half);
1701                 CUdeviceptr d_buffer = cuda_device_ptr(buffer);
1702
1703                 /* get kernel function */
1704                 if(rgba_half) {
1705                         cuda_assert(cuModuleGetFunction(&cuFilmConvert, cuModule, "kernel_cuda_convert_to_half_float"));
1706                 }
1707                 else {
1708                         cuda_assert(cuModuleGetFunction(&cuFilmConvert, cuModule, "kernel_cuda_convert_to_byte"));
1709                 }
1710
1711
1712                 float sample_scale = 1.0f/(task.sample + 1);
1713
1714                 /* pass in parameters */
1715                 void *args[] = {&d_rgba,
1716                                 &d_buffer,
1717                                 &sample_scale,
1718                                 &task.x,
1719                                 &task.y,
1720                                 &task.w,
1721                                 &task.h,
1722                                 &task.offset,
1723                                 &task.stride};
1724
1725                 /* launch kernel */
1726                 int threads_per_block;
1727                 cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, cuFilmConvert));
1728
1729                 int xthreads = (int)sqrt(threads_per_block);
1730                 int ythreads = (int)sqrt(threads_per_block);
1731                 int xblocks = (task.w + xthreads - 1)/xthreads;
1732                 int yblocks = (task.h + ythreads - 1)/ythreads;
1733
1734                 cuda_assert(cuFuncSetCacheConfig(cuFilmConvert, CU_FUNC_CACHE_PREFER_L1));
1735
1736                 cuda_assert(cuLaunchKernel(cuFilmConvert,
1737                                            xblocks , yblocks, 1, /* blocks */
1738                                            xthreads, ythreads, 1, /* threads */
1739                                            0, 0, args, 0));
1740
1741                 unmap_pixels((rgba_byte)? rgba_byte: rgba_half);
1742
1743                 cuda_assert(cuCtxSynchronize());
1744         }
1745
1746         void shader(DeviceTask& task)
1747         {
1748                 if(have_error())
1749                         return;
1750
1751                 CUDAContextScope scope(this);
1752
1753                 CUfunction cuShader;
1754                 CUdeviceptr d_input = cuda_device_ptr(task.shader_input);
1755                 CUdeviceptr d_output = cuda_device_ptr(task.shader_output);
1756
1757                 /* get kernel function */
1758                 if(task.shader_eval_type >= SHADER_EVAL_BAKE) {
1759                         cuda_assert(cuModuleGetFunction(&cuShader, cuModule, "kernel_cuda_bake"));
1760                 }
1761                 else if(task.shader_eval_type == SHADER_EVAL_DISPLACE) {
1762                         cuda_assert(cuModuleGetFunction(&cuShader, cuModule, "kernel_cuda_displace"));
1763                 }
1764                 else {
1765                         cuda_assert(cuModuleGetFunction(&cuShader, cuModule, "kernel_cuda_background"));
1766                 }
1767
1768                 /* do tasks in smaller chunks, so we can cancel it */
1769                 const int shader_chunk_size = 65536;
1770                 const int start = task.shader_x;
1771                 const int end = task.shader_x + task.shader_w;
1772                 int offset = task.offset;
1773
1774                 bool canceled = false;
1775                 for(int sample = 0; sample < task.num_samples && !canceled; sample++) {
1776                         for(int shader_x = start; shader_x < end; shader_x += shader_chunk_size) {
1777                                 int shader_w = min(shader_chunk_size, end - shader_x);
1778
1779                                 /* pass in parameters */
1780                                 void *args[8];
1781                                 int arg = 0;
1782                                 args[arg++] = &d_input;
1783                                 args[arg++] = &d_output;
1784                                 args[arg++] = &task.shader_eval_type;
1785                                 if(task.shader_eval_type >= SHADER_EVAL_BAKE) {
1786                                         args[arg++] = &task.shader_filter;
1787                                 }
1788                                 args[arg++] = &shader_x;
1789                                 args[arg++] = &shader_w;
1790                                 args[arg++] = &offset;
1791                                 args[arg++] = &sample;
1792
1793                                 /* launch kernel */
1794                                 int threads_per_block;
1795                                 cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, cuShader));
1796
1797                                 int xblocks = (shader_w + threads_per_block - 1)/threads_per_block;
1798
1799                                 cuda_assert(cuFuncSetCacheConfig(cuShader, CU_FUNC_CACHE_PREFER_L1));
1800                                 cuda_assert(cuLaunchKernel(cuShader,
1801                                                            xblocks , 1, 1, /* blocks */
1802                                                            threads_per_block, 1, 1, /* threads */
1803                                                            0, 0, args, 0));
1804
1805                                 cuda_assert(cuCtxSynchronize());
1806
1807                                 if(task.get_cancel()) {
1808                                         canceled = true;
1809                                         break;
1810                                 }
1811                         }
1812
1813                         task.update_progress(NULL);
1814                 }
1815         }
1816
1817         CUdeviceptr map_pixels(device_ptr mem)
1818         {
1819                 if(!background) {
1820                         PixelMem pmem = pixel_mem_map[mem];
1821                         CUdeviceptr buffer;
1822
1823                         size_t bytes;
1824                         cuda_assert(cuGraphicsMapResources(1, &pmem.cuPBOresource, 0));
1825                         cuda_assert(cuGraphicsResourceGetMappedPointer(&buffer, &bytes, pmem.cuPBOresource));
1826
1827                         return buffer;
1828                 }
1829
1830                 return cuda_device_ptr(mem);
1831         }
1832
1833         void unmap_pixels(device_ptr mem)
1834         {
1835                 if(!background) {
1836                         PixelMem pmem = pixel_mem_map[mem];
1837
1838                         cuda_assert(cuGraphicsUnmapResources(1, &pmem.cuPBOresource, 0));
1839                 }
1840         }
1841
1842         void pixels_alloc(device_memory& mem)
1843         {
1844                 PixelMem pmem;
1845
1846                 pmem.w = mem.data_width;
1847                 pmem.h = mem.data_height;
1848
1849                 CUDAContextScope scope(this);
1850
1851                 glGenBuffers(1, &pmem.cuPBO);
1852                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO);
1853                 if(mem.data_type == TYPE_HALF)
1854                         glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(GLhalf)*4, NULL, GL_DYNAMIC_DRAW);
1855                 else
1856                         glBufferData(GL_PIXEL_UNPACK_BUFFER, pmem.w*pmem.h*sizeof(uint8_t)*4, NULL, GL_DYNAMIC_DRAW);
1857
1858                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
1859
1860                 glGenTextures(1, &pmem.cuTexId);
1861                 glBindTexture(GL_TEXTURE_2D, pmem.cuTexId);
1862                 if(mem.data_type == TYPE_HALF)
1863                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA16F, pmem.w, pmem.h, 0, GL_RGBA, GL_HALF_FLOAT, NULL);
1864                 else
1865                         glTexImage2D(GL_TEXTURE_2D, 0, GL_RGBA8, pmem.w, pmem.h, 0, GL_RGBA, GL_UNSIGNED_BYTE, NULL);
1866                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
1867                 glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
1868                 glBindTexture(GL_TEXTURE_2D, 0);
1869
1870                 CUresult result = cuGraphicsGLRegisterBuffer(&pmem.cuPBOresource, pmem.cuPBO, CU_GRAPHICS_MAP_RESOURCE_FLAGS_NONE);
1871
1872                 if(result == CUDA_SUCCESS) {
1873                         mem.device_pointer = pmem.cuTexId;
1874                         pixel_mem_map[mem.device_pointer] = pmem;
1875
1876                         mem.device_size = mem.memory_size();
1877                         stats.mem_alloc(mem.device_size);
1878
1879                         return;
1880                 }
1881                 else {
1882                         /* failed to register buffer, fallback to no interop */
1883                         glDeleteBuffers(1, &pmem.cuPBO);
1884                         glDeleteTextures(1, &pmem.cuTexId);
1885
1886                         background = true;
1887                 }
1888         }
1889
1890         void pixels_copy_from(device_memory& mem, int y, int w, int h)
1891         {
1892                 PixelMem pmem = pixel_mem_map[mem.device_pointer];
1893
1894                 CUDAContextScope scope(this);
1895
1896                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO);
1897                 uchar *pixels = (uchar*)glMapBuffer(GL_PIXEL_UNPACK_BUFFER, GL_READ_ONLY);
1898                 size_t offset = sizeof(uchar)*4*y*w;
1899                 memcpy((uchar*)mem.host_pointer + offset, pixels + offset, sizeof(uchar)*4*w*h);
1900                 glUnmapBuffer(GL_PIXEL_UNPACK_BUFFER);
1901                 glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
1902         }
1903
1904         void pixels_free(device_memory& mem)
1905         {
1906                 if(mem.device_pointer) {
1907                         PixelMem pmem = pixel_mem_map[mem.device_pointer];
1908
1909                         CUDAContextScope scope(this);
1910
1911                         cuda_assert(cuGraphicsUnregisterResource(pmem.cuPBOresource));
1912                         glDeleteBuffers(1, &pmem.cuPBO);
1913                         glDeleteTextures(1, &pmem.cuTexId);
1914
1915                         pixel_mem_map.erase(pixel_mem_map.find(mem.device_pointer));
1916                         mem.device_pointer = 0;
1917
1918                         stats.mem_free(mem.device_size);
1919                         mem.device_size = 0;
1920                 }
1921         }
1922
1923         void draw_pixels(
1924             device_memory& mem, int y,
1925             int w, int h, int width, int height,
1926             int dx, int dy, int dw, int dh, bool transparent,
1927                 const DeviceDrawParams &draw_params)
1928         {
1929                 assert(mem.type == MEM_PIXELS);
1930
1931                 if(!background) {
1932                         const bool use_fallback_shader = (draw_params.bind_display_space_shader_cb == NULL);
1933                         PixelMem pmem = pixel_mem_map[mem.device_pointer];
1934                         float *vpointer;
1935
1936                         CUDAContextScope scope(this);
1937
1938                         /* for multi devices, this assumes the inefficient method that we allocate
1939                          * all pixels on the device even though we only render to a subset */
1940                         size_t offset = 4*y*w;
1941
1942                         if(mem.data_type == TYPE_HALF)
1943                                 offset *= sizeof(GLhalf);
1944                         else
1945                                 offset *= sizeof(uint8_t);
1946
1947                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, pmem.cuPBO);
1948                         glBindTexture(GL_TEXTURE_2D, pmem.cuTexId);
1949                         if(mem.data_type == TYPE_HALF) {
1950                                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, w, h, GL_RGBA, GL_HALF_FLOAT, (void*)offset);
1951                         }
1952                         else {
1953                                 glTexSubImage2D(GL_TEXTURE_2D, 0, 0, 0, w, h, GL_RGBA, GL_UNSIGNED_BYTE, (void*)offset);
1954                         }
1955                         glBindBuffer(GL_PIXEL_UNPACK_BUFFER, 0);
1956
1957                         if(transparent) {
1958                                 glEnable(GL_BLEND);
1959                                 glBlendFunc(GL_ONE, GL_ONE_MINUS_SRC_ALPHA);
1960                         }
1961
1962                         GLint shader_program;
1963                         if(use_fallback_shader) {
1964                                 if(!bind_fallback_display_space_shader(dw, dh)) {
1965                                         return;
1966                                 }
1967                                 shader_program = fallback_shader_program;
1968                         }
1969                         else {
1970                                 draw_params.bind_display_space_shader_cb();
1971                                 glGetIntegerv(GL_CURRENT_PROGRAM, &shader_program);
1972                         }
1973
1974                         if(!vertex_buffer) {
1975                                 glGenBuffers(1, &vertex_buffer);
1976                         }
1977
1978                         glBindBuffer(GL_ARRAY_BUFFER, vertex_buffer);
1979                         /* invalidate old contents - avoids stalling if buffer is still waiting in queue to be rendered */
1980                         glBufferData(GL_ARRAY_BUFFER, 16 * sizeof(float), NULL, GL_STREAM_DRAW);
1981
1982                         vpointer = (float *)glMapBuffer(GL_ARRAY_BUFFER, GL_WRITE_ONLY);
1983
1984                         if(vpointer) {
1985                                 /* texture coordinate - vertex pair */
1986                                 vpointer[0] = 0.0f;
1987                                 vpointer[1] = 0.0f;
1988                                 vpointer[2] = dx;
1989                                 vpointer[3] = dy;
1990
1991                                 vpointer[4] = (float)w/(float)pmem.w;
1992                                 vpointer[5] = 0.0f;
1993                                 vpointer[6] = (float)width + dx;
1994                                 vpointer[7] = dy;
1995
1996                                 vpointer[8] = (float)w/(float)pmem.w;
1997                                 vpointer[9] = (float)h/(float)pmem.h;
1998                                 vpointer[10] = (float)width + dx;
1999                                 vpointer[11] = (float)height + dy;
2000
2001                                 vpointer[12] = 0.0f;
2002                                 vpointer[13] = (float)h/(float)pmem.h;
2003                                 vpointer[14] = dx;
2004                                 vpointer[15] = (float)height + dy;
2005
2006                                 glUnmapBuffer(GL_ARRAY_BUFFER);
2007                         }
2008
2009                         GLuint vertex_array_object;
2010                         GLuint position_attribute, texcoord_attribute;
2011
2012                         glGenVertexArrays(1, &vertex_array_object);
2013                         glBindVertexArray(vertex_array_object);
2014
2015                         texcoord_attribute = glGetAttribLocation(shader_program, "texCoord");
2016                         position_attribute = glGetAttribLocation(shader_program, "pos");
2017
2018                         glEnableVertexAttribArray(texcoord_attribute);
2019                         glEnableVertexAttribArray(position_attribute);
2020
2021                         glVertexAttribPointer(texcoord_attribute, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (const GLvoid *)0);
2022                         glVertexAttribPointer(position_attribute, 2, GL_FLOAT, GL_FALSE, 4 * sizeof(float), (const GLvoid *)(sizeof(float) * 2));
2023
2024                         glDrawArrays(GL_TRIANGLE_FAN, 0, 4);
2025
2026                         if(use_fallback_shader) {
2027                                 glUseProgram(0);
2028                         }
2029                         else {
2030                                 draw_params.unbind_display_space_shader_cb();
2031                         }
2032
2033                         if(transparent) {
2034                                 glDisable(GL_BLEND);
2035                         }
2036
2037                         glBindTexture(GL_TEXTURE_2D, 0);
2038
2039                         return;
2040                 }
2041
2042                 Device::draw_pixels(mem, y, w, h, width, height, dx, dy, dw, dh, transparent, draw_params);
2043         }
2044
2045         void thread_run(DeviceTask *task)
2046         {
2047                 CUDAContextScope scope(this);
2048
2049                 if(task->type == DeviceTask::RENDER) {
2050                         DeviceRequestedFeatures requested_features;
2051                         if(use_split_kernel()) {
2052                                 if(split_kernel == NULL) {
2053                                         split_kernel = new CUDASplitKernel(this);
2054                                         split_kernel->load_kernels(requested_features);
2055                                 }
2056                         }
2057
2058                         device_vector<WorkTile> work_tiles(this, "work_tiles", MEM_READ_ONLY);
2059
2060                         /* keep rendering tiles until done */
2061                         RenderTile tile;
2062                         DenoisingTask denoising(this, *task);
2063
2064                         while(task->acquire_tile(this, tile)) {
2065                                 if(tile.task == RenderTile::PATH_TRACE) {
2066                                         if(use_split_kernel()) {
2067                                                 device_only_memory<uchar> void_buffer(this, "void_buffer");
2068                                                 split_kernel->path_trace(task, tile, void_buffer, void_buffer);
2069                                         }
2070                                         else {
2071                                                 path_trace(*task, tile, work_tiles);
2072                                         }
2073                                 }
2074                                 else if(tile.task == RenderTile::DENOISE) {
2075                                         tile.sample = tile.start_sample + tile.num_samples;
2076
2077                                         denoise(tile, denoising);
2078
2079                                         task->update_progress(&tile, tile.w*tile.h);
2080                                 }
2081
2082                                 task->release_tile(tile);
2083
2084                                 if(task->get_cancel()) {
2085                                         if(task->need_finish_queue == false)
2086                                                 break;
2087                                 }
2088                         }
2089
2090                         work_tiles.free();
2091                 }
2092                 else if(task->type == DeviceTask::SHADER) {
2093                         shader(*task);
2094
2095                         cuda_assert(cuCtxSynchronize());
2096                 }
2097         }
2098
2099         class CUDADeviceTask : public DeviceTask {
2100         public:
2101                 CUDADeviceTask(CUDADevice *device, DeviceTask& task)
2102                 : DeviceTask(task)
2103                 {
2104                         run = function_bind(&CUDADevice::thread_run, device, this);
2105                 }
2106         };
2107
2108         int get_split_task_count(DeviceTask& /*task*/)
2109         {
2110                 return 1;
2111         }
2112
2113         void task_add(DeviceTask& task)
2114         {
2115                 CUDAContextScope scope(this);
2116
2117                 /* Load texture info. */
2118                 load_texture_info();
2119
2120                 /* Synchronize all memory copies before executing task. */
2121                 cuda_assert(cuCtxSynchronize());
2122
2123                 if(task.type == DeviceTask::FILM_CONVERT) {
2124                         /* must be done in main thread due to opengl access */
2125                         film_convert(task, task.buffer, task.rgba_byte, task.rgba_half);
2126                 }
2127                 else {
2128                         task_pool.push(new CUDADeviceTask(this, task));
2129                 }
2130         }
2131
2132         void task_wait()
2133         {
2134                 task_pool.wait();
2135         }
2136
2137         void task_cancel()
2138         {
2139                 task_pool.cancel();
2140         }
2141
2142         friend class CUDASplitKernelFunction;
2143         friend class CUDASplitKernel;
2144         friend class CUDAContextScope;
2145 };
2146
2147 /* redefine the cuda_assert macro so it can be used outside of the CUDADevice class
2148  * now that the definition of that class is complete
2149  */
2150 #undef cuda_assert
2151 #define cuda_assert(stmt) \
2152         { \
2153                 CUresult result = stmt; \
2154                 \
2155                 if(result != CUDA_SUCCESS) { \
2156                         string message = string_printf("CUDA error: %s in %s", cuewErrorString(result), #stmt); \
2157                         if(device->error_msg == "") \
2158                                 device->error_msg = message; \
2159                         fprintf(stderr, "%s\n", message.c_str()); \
2160                         /*cuda_abort();*/ \
2161                         device->cuda_error_documentation(); \
2162                 } \
2163         } (void)0
2164
2165
2166 /* CUDA context scope. */
2167
2168 CUDAContextScope::CUDAContextScope(CUDADevice *device)
2169 : device(device)
2170 {
2171         cuda_assert(cuCtxPushCurrent(device->cuContext));
2172 }
2173
2174 CUDAContextScope::~CUDAContextScope()
2175 {
2176         cuda_assert(cuCtxPopCurrent(NULL));
2177 }
2178
2179 /* split kernel */
2180
2181 class CUDASplitKernelFunction : public SplitKernelFunction{
2182         CUDADevice* device;
2183         CUfunction func;
2184 public:
2185         CUDASplitKernelFunction(CUDADevice *device, CUfunction func) : device(device), func(func) {}
2186
2187         /* enqueue the kernel, returns false if there is an error */
2188         bool enqueue(const KernelDimensions &dim, device_memory &/*kg*/, device_memory &/*data*/)
2189         {
2190                 return enqueue(dim, NULL);
2191         }
2192
2193         /* enqueue the kernel, returns false if there is an error */
2194         bool enqueue(const KernelDimensions &dim, void *args[])
2195         {
2196                 if(device->have_error())
2197                         return false;
2198
2199                 CUDAContextScope scope(device);
2200
2201                 /* we ignore dim.local_size for now, as this is faster */
2202                 int threads_per_block;
2203                 cuda_assert(cuFuncGetAttribute(&threads_per_block, CU_FUNC_ATTRIBUTE_MAX_THREADS_PER_BLOCK, func));
2204
2205                 int xblocks = (dim.global_size[0]*dim.global_size[1] + threads_per_block - 1)/threads_per_block;
2206
2207                 cuda_assert(cuFuncSetCacheConfig(func, CU_FUNC_CACHE_PREFER_L1));
2208
2209                 cuda_assert(cuLaunchKernel(func,
2210                                            xblocks, 1, 1, /* blocks */
2211                                            threads_per_block, 1, 1, /* threads */
2212                                            0, 0, args, 0));
2213
2214                 return !device->have_error();
2215         }
2216 };
2217
2218 CUDASplitKernel::CUDASplitKernel(CUDADevice *device) : DeviceSplitKernel(device), device(device)
2219 {
2220 }
2221
2222 uint64_t CUDASplitKernel::state_buffer_size(device_memory& /*kg*/, device_memory& /*data*/, size_t num_threads)
2223 {
2224         CUDAContextScope scope(device);
2225
2226         device_vector<uint64_t> size_buffer(device, "size_buffer", MEM_READ_WRITE);
2227         size_buffer.alloc(1);
2228         size_buffer.zero_to_device();
2229
2230         uint threads = num_threads;
2231         CUdeviceptr d_size = device->cuda_device_ptr(size_buffer.device_pointer);
2232
2233         struct args_t {
2234                 uint* num_threads;
2235                 CUdeviceptr* size;
2236         };
2237
2238         args_t args = {
2239                 &threads,
2240                 &d_size
2241         };
2242
2243         CUfunction state_buffer_size;
2244         cuda_assert(cuModuleGetFunction(&state_buffer_size, device->cuModule, "kernel_cuda_state_buffer_size"));
2245
2246         cuda_assert(cuLaunchKernel(state_buffer_size,
2247                                    1, 1, 1,
2248                                    1, 1, 1,
2249                                    0, 0, (void**)&args, 0));
2250
2251         size_buffer.copy_from_device(0, 1, 1);
2252         size_t size = size_buffer[0];
2253         size_buffer.free();
2254
2255         return size;
2256 }
2257
2258 bool CUDASplitKernel::enqueue_split_kernel_data_init(const KernelDimensions& dim,
2259                                     RenderTile& rtile,
2260                                     int num_global_elements,
2261                                     device_memory& /*kernel_globals*/,
2262                                     device_memory& /*kernel_data*/,
2263                                     device_memory& split_data,
2264                                     device_memory& ray_state,
2265                                     device_memory& queue_index,
2266                                     device_memory& use_queues_flag,
2267                                     device_memory& work_pool_wgs)
2268 {
2269         CUDAContextScope scope(device);
2270
2271         CUdeviceptr d_split_data = device->cuda_device_ptr(split_data.device_pointer);
2272         CUdeviceptr d_ray_state = device->cuda_device_ptr(ray_state.device_pointer);
2273         CUdeviceptr d_queue_index = device->cuda_device_ptr(queue_index.device_pointer);
2274         CUdeviceptr d_use_queues_flag = device->cuda_device_ptr(use_queues_flag.device_pointer);
2275         CUdeviceptr d_work_pool_wgs = device->cuda_device_ptr(work_pool_wgs.device_pointer);
2276
2277         CUdeviceptr d_buffer = device->cuda_device_ptr(rtile.buffer);
2278
2279         int end_sample = rtile.start_sample + rtile.num_samples;
2280         int queue_size = dim.global_size[0] * dim.global_size[1];
2281
2282         struct args_t {
2283                 CUdeviceptr* split_data_buffer;
2284                 int* num_elements;
2285                 CUdeviceptr* ray_state;
2286                 int* start_sample;
2287                 int* end_sample;
2288                 int* sx;
2289                 int* sy;
2290                 int* sw;
2291                 int* sh;
2292                 int* offset;
2293                 int* stride;
2294                 CUdeviceptr* queue_index;
2295                 int* queuesize;
2296                 CUdeviceptr* use_queues_flag;
2297                 CUdeviceptr* work_pool_wgs;
2298                 int* num_samples;
2299                 CUdeviceptr* buffer;
2300         };
2301
2302         args_t args = {
2303                 &d_split_data,
2304                 &num_global_elements,
2305                 &d_ray_state,
2306                 &rtile.start_sample,
2307                 &end_sample,
2308                 &rtile.x,
2309                 &rtile.y,
2310                 &rtile.w,
2311                 &rtile.h,
2312                 &rtile.offset,
2313                 &rtile.stride,
2314                 &d_queue_index,
2315                 &queue_size,
2316                 &d_use_queues_flag,
2317                 &d_work_pool_wgs,
2318                 &rtile.num_samples,
2319                 &d_buffer
2320         };
2321
2322         CUfunction data_init;
2323         cuda_assert(cuModuleGetFunction(&data_init, device->cuModule, "kernel_cuda_path_trace_data_init"));
2324         if(device->have_error()) {
2325                 return false;
2326         }
2327
2328         CUDASplitKernelFunction(device, data_init).enqueue(dim, (void**)&args);
2329
2330         return !device->have_error();
2331 }
2332
2333 SplitKernelFunction* CUDASplitKernel::get_split_kernel_function(const string& kernel_name,
2334                                                                 const DeviceRequestedFeatures&)
2335 {
2336         CUDAContextScope scope(device);
2337         CUfunction func;
2338
2339         cuda_assert(cuModuleGetFunction(&func, device->cuModule, (string("kernel_cuda_") + kernel_name).data()));
2340         if(device->have_error()) {
2341                 device->cuda_error_message(string_printf("kernel \"kernel_cuda_%s\" not found in module", kernel_name.data()));
2342                 return NULL;
2343         }
2344
2345         return new CUDASplitKernelFunction(device, func);
2346 }
2347
2348 int2 CUDASplitKernel::split_kernel_local_size()
2349 {
2350         return make_int2(32, 1);
2351 }
2352
2353 int2 CUDASplitKernel::split_kernel_global_size(device_memory& kg, device_memory& data, DeviceTask * /*task*/)
2354 {
2355         CUDAContextScope scope(device);
2356         size_t free;
2357         size_t total;
2358
2359         cuda_assert(cuMemGetInfo(&free, &total));
2360
2361         VLOG(1) << "Maximum device allocation size: "
2362                 << string_human_readable_number(free) << " bytes. ("
2363                 << string_human_readable_size(free) << ").";
2364
2365         size_t num_elements = max_elements_for_max_buffer_size(kg, data, free / 2);
2366         size_t side = round_down((int)sqrt(num_elements), 32);
2367         int2 global_size = make_int2(side, round_down(num_elements / side, 16));
2368         VLOG(1) << "Global size: " << global_size << ".";
2369         return global_size;
2370 }
2371
2372 bool device_cuda_init(void)
2373 {
2374 #ifdef WITH_CUDA_DYNLOAD
2375         static bool initialized = false;
2376         static bool result = false;
2377
2378         if(initialized)
2379                 return result;
2380
2381         initialized = true;
2382         int cuew_result = cuewInit(CUEW_INIT_CUDA);
2383         if(cuew_result == CUEW_SUCCESS) {
2384                 VLOG(1) << "CUEW initialization succeeded";
2385                 if(CUDADevice::have_precompiled_kernels()) {
2386                         VLOG(1) << "Found precompiled kernels";
2387                         result = true;
2388                 }
2389 #ifndef _WIN32
2390                 else if(cuewCompilerPath() != NULL) {
2391                         VLOG(1) << "Found CUDA compiler " << cuewCompilerPath();
2392                         result = true;
2393                 }
2394                 else {
2395                         VLOG(1) << "Neither precompiled kernels nor CUDA compiler was found,"
2396                                 << " unable to use CUDA";
2397                 }
2398 #endif
2399         }
2400         else {
2401                 VLOG(1) << "CUEW initialization failed: "
2402                         << ((cuew_result == CUEW_ERROR_ATEXIT_FAILED)
2403                             ? "Error setting up atexit() handler"
2404                             : "Error opening the library");
2405         }
2406
2407         return result;
2408 #else  /* WITH_CUDA_DYNLOAD */
2409         return true;
2410 #endif /* WITH_CUDA_DYNLOAD */
2411 }
2412
2413 Device *device_cuda_create(DeviceInfo& info, Stats &stats, bool background)
2414 {
2415         return new CUDADevice(info, stats, background);
2416 }
2417
2418 static CUresult device_cuda_safe_init()
2419 {
2420 #ifdef _WIN32
2421         __try {
2422                 return cuInit(0);
2423         }
2424         __except(EXCEPTION_EXECUTE_HANDLER) {
2425                 /* Ignore crashes inside the CUDA driver and hope we can
2426                  * survive even with corrupted CUDA installs. */
2427                 fprintf(stderr, "Cycles CUDA: driver crashed, continuing without CUDA.\n");
2428         }
2429
2430         return CUDA_ERROR_NO_DEVICE;
2431 #else
2432         return cuInit(0);
2433 #endif
2434 }
2435
2436 void device_cuda_info(vector<DeviceInfo>& devices)
2437 {
2438         CUresult result = device_cuda_safe_init();
2439         if(result != CUDA_SUCCESS) {
2440                 if(result != CUDA_ERROR_NO_DEVICE)
2441                         fprintf(stderr, "CUDA cuInit: %s\n", cuewErrorString(result));
2442                 return;
2443         }
2444
2445         int count = 0;
2446         result = cuDeviceGetCount(&count);
2447         if(result != CUDA_SUCCESS) {
2448                 fprintf(stderr, "CUDA cuDeviceGetCount: %s\n", cuewErrorString(result));
2449                 return;
2450         }
2451
2452         vector<DeviceInfo> display_devices;
2453
2454         for(int num = 0; num < count; num++) {
2455                 char name[256];
2456
2457                 result = cuDeviceGetName(name, 256, num);
2458                 if(result != CUDA_SUCCESS) {
2459                         fprintf(stderr, "CUDA cuDeviceGetName: %s\n", cuewErrorString(result));
2460                         continue;
2461                 }
2462
2463                 int major;
2464                 cuDeviceGetAttribute(&major, CU_DEVICE_ATTRIBUTE_COMPUTE_CAPABILITY_MAJOR, num);
2465                 if(major < 3) {
2466                         VLOG(1) << "Ignoring device \"" << name
2467                                 << "\", this graphics card is no longer supported.";
2468                         continue;
2469                 }
2470
2471                 DeviceInfo info;
2472
2473                 info.type = DEVICE_CUDA;
2474                 info.description = string(name);
2475                 info.num = num;
2476
2477                 info.advanced_shading = (major >= 3);
2478                 info.has_half_images = (major >= 3);
2479                 info.has_volume_decoupled = false;
2480                 info.bvh_layout_mask = BVH_LAYOUT_BVH2;
2481
2482                 int pci_location[3] = {0, 0, 0};
2483                 cuDeviceGetAttribute(&pci_location[0], CU_DEVICE_ATTRIBUTE_PCI_DOMAIN_ID, num);
2484                 cuDeviceGetAttribute(&pci_location[1], CU_DEVICE_ATTRIBUTE_PCI_BUS_ID, num);
2485                 cuDeviceGetAttribute(&pci_location[2], CU_DEVICE_ATTRIBUTE_PCI_DEVICE_ID, num);
2486                 info.id = string_printf("CUDA_%s_%04x:%02x:%02x",
2487                                         name,
2488                                         (unsigned int)pci_location[0],
2489                                         (unsigned int)pci_location[1],
2490                                         (unsigned int)pci_location[2]);
2491
2492                 /* If device has a kernel timeout and no compute preemption, we assume
2493                  * it is connected to a display and will freeze the display while doing
2494                  * computations. */
2495                 int timeout_attr = 0, preempt_attr = 0;
2496                 cuDeviceGetAttribute(&timeout_attr, CU_DEVICE_ATTRIBUTE_KERNEL_EXEC_TIMEOUT, num);
2497                 cuDeviceGetAttribute(&preempt_attr, CU_DEVICE_ATTRIBUTE_COMPUTE_PREEMPTION_SUPPORTED, num);
2498
2499                 if(timeout_attr && !preempt_attr) {
2500                         VLOG(1) << "Device is recognized as display.";
2501                         info.description += " (Display)";
2502                         info.display_device = true;
2503                         display_devices.push_back(info);
2504                 }
2505                 else {
2506                         devices.push_back(info);
2507                 }
2508                 VLOG(1) << "Added device \"" << name << "\" with id \"" << info.id << "\".";
2509         }
2510
2511         if(!display_devices.empty())
2512                 devices.insert(devices.end(), display_devices.begin(), display_devices.end());
2513 }
2514
2515 string device_cuda_capabilities(void)
2516 {
2517         CUresult result = device_cuda_safe_init();
2518         if(result != CUDA_SUCCESS) {
2519                 if(result != CUDA_ERROR_NO_DEVICE) {
2520                         return string("Error initializing CUDA: ") + cuewErrorString(result);
2521                 }
2522                 return "No CUDA device found\n";
2523         }
2524
2525         int count;
2526         result = cuDeviceGetCount(&count);
2527         if(result != CUDA_SUCCESS) {
2528                 return string("Error getting devices: ") + cuewErrorString(result);
2529         }
2530
2531         string capabilities = "";
2532         for(int num = 0; num < count; num++) {
2533                 char name[256];
2534                 if(cuDeviceGetName(name, 256, num) != CUDA_SUCCESS) {
2535                         continue;
2536                 }
2537                 capabilities += string("\t") + name + "\n";
2538                 int value;
2539 #define GET_ATTR(attr) \
2540                 { \
2541                         if(cuDeviceGetAttribute(&value, \
2542                                                 CU_DEVICE_ATTRIBUTE_##attr, \
2543                                                 num) == CUDA_SUCCESS) \
2544                         { \
2545                                 capabilities += string_printf("\t\tCU_DEVICE_ATTRIBUTE_" #attr "\t\t\t%d\n", \
2546                                                               value); \
2547                         } \
2548                 } (void)0
2549                 /* TODO(sergey): Strip all attributes which are not useful for us
2550                  * or does not depend on the driver.
2551                  */
2552                 GET_ATTR(MAX_THREADS_PER_BLOCK);
2553                 GET_ATTR(MAX_BLOCK_DIM_X);
2554                 GET_ATTR(MAX_BLOCK_DIM_Y);
2555                 GET_ATTR(MAX_BLOCK_DIM_Z);
2556                 GET_ATTR(MAX_GRID_DIM_X);
2557                 GET_ATTR(MAX_GRID_DIM_Y);
2558                 GET_ATTR(MAX_GRID_DIM_Z);
2559                 GET_ATTR(MAX_SHARED_MEMORY_PER_BLOCK);
2560                 GET_ATTR(SHARED_MEMORY_PER_BLOCK);
2561                 GET_ATTR(TOTAL_CONSTANT_MEMORY);
2562                 GET_ATTR(WARP_SIZE);
2563                 GET_ATTR(MAX_PITCH);
2564                 GET_ATTR(MAX_REGISTERS_PER_BLOCK);
2565                 GET_ATTR(REGISTERS_PER_BLOCK);
2566                 GET_ATTR(CLOCK_RATE);
2567                 GET_ATTR(TEXTURE_ALIGNMENT);
2568                 GET_ATTR(GPU_OVERLAP);
2569                 GET_ATTR(MULTIPROCESSOR_COUNT);
2570                 GET_ATTR(KERNEL_EXEC_TIMEOUT);
2571                 GET_ATTR(INTEGRATED);
2572                 GET_ATTR(CAN_MAP_HOST_MEMORY);
2573                 GET_ATTR(COMPUTE_MODE);
2574                 GET_ATTR(MAXIMUM_TEXTURE1D_WIDTH);
2575                 GET_ATTR(MAXIMUM_TEXTURE2D_WIDTH);
2576                 GET_ATTR(MAXIMUM_TEXTURE2D_HEIGHT);
2577                 GET_ATTR(MAXIMUM_TEXTURE3D_WIDTH);
2578                 GET_ATTR(MAXIMUM_TEXTURE3D_HEIGHT);
2579                 GET_ATTR(MAXIMUM_TEXTURE3D_DEPTH);
2580                 GET_ATTR(MAXIMUM_TEXTURE2D_LAYERED_WIDTH);
2581                 GET_ATTR(MAXIMUM_TEXTURE2D_LAYERED_HEIGHT);
2582                 GET_ATTR(MAXIMUM_TEXTURE2D_LAYERED_LAYERS);
2583                 GET_ATTR(MAXIMUM_TEXTURE2D_ARRAY_WIDTH);
2584                 GET_ATTR(MAXIMUM_TEXTURE2D_ARRAY_HEIGHT);
2585                 GET_ATTR(MAXIMUM_TEXTURE2D_ARRAY_NUMSLICES);
2586                 GET_ATTR(SURFACE_ALIGNMENT);
2587                 GET_ATTR(CONCURRENT_KERNELS);
2588                 GET_ATTR(ECC_ENABLED);
2589                 GET_ATTR(TCC_DRIVER);
2590                 GET_ATTR(MEMORY_CLOCK_RATE);
2591                 GET_ATTR(GLOBAL_MEMORY_BUS_WIDTH);
2592                 GET_ATTR(L2_CACHE_SIZE);
2593                 GET_ATTR(MAX_THREADS_PER_MULTIPROCESSOR);
2594                 GET_ATTR(ASYNC_ENGINE_COUNT);
2595                 GET_ATTR(UNIFIED_ADDRESSING);
2596                 GET_ATTR(MAXIMUM_TEXTURE1D_LAYERED_WIDTH);
2597                 GET_ATTR(MAXIMUM_TEXTURE1D_LAYERED_LAYERS);
2598                 GET_ATTR(CAN_TEX2D_GATHER);
2599                 GET_ATTR(MAXIMUM_TEXTURE2D_GATHER_WIDTH);
2600                 GET_ATTR(MAXIMUM_TEXTURE2D_GATHER_HEIGHT);
2601                 GET_ATTR(MAXIMUM_TEXTURE3D_WIDTH_ALTERNATE);
2602                 GET_ATTR(MAXIMUM_TEXTURE3D_HEIGHT_ALTERNATE);
2603                 GET_ATTR(MAXIMUM_TEXTURE3D_DEPTH_ALTERNATE);
2604                 GET_ATTR(TEXTURE_PITCH_ALIGNMENT);
2605                 GET_ATTR(MAXIMUM_TEXTURECUBEMAP_WIDTH);
2606                 GET_ATTR(MAXIMUM_TEXTURECUBEMAP_LAYERED_WIDTH);
2607                 GET_ATTR(MAXIMUM_TEXTURECUBEMAP_LAYERED_LAYERS);
2608                 GET_ATTR(MAXIMUM_SURFACE1D_WIDTH);
2609                 GET_ATTR(MAXIMUM_SURFACE2D_WIDTH);
2610                 GET_ATTR(MAXIMUM_SURFACE2D_HEIGHT);
2611                 GET_ATTR(MAXIMUM_SURFACE3D_WIDTH);
2612                 GET_ATTR(MAXIMUM_SURFACE3D_HEIGHT);
2613                 GET_ATTR(MAXIMUM_SURFACE3D_DEPTH);
2614                 GET_ATTR(MAXIMUM_SURFACE1D_LAYERED_WIDTH);
2615                 GET_ATTR(MAXIMUM_SURFACE1D_LAYERED_LAYERS);
2616                 GET_ATTR(MAXIMUM_SURFACE2D_LAYERED_WIDTH);
2617                 GET_ATTR(MAXIMUM_SURFACE2D_LAYERED_HEIGHT);
2618                 GET_ATTR(MAXIMUM_SURFACE2D_LAYERED_LAYERS);
2619                 GET_ATTR(MAXIMUM_SURFACECUBEMAP_WIDTH);
2620                 GET_ATTR(MAXIMUM_SURFACECUBEMAP_LAYERED_WIDTH);
2621                 GET_ATTR(MAXIMUM_SURFACECUBEMAP_LAYERED_LAYERS);
2622                 GET_ATTR(MAXIMUM_TEXTURE1D_LINEAR_WIDTH);
2623                 GET_ATTR(MAXIMUM_TEXTURE2D_LINEAR_WIDTH);
2624                 GET_ATTR(MAXIMUM_TEXTURE2D_LINEAR_HEIGHT);
2625                 GET_ATTR(MAXIMUM_TEXTURE2D_LINEAR_PITCH);
2626                 GET_ATTR(MAXIMUM_TEXTURE2D_MIPMAPPED_WIDTH);
2627                 GET_ATTR(MAXIMUM_TEXTURE2D_MIPMAPPED_HEIGHT);
2628                 GET_ATTR(COMPUTE_CAPABILITY_MAJOR);
2629                 GET_ATTR(COMPUTE_CAPABILITY_MINOR);
2630                 GET_ATTR(MAXIMUM_TEXTURE1D_MIPMAPPED_WIDTH);
2631                 GET_ATTR(STREAM_PRIORITIES_SUPPORTED);
2632                 GET_ATTR(GLOBAL_L1_CACHE_SUPPORTED);
2633                 GET_ATTR(LOCAL_L1_CACHE_SUPPORTED);
2634                 GET_ATTR(MAX_SHARED_MEMORY_PER_MULTIPROCESSOR);
2635                 GET_ATTR(MAX_REGISTERS_PER_MULTIPROCESSOR);
2636                 GET_ATTR(MANAGED_MEMORY);
2637                 GET_ATTR(MULTI_GPU_BOARD);
2638                 GET_ATTR(MULTI_GPU_BOARD_GROUP_ID);
2639 #undef GET_ATTR
2640                 capabilities += "\n";
2641         }
2642
2643         return capabilities;
2644 }
2645
2646 CCL_NAMESPACE_END