2 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
6 % OOO PPPP EEEEE N N CCCC L %
8 % O O PPPP EEE N N N C L %
10 % OOO P EEEEE N N CCCC LLLLL %
13 % MagickCore OpenCL Methods %
20 % Copyright 1999-2018 ImageMagick Studio LLC, a non-profit organization %
21 % dedicated to making software imaging solutions freely available. %
23 % You may not use this file except in compliance with the License. You may %
24 % obtain a copy of the License at %
26 % https://imagemagick.org/script/license.php %
28 % Unless required by applicable law or agreed to in writing, software %
29 % distributed under the License is distributed on an "AS IS" BASIS, %
30 % WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. %
31 % See the License for the specific language governing permissions and %
32 % limitations under the License. %
34 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
43 #include "MagickCore/studio.h"
44 #include "MagickCore/artifact.h"
45 #include "MagickCore/cache.h"
46 #include "MagickCore/cache-private.h"
47 #include "MagickCore/color.h"
48 #include "MagickCore/compare.h"
49 #include "MagickCore/constitute.h"
50 #include "MagickCore/configure.h"
51 #include "MagickCore/distort.h"
52 #include "MagickCore/draw.h"
53 #include "MagickCore/effect.h"
54 #include "MagickCore/exception.h"
55 #include "MagickCore/exception-private.h"
56 #include "MagickCore/fx.h"
57 #include "MagickCore/gem.h"
58 #include "MagickCore/geometry.h"
59 #include "MagickCore/image.h"
60 #include "MagickCore/image-private.h"
61 #include "MagickCore/layer.h"
62 #include "MagickCore/mime-private.h"
63 #include "MagickCore/memory_.h"
64 #include "MagickCore/memory-private.h"
65 #include "MagickCore/monitor.h"
66 #include "MagickCore/montage.h"
67 #include "MagickCore/morphology.h"
68 #include "MagickCore/nt-base.h"
69 #include "MagickCore/nt-base-private.h"
70 #include "MagickCore/opencl.h"
71 #include "MagickCore/opencl-private.h"
72 #include "MagickCore/option.h"
73 #include "MagickCore/policy.h"
74 #include "MagickCore/property.h"
75 #include "MagickCore/quantize.h"
76 #include "MagickCore/quantum.h"
77 #include "MagickCore/random_.h"
78 #include "MagickCore/random-private.h"
79 #include "MagickCore/resample.h"
80 #include "MagickCore/resource_.h"
81 #include "MagickCore/splay-tree.h"
82 #include "MagickCore/semaphore.h"
83 #include "MagickCore/statistic.h"
84 #include "MagickCore/string_.h"
85 #include "MagickCore/string-private.h"
86 #include "MagickCore/token.h"
87 #include "MagickCore/utility.h"
88 #include "MagickCore/utility-private.h"
90 #if defined(MAGICKCORE_OPENCL_SUPPORT)
91 #if defined(MAGICKCORE_LTDL_DELEGATE)
95 #ifndef MAGICKCORE_WINDOWS_SUPPORT
99 #ifdef MAGICKCORE_HAVE_OPENCL_CL_H
100 #define MAGICKCORE_OPENCL_MACOSX 1
106 #define IMAGEMAGICK_PROFILE_FILE "ImagemagickOpenCLDeviceProfile.xml"
109 Typedef declarations.
132 } MagickCLDeviceBenchmark;
135 Forward declarations.
138 static MagickBooleanType
139 HasOpenCLDevices(MagickCLEnv,ExceptionInfo *),
140 LoadOpenCLLibrary(void);
142 static MagickCLDevice
143 RelinquishMagickCLDevice(MagickCLDevice);
146 RelinquishMagickCLEnv(MagickCLEnv);
149 BenchmarkOpenCLDevices(MagickCLEnv);
152 *accelerateKernels, *accelerateKernels2;
158 /* Default OpenCL environment */
166 /* Cached location of the OpenCL cache files */
170 *cache_directory_lock;
172 static inline MagickBooleanType IsSameOpenCLDevice(MagickCLDevice a,
175 if ((LocaleCompare(a->platform_name,b->platform_name) == 0) &&
176 (LocaleCompare(a->vendor_name,b->vendor_name) == 0) &&
177 (LocaleCompare(a->name,b->name) == 0) &&
178 (LocaleCompare(a->version,b->version) == 0) &&
179 (a->max_clock_frequency == b->max_clock_frequency) &&
180 (a->max_compute_units == b->max_compute_units))
186 static inline MagickBooleanType IsBenchmarkedOpenCLDevice(MagickCLDevice a,
187 MagickCLDeviceBenchmark *b)
189 if ((LocaleCompare(a->platform_name,b->platform_name) == 0) &&
190 (LocaleCompare(a->vendor_name,b->vendor_name) == 0) &&
191 (LocaleCompare(a->name,b->name) == 0) &&
192 (LocaleCompare(a->version,b->version) == 0) &&
193 (a->max_clock_frequency == b->max_clock_frequency) &&
194 (a->max_compute_units == b->max_compute_units))
200 static inline void RelinquishMagickCLDevices(MagickCLEnv clEnv)
205 if (clEnv->devices != (MagickCLDevice *) NULL)
207 for (i = 0; i < clEnv->number_devices; i++)
208 clEnv->devices[i]=RelinquishMagickCLDevice(clEnv->devices[i]);
209 clEnv->devices=(MagickCLDevice *) RelinquishMagickMemory(clEnv->devices);
211 clEnv->number_devices=0;
214 static inline MagickBooleanType MagickCreateDirectory(const char *path)
219 #ifdef MAGICKCORE_WINDOWS_SUPPORT
222 status=mkdir(path, 0777);
224 return(status == 0 ? MagickTrue : MagickFalse);
227 static inline void InitAccelerateTimer(AccelerateTimer *timer)
230 QueryPerformanceFrequency((LARGE_INTEGER*)&timer->freq);
232 timer->freq=(long long)1.0E3;
238 static inline double ReadAccelerateTimer(AccelerateTimer *timer)
240 return (double)timer->clocks/(double)timer->freq;
243 static inline void StartAccelerateTimer(AccelerateTimer* timer)
246 QueryPerformanceCounter((LARGE_INTEGER*)&timer->start);
251 timer->start=(long long)s.tv_sec*(long long)1.0E3+(long long)s.tv_usec/
256 static inline void StopAccelerateTimer(AccelerateTimer *timer)
263 QueryPerformanceCounter((LARGE_INTEGER*)&(n));
268 n=(long long)s.tv_sec*(long long)1.0E3+(long long)s.tv_usec/
276 static const char *GetOpenCLCacheDirectory()
278 if (cache_directory == (char *) NULL)
280 if (cache_directory_lock == (SemaphoreInfo *) NULL)
281 ActivateSemaphoreInfo(&cache_directory_lock);
282 LockSemaphoreInfo(cache_directory_lock);
283 if (cache_directory == (char *) NULL)
287 path[MagickPathExtent],
297 home=GetEnvironmentValue("MAGICK_OPENCL_CACHE_DIR");
298 if (home == (char *) NULL)
300 home=GetEnvironmentValue("XDG_CACHE_HOME");
301 if (home == (char *) NULL)
302 home=GetEnvironmentValue("LOCALAPPDATA");
303 if (home == (char *) NULL)
304 home=GetEnvironmentValue("APPDATA");
305 if (home == (char *) NULL)
306 home=GetEnvironmentValue("USERPROFILE");
309 if (home != (char *) NULL)
311 /* first check if $HOME exists */
312 (void) FormatLocaleString(path,MagickPathExtent,"%s",home);
313 status=GetPathAttributes(path,&attributes);
314 if (status == MagickFalse)
315 status=MagickCreateDirectory(path);
317 /* first check if $HOME/ImageMagick exists */
318 if (status != MagickFalse)
320 (void) FormatLocaleString(path,MagickPathExtent,
321 "%s%sImageMagick",home,DirectorySeparator);
323 status=GetPathAttributes(path,&attributes);
324 if (status == MagickFalse)
325 status=MagickCreateDirectory(path);
328 if (status != MagickFalse)
330 temp=(char*) AcquireCriticalMemory(strlen(path)+1);
331 CopyMagickString(temp,path,strlen(path)+1);
333 home=DestroyString(home);
337 home=GetEnvironmentValue("HOME");
338 if (home != (char *) NULL)
340 /* first check if $HOME/.cache exists */
341 (void) FormatLocaleString(path,MagickPathExtent,"%s%s.cache",
342 home,DirectorySeparator);
343 status=GetPathAttributes(path,&attributes);
344 if (status == MagickFalse)
345 status=MagickCreateDirectory(path);
347 /* first check if $HOME/.cache/ImageMagick exists */
348 if (status != MagickFalse)
350 (void) FormatLocaleString(path,MagickPathExtent,
351 "%s%s.cache%sImageMagick",home,DirectorySeparator,
353 status=GetPathAttributes(path,&attributes);
354 if (status == MagickFalse)
355 status=MagickCreateDirectory(path);
358 if (status != MagickFalse)
360 temp=(char*) AcquireCriticalMemory(strlen(path)+1);
361 CopyMagickString(temp,path,strlen(path)+1);
363 home=DestroyString(home);
366 if (temp == (char *) NULL)
367 temp=AcquireString("?");
368 cache_directory=temp;
370 UnlockSemaphoreInfo(cache_directory_lock);
372 if (*cache_directory == '?')
373 return((const char *) NULL);
374 return(cache_directory);
377 static void SelectOpenCLDevice(MagickCLEnv clEnv,cl_device_type type)
386 for (i = 0; i < clEnv->number_devices; i++)
387 clEnv->devices[i]->enabled=MagickFalse;
389 for (i = 0; i < clEnv->number_devices; i++)
391 device=clEnv->devices[i];
392 if (device->type != type)
395 device->enabled=MagickTrue;
396 for (j = i+1; j < clEnv->number_devices; j++)
401 other_device=clEnv->devices[j];
402 if (IsSameOpenCLDevice(device,other_device))
403 other_device->enabled=MagickTrue;
408 static size_t StringSignature(const char* string)
423 stringLength=(size_t) strlen(string);
424 signature=stringLength;
425 n=stringLength/sizeof(size_t);
427 for (i = 0; i < n; i++)
429 if (n * sizeof(size_t) != stringLength)
435 for (i = 0; i < 4; i++, j++)
437 if (j < stringLength)
448 static void DestroyMagickCLCacheInfo(MagickCLCacheInfo info)
453 for (i=0; i < (ssize_t) info->event_count; i++)
454 openCL_library->clReleaseEvent(info->events[i]);
455 info->events=(cl_event *) RelinquishMagickMemory(info->events);
456 if (info->buffer != (cl_mem) NULL)
457 openCL_library->clReleaseMemObject(info->buffer);
458 RelinquishSemaphoreInfo(&info->events_semaphore);
459 ReleaseOpenCLDevice(info->device);
460 RelinquishMagickMemory(info);
464 Provide call to OpenCL library methods
467 MagickPrivate cl_mem CreateOpenCLBuffer(MagickCLDevice device,
468 cl_mem_flags flags,size_t size,void *host_ptr)
470 return(openCL_library->clCreateBuffer(device->context,flags,size,host_ptr,
474 MagickPrivate void ReleaseOpenCLKernel(cl_kernel kernel)
476 (void) openCL_library->clReleaseKernel(kernel);
479 MagickPrivate void ReleaseOpenCLMemObject(cl_mem memobj)
481 (void) openCL_library->clReleaseMemObject(memobj);
484 MagickPrivate void RetainOpenCLMemObject(cl_mem memobj)
486 (void) openCL_library->clRetainMemObject(memobj);
489 MagickPrivate cl_int SetOpenCLKernelArg(cl_kernel kernel,size_t arg_index,
490 size_t arg_size,const void *arg_value)
492 return(openCL_library->clSetKernelArg(kernel,(cl_uint) arg_index,arg_size,
497 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
501 + A c q u i r e M a g i c k C L C a c h e I n f o %
505 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
507 % AcquireMagickCLCacheInfo() acquires an OpenCL cache info structure.
509 % The format of the AcquireMagickCLCacheInfo method is:
511 % MagickCLCacheInfo AcquireMagickCLCacheInfo(MagickCLDevice device,
512 % Quantum *pixels,const MagickSizeType length)
514 % A description of each parameter follows:
516 % o device: the OpenCL device.
518 % o pixels: the pixel buffer of the image.
520 % o length: the length of the pixel buffer.
524 MagickPrivate MagickCLCacheInfo AcquireMagickCLCacheInfo(MagickCLDevice device,
525 Quantum *pixels,const MagickSizeType length)
533 info=(MagickCLCacheInfo) AcquireCriticalMemory(sizeof(*info));
534 (void) memset(info,0,sizeof(*info));
535 LockSemaphoreInfo(openCL_lock);
537 UnlockSemaphoreInfo(openCL_lock);
541 info->events_semaphore=AcquireSemaphoreInfo();
542 info->buffer=openCL_library->clCreateBuffer(device->context,
543 CL_MEM_READ_WRITE | CL_MEM_USE_HOST_PTR,(size_t) length,(void *) pixels,
545 if (status == CL_SUCCESS)
547 DestroyMagickCLCacheInfo(info);
548 return((MagickCLCacheInfo) NULL);
552 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
556 % A c q u i r e M a g i c k C L D e v i c e %
560 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
562 % AcquireMagickCLDevice() acquires an OpenCL device
564 % The format of the AcquireMagickCLDevice method is:
566 % MagickCLDevice AcquireMagickCLDevice()
570 static MagickCLDevice AcquireMagickCLDevice()
575 device=(MagickCLDevice) AcquireMagickMemory(sizeof(*device));
578 (void) memset(device,0,sizeof(*device));
579 ActivateSemaphoreInfo(&device->lock);
580 device->score=MAGICKCORE_OPENCL_UNDEFINED_SCORE;
581 device->command_queues_index=-1;
582 device->enabled=MagickTrue;
588 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
592 % A c q u i r e M a g i c k C L E n v %
596 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
598 % AcquireMagickCLEnv() allocates the MagickCLEnv structure
602 static MagickCLEnv AcquireMagickCLEnv(void)
610 clEnv=(MagickCLEnv) AcquireMagickMemory(sizeof(*clEnv));
611 if (clEnv != (MagickCLEnv) NULL)
613 (void) memset(clEnv,0,sizeof(*clEnv));
614 ActivateSemaphoreInfo(&clEnv->lock);
615 clEnv->cpu_score=MAGICKCORE_OPENCL_UNDEFINED_SCORE;
616 clEnv->enabled=MagickTrue;
617 option=getenv("MAGICK_OCL_DEVICE");
618 if ((option != (const char *) NULL) && (strcmp(option,"OFF") == 0))
619 clEnv->enabled=MagickFalse;
625 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
629 + A c q u i r e O p e n C L C o m m a n d Q u e u e %
633 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
635 % AcquireOpenCLCommandQueue() acquires an OpenCL command queue
637 % The format of the AcquireOpenCLCommandQueue method is:
639 % cl_command_queue AcquireOpenCLCommandQueue(MagickCLDevice device)
641 % A description of each parameter follows:
643 % o device: the OpenCL device.
647 MagickPrivate cl_command_queue AcquireOpenCLCommandQueue(MagickCLDevice device)
652 cl_command_queue_properties
655 assert(device != (MagickCLDevice) NULL);
656 LockSemaphoreInfo(device->lock);
657 if ((device->profile_kernels == MagickFalse) &&
658 (device->command_queues_index >= 0))
660 queue=device->command_queues[device->command_queues_index--];
661 UnlockSemaphoreInfo(device->lock);
665 UnlockSemaphoreInfo(device->lock);
667 if (device->profile_kernels != MagickFalse)
668 properties=CL_QUEUE_PROFILING_ENABLE;
669 queue=openCL_library->clCreateCommandQueue(device->context,
670 device->deviceID,properties,(cl_int *) NULL);
676 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
680 + A c q u i r e O p e n C L K e r n e l %
684 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
686 % AcquireOpenCLKernel() acquires an OpenCL kernel
688 % The format of the AcquireOpenCLKernel method is:
690 % cl_kernel AcquireOpenCLKernel(MagickCLEnv clEnv,
691 % MagickOpenCLProgram program, const char* kernelName)
693 % A description of each parameter follows:
695 % o clEnv: the OpenCL environment.
697 % o program: the OpenCL program module that the kernel belongs to.
699 % o kernelName: the name of the kernel
703 MagickPrivate cl_kernel AcquireOpenCLKernel(MagickCLDevice device,
704 const char *kernel_name)
709 assert(device != (MagickCLDevice) NULL);
710 kernel=openCL_library->clCreateKernel(device->program,kernel_name,
716 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
720 % A u t o S e l e c t O p e n C L D e v i c e s %
724 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
726 % AutoSelectOpenCLDevices() determines the best device based on the
727 % information from the micro-benchmark.
729 % The format of the AutoSelectOpenCLDevices method is:
731 % void AcquireOpenCLKernel(MagickCLEnv clEnv,ExceptionInfo *exception)
733 % A description of each parameter follows:
735 % o clEnv: the OpenCL environment.
737 % o exception: return any errors or warnings in this structure.
741 static void LoadOpenCLDeviceBenchmark(MagickCLEnv clEnv,const char *xml)
744 keyword[MagickPathExtent],
750 MagickCLDeviceBenchmark
757 if (xml == (char *) NULL)
759 device_benchmark=(MagickCLDeviceBenchmark *) NULL;
760 token=AcquireString(xml);
761 extent=strlen(token)+MagickPathExtent;
762 for (q=(char *) xml; *q != '\0'; )
767 GetNextToken(q,&q,extent,token);
770 (void) CopyMagickString(keyword,token,MagickPathExtent);
771 if (LocaleNCompare(keyword,"<!DOCTYPE",9) == 0)
776 while ((LocaleNCompare(q,"]>",2) != 0) && (*q != '\0'))
777 GetNextToken(q,&q,extent,token);
780 if (LocaleNCompare(keyword,"<!--",4) == 0)
785 while ((LocaleNCompare(q,"->",2) != 0) && (*q != '\0'))
786 GetNextToken(q,&q,extent,token);
789 if (LocaleCompare(keyword,"<device") == 0)
794 device_benchmark=(MagickCLDeviceBenchmark *) AcquireMagickMemory(
795 sizeof(*device_benchmark));
796 if (device_benchmark == (MagickCLDeviceBenchmark *) NULL)
798 (void) memset(device_benchmark,0,sizeof(*device_benchmark));
799 device_benchmark->score=MAGICKCORE_OPENCL_UNDEFINED_SCORE;
802 if (device_benchmark == (MagickCLDeviceBenchmark *) NULL)
804 if (LocaleCompare(keyword,"/>") == 0)
806 if (device_benchmark->score != MAGICKCORE_OPENCL_UNDEFINED_SCORE)
808 if (LocaleCompare(device_benchmark->name, "CPU") == 0)
809 clEnv->cpu_score=device_benchmark->score;
816 Set the score for all devices that match this device.
818 for (i = 0; i < clEnv->number_devices; i++)
820 device=clEnv->devices[i];
821 if (IsBenchmarkedOpenCLDevice(device,device_benchmark))
822 device->score=device_benchmark->score;
827 device_benchmark->platform_name=RelinquishMagickMemory(
828 device_benchmark->platform_name);
829 device_benchmark->vendor_name=RelinquishMagickMemory(
830 device_benchmark->vendor_name);
831 device_benchmark->name=RelinquishMagickMemory(device_benchmark->name);
832 device_benchmark->version=RelinquishMagickMemory(
833 device_benchmark->version);
834 device_benchmark=(MagickCLDeviceBenchmark *) RelinquishMagickMemory(
838 GetNextToken(q,(const char **) NULL,extent,token);
841 GetNextToken(q,&q,extent,token);
842 GetNextToken(q,&q,extent,token);
848 if (LocaleCompare((char *) keyword,"maxClockFrequency") == 0)
850 device_benchmark->max_clock_frequency=StringToInteger(token);
853 if (LocaleCompare((char *) keyword,"maxComputeUnits") == 0)
855 device_benchmark->max_compute_units=StringToInteger(token);
863 if (LocaleCompare((char *) keyword,"name") == 0)
864 device_benchmark->name=ConstantString(token);
870 if (LocaleCompare((char *) keyword,"platform") == 0)
871 device_benchmark->platform_name=ConstantString(token);
877 if (LocaleCompare((char *) keyword,"score") == 0)
878 device_benchmark->score=StringToDouble(token,(char **) NULL);
884 if (LocaleCompare((char *) keyword,"vendor") == 0)
885 device_benchmark->vendor_name=ConstantString(token);
886 if (LocaleCompare((char *) keyword,"version") == 0)
887 device_benchmark->version=ConstantString(token);
894 token=(char *) RelinquishMagickMemory(token);
895 device_benchmark=(MagickCLDeviceBenchmark *) RelinquishMagickMemory(
899 static MagickBooleanType CanWriteProfileToFile(const char *filename)
904 profileFile=fopen(filename,"ab");
906 if (profileFile == (FILE *)NULL)
913 static MagickBooleanType LoadOpenCLBenchmarks(MagickCLEnv clEnv)
916 filename[MagickPathExtent];
924 (void) FormatLocaleString(filename,MagickPathExtent,"%s%s%s",
925 GetOpenCLCacheDirectory(),DirectorySeparator,IMAGEMAGICK_PROFILE_FILE);
928 We don't run the benchmark when we can not write out a device profile. The
929 first GPU device will be used.
931 #if !defined(MAGICKCORE_ZERO_CONFIGURATION_SUPPORT)
932 if (CanWriteProfileToFile(filename) == MagickFalse)
935 for (i = 0; i < clEnv->number_devices; i++)
936 clEnv->devices[i]->score=1.0;
938 SelectOpenCLDevice(clEnv,CL_DEVICE_TYPE_GPU);
942 option=ConfigureFileToStringInfo(filename);
943 LoadOpenCLDeviceBenchmark(clEnv,(const char *) GetStringInfoDatum(option));
944 option=DestroyStringInfo(option);
948 static void AutoSelectOpenCLDevices(MagickCLEnv clEnv)
962 option=getenv("MAGICK_OCL_DEVICE");
963 if (option != (const char *) NULL)
965 if (strcmp(option,"GPU") == 0)
966 SelectOpenCLDevice(clEnv,CL_DEVICE_TYPE_GPU);
967 else if (strcmp(option,"CPU") == 0)
968 SelectOpenCLDevice(clEnv,CL_DEVICE_TYPE_CPU);
969 else if (strcmp(option,"OFF") == 0)
971 for (i = 0; i < clEnv->number_devices; i++)
972 clEnv->devices[i]->enabled=MagickFalse;
973 clEnv->enabled=MagickFalse;
977 if (LoadOpenCLBenchmarks(clEnv) == MagickFalse)
980 benchmark=MagickFalse;
981 if (clEnv->cpu_score == MAGICKCORE_OPENCL_UNDEFINED_SCORE)
982 benchmark=MagickTrue;
985 for (i = 0; i < clEnv->number_devices; i++)
987 if (clEnv->devices[i]->score == MAGICKCORE_OPENCL_UNDEFINED_SCORE)
989 benchmark=MagickTrue;
995 if (benchmark != MagickFalse)
996 BenchmarkOpenCLDevices(clEnv);
998 best_score=clEnv->cpu_score;
999 for (i = 0; i < clEnv->number_devices; i++)
1000 best_score=MagickMin(clEnv->devices[i]->score,best_score);
1002 for (i = 0; i < clEnv->number_devices; i++)
1004 if (clEnv->devices[i]->score != best_score)
1005 clEnv->devices[i]->enabled=MagickFalse;
1010 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1014 % B e n c h m a r k O p e n C L D e v i c e s %
1018 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1020 % BenchmarkOpenCLDevices() benchmarks the OpenCL devices and the CPU to help
1021 % the automatic selection of the best device.
1023 % The format of the BenchmarkOpenCLDevices method is:
1025 % void BenchmarkOpenCLDevices(MagickCLEnv clEnv,ExceptionInfo *exception)
1027 % A description of each parameter follows:
1029 % o clEnv: the OpenCL environment.
1031 % o exception: return any errors or warnings
1034 static double RunOpenCLBenchmark(MagickBooleanType is_cpu)
1051 exception=AcquireExceptionInfo();
1052 imageInfo=AcquireImageInfo();
1053 CloneString(&imageInfo->size,"2048x1536");
1054 CopyMagickString(imageInfo->filename,"xc:none",MagickPathExtent);
1055 inputImage=ReadImage(imageInfo,exception);
1057 InitAccelerateTimer(&timer);
1059 for (i=0; i<=2; i++)
1067 StartAccelerateTimer(&timer);
1069 bluredImage=BlurImage(inputImage,10.0f,3.5f,exception);
1070 unsharpedImage=UnsharpMaskImage(bluredImage,2.0f,2.0f,50.0f,10.0f,
1072 resizedImage=ResizeImage(unsharpedImage,640,480,LanczosFilter,
1076 We need this to get a proper performance benchmark, the operations
1077 are executed asynchronous.
1079 if (is_cpu == MagickFalse)
1084 cache_info=(CacheInfo *) resizedImage->cache;
1085 if (cache_info->opencl != (MagickCLCacheInfo) NULL)
1086 openCL_library->clWaitForEvents(cache_info->opencl->event_count,
1087 cache_info->opencl->events);
1091 StopAccelerateTimer(&timer);
1093 if (bluredImage != (Image *) NULL)
1094 DestroyImage(bluredImage);
1095 if (unsharpedImage != (Image *) NULL)
1096 DestroyImage(unsharpedImage);
1097 if (resizedImage != (Image *) NULL)
1098 DestroyImage(resizedImage);
1100 DestroyImage(inputImage);
1101 return(ReadAccelerateTimer(&timer));
1104 static void RunDeviceBenckmark(MagickCLEnv clEnv,MagickCLEnv testEnv,
1105 MagickCLDevice device)
1107 testEnv->devices[0]=device;
1108 default_CLEnv=testEnv;
1109 device->score=RunOpenCLBenchmark(MagickFalse);
1110 default_CLEnv=clEnv;
1111 testEnv->devices[0]=(MagickCLDevice) NULL;
1114 static void CacheOpenCLBenchmarks(MagickCLEnv clEnv)
1117 filename[MagickPathExtent];
1129 (void) FormatLocaleString(filename,MagickPathExtent,"%s%s%s",
1130 GetOpenCLCacheDirectory(),DirectorySeparator,
1131 IMAGEMAGICK_PROFILE_FILE);
1133 cache_file=fopen_utf8(filename,"wb");
1134 if (cache_file == (FILE *) NULL)
1136 fwrite("<devices>\n",sizeof(char),10,cache_file);
1137 fprintf(cache_file," <device name=\"CPU\" score=\"%.4g\"/>\n",
1139 for (i = 0; i < clEnv->number_devices; i++)
1144 device=clEnv->devices[i];
1145 duplicate=MagickFalse;
1146 for (j = 0; j < i; j++)
1148 if (IsSameOpenCLDevice(clEnv->devices[j],device))
1150 duplicate=MagickTrue;
1158 if (device->score != MAGICKCORE_OPENCL_UNDEFINED_SCORE)
1159 fprintf(cache_file," <device platform=\"%s\" vendor=\"%s\" name=\"%s\"\
1160 version=\"%s\" maxClockFrequency=\"%d\" maxComputeUnits=\"%d\"\
1161 score=\"%.4g\"/>\n",
1162 device->platform_name,device->vendor_name,device->name,device->version,
1163 (int)device->max_clock_frequency,(int)device->max_compute_units,
1166 fwrite("</devices>",sizeof(char),10,cache_file);
1171 static void BenchmarkOpenCLDevices(MagickCLEnv clEnv)
1183 testEnv=AcquireMagickCLEnv();
1184 testEnv->library=openCL_library;
1185 testEnv->devices=(MagickCLDevice *) AcquireCriticalMemory(
1186 sizeof(MagickCLDevice));
1187 testEnv->number_devices=1;
1188 testEnv->benchmark_thread_id=GetMagickThreadId();
1189 testEnv->initialized=MagickTrue;
1191 for (i = 0; i < clEnv->number_devices; i++)
1192 clEnv->devices[i]->score=MAGICKCORE_OPENCL_UNDEFINED_SCORE;
1194 for (i = 0; i < clEnv->number_devices; i++)
1196 device=clEnv->devices[i];
1197 if (device->score == MAGICKCORE_OPENCL_UNDEFINED_SCORE)
1198 RunDeviceBenckmark(clEnv,testEnv,device);
1200 /* Set the score on all the other devices that are the same */
1201 for (j = i+1; j < clEnv->number_devices; j++)
1206 other_device=clEnv->devices[j];
1207 if (IsSameOpenCLDevice(device,other_device))
1208 other_device->score=device->score;
1212 testEnv->enabled=MagickFalse;
1213 default_CLEnv=testEnv;
1214 clEnv->cpu_score=RunOpenCLBenchmark(MagickTrue);
1215 default_CLEnv=clEnv;
1217 testEnv=RelinquishMagickCLEnv(testEnv);
1218 CacheOpenCLBenchmarks(clEnv);
1222 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1226 % C o m p i l e O p e n C L K e r n e l %
1230 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1232 % CompileOpenCLKernel() compiles the kernel for the specified device. The
1233 % kernel will be cached on disk to reduce the compilation time.
1235 % The format of the CompileOpenCLKernel method is:
1237 % MagickBooleanType AcquireOpenCLKernel(MagickCLDevice clEnv,
1238 % unsigned int signature,const char *kernel,const char *options,
1239 % ExceptionInfo *exception)
1241 % A description of each parameter follows:
1243 % o device: the OpenCL device.
1245 % o kernel: the source code of the kernel.
1247 % o options: options for the compiler.
1249 % o signature: a number to uniquely identify the kernel
1251 % o exception: return any errors or warnings in this structure.
1255 static void CacheOpenCLKernel(MagickCLDevice device,char *filename,
1256 ExceptionInfo *exception)
1267 status=openCL_library->clGetProgramInfo(device->program,
1268 CL_PROGRAM_BINARY_SIZES,sizeof(size_t),&binaryProgramSize,NULL);
1269 if (status != CL_SUCCESS)
1271 binaryProgram=(unsigned char*) AcquireMagickMemory(binaryProgramSize);
1272 if (binaryProgram == (unsigned char *) NULL)
1274 (void) ThrowMagickException(exception,GetMagickModule(),
1275 ResourceLimitError,"MemoryAllocationFailed","`%s'",filename);
1278 status=openCL_library->clGetProgramInfo(device->program,
1279 CL_PROGRAM_BINARIES,sizeof(unsigned char*),&binaryProgram,NULL);
1280 if (status == CL_SUCCESS)
1281 (void) BlobToFile(filename,binaryProgram,binaryProgramSize,exception);
1282 binaryProgram=(unsigned char *) RelinquishMagickMemory(binaryProgram);
1285 static MagickBooleanType LoadCachedOpenCLKernel(MagickCLDevice device,
1286 const char *filename)
1301 sans_exception=AcquireExceptionInfo();
1302 binaryProgram=(unsigned char *) FileToBlob(filename,~0UL,&length,
1304 sans_exception=DestroyExceptionInfo(sans_exception);
1305 if (binaryProgram == (unsigned char *) NULL)
1306 return(MagickFalse);
1307 device->program=openCL_library->clCreateProgramWithBinary(device->context,1,
1308 &device->deviceID,&length,(const unsigned char**)&binaryProgram,
1309 &binaryStatus,&status);
1310 binaryProgram=(unsigned char *) RelinquishMagickMemory(binaryProgram);
1311 return((status != CL_SUCCESS) || (binaryStatus != CL_SUCCESS) ? MagickFalse :
1315 static void LogOpenCLBuildFailure(MagickCLDevice device,const char *kernel,
1316 ExceptionInfo *exception)
1319 filename[MagickPathExtent],
1325 (void) FormatLocaleString(filename,MagickPathExtent,"%s%s%s",
1326 GetOpenCLCacheDirectory(),DirectorySeparator,"magick_badcl.cl");
1328 (void) remove_utf8(filename);
1329 (void) BlobToFile(filename,kernel,strlen(kernel),exception);
1331 openCL_library->clGetProgramBuildInfo(device->program,device->deviceID,
1332 CL_PROGRAM_BUILD_LOG,0,NULL,&log_size);
1333 log=(char*)AcquireCriticalMemory(log_size);
1334 openCL_library->clGetProgramBuildInfo(device->program,device->deviceID,
1335 CL_PROGRAM_BUILD_LOG,log_size,log,&log_size);
1337 (void) FormatLocaleString(filename,MagickPathExtent,"%s%s%s",
1338 GetOpenCLCacheDirectory(),DirectorySeparator,"magick_badcl.log");
1340 (void) remove_utf8(filename);
1341 (void) BlobToFile(filename,log,log_size,exception);
1342 log=(char*)RelinquishMagickMemory(log);
1345 static MagickBooleanType CompileOpenCLKernel(MagickCLDevice device,
1346 const char *kernel,const char *options,size_t signature,
1347 ExceptionInfo *exception)
1350 deviceName[MagickPathExtent],
1351 filename[MagickPathExtent],
1363 (void) CopyMagickString(deviceName,device->name,MagickPathExtent);
1365 /* Strip out illegal characters for file names */
1366 while (*ptr != '\0')
1368 if ((*ptr == ' ') || (*ptr == '\\') || (*ptr == '/') || (*ptr == ':') ||
1369 (*ptr == '*') || (*ptr == '?') || (*ptr == '"') || (*ptr == '<') ||
1370 (*ptr == '>' || *ptr == '|'))
1374 (void) FormatLocaleString(filename,MagickPathExtent,
1375 "%s%s%s_%s_%08x_%.20g.bin",GetOpenCLCacheDirectory(),
1376 DirectorySeparator,"magick_opencl",deviceName,(unsigned int) signature,
1377 (double) sizeof(char*)*8);
1378 loaded=LoadCachedOpenCLKernel(device,filename);
1379 if (loaded == MagickFalse)
1381 /* Binary CL program unavailable, compile the program from source */
1382 length=strlen(kernel);
1383 device->program=openCL_library->clCreateProgramWithSource(
1384 device->context,1,&kernel,&length,&status);
1385 if (status != CL_SUCCESS)
1386 return(MagickFalse);
1389 status=openCL_library->clBuildProgram(device->program,1,&device->deviceID,
1391 if (status != CL_SUCCESS)
1393 (void) ThrowMagickException(exception,GetMagickModule(),DelegateWarning,
1394 "clBuildProgram failed.","(%d)",(int)status);
1395 LogOpenCLBuildFailure(device,kernel,exception);
1396 return(MagickFalse);
1399 /* Save the binary to a file to avoid re-compilation of the kernels */
1400 if (loaded == MagickFalse)
1401 CacheOpenCLKernel(device,filename,exception);
1406 static cl_event* CopyOpenCLEvents(MagickCLCacheInfo first,
1407 MagickCLCacheInfo second,cl_uint *event_count)
1418 assert(first != (MagickCLCacheInfo) NULL);
1419 assert(event_count != (cl_uint *) NULL);
1420 events=(cl_event *) NULL;
1421 LockSemaphoreInfo(first->events_semaphore);
1422 if (second != (MagickCLCacheInfo) NULL)
1423 LockSemaphoreInfo(second->events_semaphore);
1424 *event_count=first->event_count;
1425 if (second != (MagickCLCacheInfo) NULL)
1426 *event_count+=second->event_count;
1427 if (*event_count > 0)
1429 events=AcquireQuantumMemory(*event_count,sizeof(*events));
1430 if (events == (cl_event *) NULL)
1435 for (i=0; i < first->event_count; i++, j++)
1436 events[j]=first->events[i];
1437 if (second != (MagickCLCacheInfo) NULL)
1439 for (i=0; i < second->event_count; i++, j++)
1440 events[j]=second->events[i];
1444 UnlockSemaphoreInfo(first->events_semaphore);
1445 if (second != (MagickCLCacheInfo) NULL)
1446 UnlockSemaphoreInfo(second->events_semaphore);
1451 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1455 + C o p y M a g i c k C L C a c h e I n f o %
1459 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1461 % CopyMagickCLCacheInfo() copies the memory from the device into host memory.
1463 % The format of the CopyMagickCLCacheInfo method is:
1465 % void CopyMagickCLCacheInfo(MagickCLCacheInfo info)
1467 % A description of each parameter follows:
1469 % o info: the OpenCL cache info.
1472 MagickPrivate MagickCLCacheInfo CopyMagickCLCacheInfo(MagickCLCacheInfo info)
1486 if (info == (MagickCLCacheInfo) NULL)
1487 return((MagickCLCacheInfo) NULL);
1488 events=CopyOpenCLEvents(info,(MagickCLCacheInfo) NULL,&event_count);
1489 if (events != (cl_event *) NULL)
1491 queue=AcquireOpenCLCommandQueue(info->device);
1492 pixels=openCL_library->clEnqueueMapBuffer(queue,info->buffer,CL_TRUE,
1493 CL_MAP_READ | CL_MAP_WRITE,0,info->length,event_count,events,
1494 (cl_event *) NULL,(cl_int *) NULL);
1495 assert(pixels == info->pixels);
1496 ReleaseOpenCLCommandQueue(info->device,queue);
1497 events=(cl_event *) RelinquishMagickMemory(events);
1499 return(RelinquishMagickCLCacheInfo(info,MagickFalse));
1503 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1507 + D u m p O p e n C L P r o f i l e D a t a %
1511 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1513 % DumpOpenCLProfileData() dumps the kernel profile data.
1515 % The format of the DumpProfileData method is:
1517 % void DumpProfileData()
1521 MagickPrivate void DumpOpenCLProfileData()
1523 #define OpenCLLog(message) \
1524 fwrite(message,sizeof(char),strlen(message),log); \
1525 fwrite("\n",sizeof(char),1,log);
1529 filename[MagickPathExtent],
1542 clEnv=GetCurrentOpenCLEnv();
1543 if (clEnv == (MagickCLEnv) NULL)
1546 for (i = 0; i < clEnv->number_devices; i++)
1547 if (clEnv->devices[i]->profile_kernels != MagickFalse)
1549 if (i == clEnv->number_devices)
1552 (void) FormatLocaleString(filename,MagickPathExtent,"%s%s%s",
1553 GetOpenCLCacheDirectory(),DirectorySeparator,"ImageMagickOpenCL.log");
1555 log=fopen_utf8(filename,"wb");
1557 for (i = 0; i < clEnv->number_devices; i++)
1562 device=clEnv->devices[i];
1563 if ((device->profile_kernels == MagickFalse) ||
1564 (device->profile_records == (KernelProfileRecord *) NULL))
1567 OpenCLLog("====================================================");
1568 fprintf(log,"Device: %s\n",device->name);
1569 fprintf(log,"Version: %s\n",device->version);
1570 OpenCLLog("====================================================");
1571 OpenCLLog(" average calls min max");
1572 OpenCLLog(" ------- ----- --- ---");
1574 while (device->profile_records[j] != (KernelProfileRecord) NULL)
1579 profile=device->profile_records[j];
1581 strncpy(indent,profile->kernel_name,MagickMin(strlen(
1582 profile->kernel_name),strlen(indent)-1));
1583 sprintf(buf,"%s %7d %7d %7d %7d",indent,(int) (profile->total/
1584 profile->count),(int) profile->count,(int) profile->min,
1585 (int) profile->max);
1589 OpenCLLog("====================================================");
1590 fwrite("\n\n",sizeof(char),2,log);
1595 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1599 + E n q u e u e O p e n C L K e r n e l %
1603 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1605 % EnqueueOpenCLKernel() enques the specified kernel and registers the OpenCL
1606 % events with the images.
1608 % The format of the EnqueueOpenCLKernel method is:
1610 % MagickBooleanType EnqueueOpenCLKernel(cl_kernel kernel,cl_uint work_dim,
1611 % const size_t *global_work_offset,const size_t *global_work_size,
1612 % const size_t *local_work_size,const Image *input_image,
1613 % const Image *output_image,ExceptionInfo *exception)
1615 % A description of each parameter follows:
1617 % o kernel: the OpenCL kernel.
1619 % o work_dim: the number of dimensions used to specify the global work-items
1620 % and work-items in the work-group.
1622 % o offset: can be used to specify an array of work_dim unsigned values
1623 % that describe the offset used to calculate the global ID of a
1626 % o gsize: points to an array of work_dim unsigned values that describe the
1627 % number of global work-items in work_dim dimensions that will
1628 % execute the kernel function.
1630 % o lsize: points to an array of work_dim unsigned values that describe the
1631 % number of work-items that make up a work-group that will execute
1632 % the kernel specified by kernel.
1634 % o input_image: the input image of the operation.
1636 % o output_image: the output or secondairy image of the operation.
1638 % o exception: return any errors or warnings in this structure.
1642 static MagickBooleanType RegisterCacheEvent(MagickCLCacheInfo info,
1645 assert(info != (MagickCLCacheInfo) NULL);
1646 assert(event != (cl_event) NULL);
1647 if (openCL_library->clRetainEvent(event) != CL_SUCCESS)
1649 openCL_library->clWaitForEvents(1,&event);
1650 return(MagickFalse);
1652 LockSemaphoreInfo(info->events_semaphore);
1653 if (info->events == (cl_event *) NULL)
1655 info->events=AcquireMagickMemory(sizeof(*info->events));
1656 info->event_count=1;
1659 info->events=ResizeQuantumMemory(info->events,++info->event_count,
1660 sizeof(*info->events));
1661 if (info->events == (cl_event *) NULL)
1662 ThrowFatalException(ResourceLimitFatalError,"MemoryAllocationFailed");
1663 info->events[info->event_count-1]=event;
1664 UnlockSemaphoreInfo(info->events_semaphore);
1668 MagickPrivate MagickBooleanType EnqueueOpenCLKernel(cl_command_queue queue,
1669 cl_kernel kernel,cl_uint work_dim,const size_t *offset,const size_t *gsize,
1670 const size_t *lsize,const Image *input_image,const Image *output_image,
1671 MagickBooleanType flush,ExceptionInfo *exception)
1687 assert(input_image != (const Image *) NULL);
1688 input_info=(CacheInfo *) input_image->cache;
1689 assert(input_info != (CacheInfo *) NULL);
1690 assert(input_info->opencl != (MagickCLCacheInfo) NULL);
1691 output_info=(CacheInfo *) NULL;
1692 if (output_image == (const Image *) NULL)
1693 events=CopyOpenCLEvents(input_info->opencl,(MagickCLCacheInfo) NULL,
1697 output_info=(CacheInfo *) output_image->cache;
1698 assert(output_info != (CacheInfo *) NULL);
1699 assert(output_info->opencl != (MagickCLCacheInfo) NULL);
1700 events=CopyOpenCLEvents(input_info->opencl,output_info->opencl,
1703 status=openCL_library->clEnqueueNDRangeKernel(queue,kernel,work_dim,offset,
1704 gsize,lsize,event_count,events,&event);
1705 /* This can fail due to memory issues and calling clFinish might help. */
1706 if ((status != CL_SUCCESS) && (event_count > 0))
1708 openCL_library->clFinish(queue);
1709 status=openCL_library->clEnqueueNDRangeKernel(queue,kernel,work_dim,
1710 offset,gsize,lsize,event_count,events,&event);
1712 events=(cl_event *) RelinquishMagickMemory(events);
1713 if (status != CL_SUCCESS)
1715 (void) OpenCLThrowMagickException(input_info->opencl->device,exception,
1716 GetMagickModule(),ResourceLimitWarning,
1717 "clEnqueueNDRangeKernel failed.","'%s'",".");
1718 return(MagickFalse);
1720 if (flush != MagickFalse)
1721 openCL_library->clFlush(queue);
1722 if (RecordProfileData(input_info->opencl->device,kernel,event) == MagickFalse)
1724 if (RegisterCacheEvent(input_info->opencl,event) != MagickFalse)
1726 if (output_info != (CacheInfo *) NULL)
1727 (void) RegisterCacheEvent(output_info->opencl,event);
1730 openCL_library->clReleaseEvent(event);
1735 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1739 + G e t C u r r u n t O p e n C L E n v %
1743 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1745 % GetCurrentOpenCLEnv() returns the current OpenCL env
1747 % The format of the GetCurrentOpenCLEnv method is:
1749 % MagickCLEnv GetCurrentOpenCLEnv()
1753 MagickPrivate MagickCLEnv GetCurrentOpenCLEnv(void)
1755 if (default_CLEnv != (MagickCLEnv) NULL)
1757 if ((default_CLEnv->benchmark_thread_id != (MagickThreadType) 0) &&
1758 (default_CLEnv->benchmark_thread_id != GetMagickThreadId()))
1759 return((MagickCLEnv) NULL);
1761 return(default_CLEnv);
1764 if (GetOpenCLCacheDirectory() == (char *) NULL)
1765 return((MagickCLEnv) NULL);
1767 if (openCL_lock == (SemaphoreInfo *) NULL)
1768 ActivateSemaphoreInfo(&openCL_lock);
1770 LockSemaphoreInfo(openCL_lock);
1771 if (default_CLEnv == (MagickCLEnv) NULL)
1772 default_CLEnv=AcquireMagickCLEnv();
1773 UnlockSemaphoreInfo(openCL_lock);
1775 return(default_CLEnv);
1779 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1783 % G e t O p e n C L D e v i c e B e n c h m a r k D u r a t i o n %
1787 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1789 % GetOpenCLDeviceBenchmarkScore() returns the score of the benchmark for the
1790 % device. The score is determined by the duration of the micro benchmark so
1791 % that means a lower score is better than a higher score.
1793 % The format of the GetOpenCLDeviceBenchmarkScore method is:
1795 % double GetOpenCLDeviceBenchmarkScore(const MagickCLDevice device)
1797 % A description of each parameter follows:
1799 % o device: the OpenCL device.
1802 MagickExport double GetOpenCLDeviceBenchmarkScore(
1803 const MagickCLDevice device)
1805 if (device == (MagickCLDevice) NULL)
1806 return(MAGICKCORE_OPENCL_UNDEFINED_SCORE);
1807 return(device->score);
1811 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1815 % G e t O p e n C L D e v i c e E n a b l e d %
1819 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1821 % GetOpenCLDeviceEnabled() returns true if the device is enabled.
1823 % The format of the GetOpenCLDeviceEnabled method is:
1825 % MagickBooleanType GetOpenCLDeviceEnabled(const MagickCLDevice device)
1827 % A description of each parameter follows:
1829 % o device: the OpenCL device.
1832 MagickExport MagickBooleanType GetOpenCLDeviceEnabled(
1833 const MagickCLDevice device)
1835 if (device == (MagickCLDevice) NULL)
1836 return(MagickFalse);
1837 return(device->enabled);
1841 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1845 % G e t O p e n C L D e v i c e N a m e %
1849 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1851 % GetOpenCLDeviceName() returns the name of the device.
1853 % The format of the GetOpenCLDeviceName method is:
1855 % const char *GetOpenCLDeviceName(const MagickCLDevice device)
1857 % A description of each parameter follows:
1859 % o device: the OpenCL device.
1862 MagickExport const char *GetOpenCLDeviceName(const MagickCLDevice device)
1864 if (device == (MagickCLDevice) NULL)
1865 return((const char *) NULL);
1866 return(device->name);
1870 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1874 % G e t O p e n C L D e v i c e V e n d o r N a m e %
1878 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1880 % GetOpenCLDeviceVendorName() returns the vendor name of the device.
1882 % The format of the GetOpenCLDeviceVendorName method is:
1884 % const char *GetOpenCLDeviceVendorName(const MagickCLDevice device)
1886 % A description of each parameter follows:
1888 % o device: the OpenCL device.
1891 MagickExport const char *GetOpenCLDeviceVendorName(const MagickCLDevice device)
1893 if (device == (MagickCLDevice) NULL)
1894 return((const char *) NULL);
1895 return(device->vendor_name);
1899 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1903 % G e t O p e n C L D e v i c e s %
1907 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1909 % GetOpenCLDevices() returns the devices of the OpenCL environment at sets the
1910 % value of length to the number of devices that are available.
1912 % The format of the GetOpenCLDevices method is:
1914 % const MagickCLDevice *GetOpenCLDevices(size_t *length,
1915 % ExceptionInfo *exception)
1917 % A description of each parameter follows:
1919 % o length: the number of device.
1921 % o exception: return any errors or warnings in this structure.
1925 MagickExport MagickCLDevice *GetOpenCLDevices(size_t *length,
1926 ExceptionInfo *exception)
1931 clEnv=GetCurrentOpenCLEnv();
1932 if (clEnv == (MagickCLEnv) NULL)
1934 if (length != (size_t *) NULL)
1936 return((MagickCLDevice *) NULL);
1938 InitializeOpenCL(clEnv,exception);
1939 if (length != (size_t *) NULL)
1940 *length=clEnv->number_devices;
1941 return(clEnv->devices);
1945 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1949 % G e t O p e n C L D e v i c e T y p e %
1953 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1955 % GetOpenCLDeviceType() returns the type of the device.
1957 % The format of the GetOpenCLDeviceType method is:
1959 % MagickCLDeviceType GetOpenCLDeviceType(const MagickCLDevice device)
1961 % A description of each parameter follows:
1963 % o device: the OpenCL device.
1966 MagickExport MagickCLDeviceType GetOpenCLDeviceType(
1967 const MagickCLDevice device)
1969 if (device == (MagickCLDevice) NULL)
1970 return(UndefinedCLDeviceType);
1971 if (device->type == CL_DEVICE_TYPE_GPU)
1972 return(GpuCLDeviceType);
1973 if (device->type == CL_DEVICE_TYPE_CPU)
1974 return(CpuCLDeviceType);
1975 return(UndefinedCLDeviceType);
1979 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1983 % G e t O p e n C L D e v i c e V e r s i o n %
1987 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
1989 % GetOpenCLDeviceVersion() returns the version of the device.
1991 % The format of the GetOpenCLDeviceName method is:
1993 % const char *GetOpenCLDeviceVersion(MagickCLDevice device)
1995 % A description of each parameter follows:
1997 % o device: the OpenCL device.
2000 MagickExport const char *GetOpenCLDeviceVersion(const MagickCLDevice device)
2002 if (device == (MagickCLDevice) NULL)
2003 return((const char *) NULL);
2004 return(device->version);
2008 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2012 % G e t O p e n C L E n a b l e d %
2016 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2018 % GetOpenCLEnabled() returns true if OpenCL acceleration is enabled.
2020 % The format of the GetOpenCLEnabled method is:
2022 % MagickBooleanType GetOpenCLEnabled()
2026 MagickExport MagickBooleanType GetOpenCLEnabled(void)
2031 clEnv=GetCurrentOpenCLEnv();
2032 if (clEnv == (MagickCLEnv) NULL)
2033 return(MagickFalse);
2034 return(clEnv->enabled);
2038 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2042 % G e t O p e n C L K e r n e l P r o f i l e R e c o r d s %
2046 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2048 % GetOpenCLKernelProfileRecords() returns the profile records for the
2049 % specified device and sets length to the number of profile records.
2051 % The format of the GetOpenCLKernelProfileRecords method is:
2053 % const KernelProfileRecord *GetOpenCLKernelProfileRecords(size *length)
2055 % A description of each parameter follows:
2057 % o length: the number of profiles records.
2060 MagickExport const KernelProfileRecord *GetOpenCLKernelProfileRecords(
2061 const MagickCLDevice device,size_t *length)
2063 if ((device == (const MagickCLDevice) NULL) || (device->profile_records ==
2064 (KernelProfileRecord *) NULL))
2066 if (length != (size_t *) NULL)
2068 return((const KernelProfileRecord *) NULL);
2070 if (length != (size_t *) NULL)
2073 LockSemaphoreInfo(device->lock);
2074 while (device->profile_records[*length] != (KernelProfileRecord) NULL)
2076 UnlockSemaphoreInfo(device->lock);
2078 return(device->profile_records);
2082 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2086 % H a s O p e n C L D e v i c e s %
2090 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2092 % HasOpenCLDevices() checks if the OpenCL environment has devices that are
2093 % enabled and compiles the kernel for the device when necessary. False will be
2094 % returned if no enabled devices could be found
2096 % The format of the HasOpenCLDevices method is:
2098 % MagickBooleanType HasOpenCLDevices(MagickCLEnv clEnv,
2099 % ExceptionInfo exception)
2101 % A description of each parameter follows:
2103 % o clEnv: the OpenCL environment.
2105 % o exception: return any errors or warnings in this structure.
2109 static MagickBooleanType HasOpenCLDevices(MagickCLEnv clEnv,
2110 ExceptionInfo *exception)
2113 *accelerateKernelsBuffer,
2114 options[MagickPathExtent];
2125 /* Check if there are enabled devices */
2126 for (i = 0; i < clEnv->number_devices; i++)
2128 if ((clEnv->devices[i]->enabled != MagickFalse))
2131 if (i == clEnv->number_devices)
2132 return(MagickFalse);
2134 /* Check if we need to compile a kernel for one of the devices */
2136 for (i = 0; i < clEnv->number_devices; i++)
2138 if ((clEnv->devices[i]->enabled != MagickFalse) &&
2139 (clEnv->devices[i]->program == (cl_program) NULL))
2145 if (status != MagickFalse)
2148 /* Get additional options */
2149 (void) FormatLocaleString(options,MaxTextExtent,CLOptions,
2150 (float)QuantumRange,(float)QuantumScale,(float)CLCharQuantumScale,
2151 (float)MagickEpsilon,(float)MagickPI,(unsigned int)MaxMap,
2152 (unsigned int)MAGICKCORE_QUANTUM_DEPTH);
2154 signature=StringSignature(options);
2155 accelerateKernelsBuffer=(char*) AcquireMagickMemory(
2156 strlen(accelerateKernels)+strlen(accelerateKernels2)+1);
2157 if (accelerateKernelsBuffer == (char*) NULL)
2158 return(MagickFalse);
2159 sprintf(accelerateKernelsBuffer,"%s%s",accelerateKernels,accelerateKernels2);
2160 signature^=StringSignature(accelerateKernelsBuffer);
2163 for (i = 0; i < clEnv->number_devices; i++)
2171 device=clEnv->devices[i];
2172 if ((device->enabled == MagickFalse) ||
2173 (device->program != (cl_program) NULL))
2176 LockSemaphoreInfo(device->lock);
2177 if (device->program != (cl_program) NULL)
2179 UnlockSemaphoreInfo(device->lock);
2182 device_signature=signature;
2183 device_signature^=StringSignature(device->platform_name);
2184 status=CompileOpenCLKernel(device,accelerateKernelsBuffer,options,
2185 device_signature,exception);
2186 UnlockSemaphoreInfo(device->lock);
2187 if (status == MagickFalse)
2190 accelerateKernelsBuffer=RelinquishMagickMemory(accelerateKernelsBuffer);
2195 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2199 + I n i t i a l i z e O p e n C L %
2203 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2205 % InitializeOpenCL() is used to initialize the OpenCL environment. This method
2206 % makes sure the devices are propertly initialized and benchmarked.
2208 % The format of the InitializeOpenCL method is:
2210 % MagickBooleanType InitializeOpenCL(ExceptionInfo exception)
2212 % A description of each parameter follows:
2214 % o exception: return any errors or warnings in this structure.
2218 static cl_uint GetOpenCLDeviceCount(MagickCLEnv clEnv,cl_platform_id platform)
2221 version[MagickPathExtent];
2226 if (clEnv->library->clGetPlatformInfo(platform,CL_PLATFORM_VERSION,
2227 MagickPathExtent,version,NULL) != CL_SUCCESS)
2229 if (strncmp(version, "OpenCL 1.0 ", 11) == 0)
2231 if (clEnv->library->clGetDeviceIDs(platform,
2232 CL_DEVICE_TYPE_CPU|CL_DEVICE_TYPE_GPU,0,NULL,&num) != CL_SUCCESS)
2237 static void LoadOpenCLDevices(MagickCLEnv clEnv)
2239 cl_context_properties
2262 if (openCL_library->clGetPlatformIDs(0,NULL,&number_platforms) != CL_SUCCESS)
2264 if (number_platforms == 0)
2266 platforms=(cl_platform_id *) AcquireMagickMemory(number_platforms*
2267 sizeof(cl_platform_id));
2268 if (platforms == (cl_platform_id *) NULL)
2270 if (openCL_library->clGetPlatformIDs(number_platforms,platforms,NULL) != CL_SUCCESS)
2272 platforms=(cl_platform_id *) RelinquishMagickMemory(platforms);
2275 for (i = 0; i < number_platforms; i++)
2277 number_devices=GetOpenCLDeviceCount(clEnv,platforms[i]);
2278 if (number_devices == 0)
2279 platforms[i]=(cl_platform_id) NULL;
2281 clEnv->number_devices+=number_devices;
2283 if (clEnv->number_devices == 0)
2285 platforms=(cl_platform_id *) RelinquishMagickMemory(platforms);
2288 clEnv->devices=(MagickCLDevice *) AcquireQuantumMemory(clEnv->number_devices,
2289 sizeof(MagickCLDevice));
2290 if (clEnv->devices == (MagickCLDevice *) NULL)
2292 RelinquishMagickCLDevices(clEnv);
2293 platforms=(cl_platform_id *) RelinquishMagickMemory(platforms);
2296 (void) memset(clEnv->devices,0,clEnv->number_devices*
2297 sizeof(MagickCLDevice));
2298 devices=(cl_device_id *) AcquireQuantumMemory(clEnv->number_devices,
2299 sizeof(cl_device_id));
2300 if (devices == (cl_device_id *) NULL)
2302 platforms=(cl_platform_id *) RelinquishMagickMemory(platforms);
2303 RelinquishMagickCLDevices(clEnv);
2306 clEnv->number_contexts=(size_t) number_platforms;
2307 clEnv->contexts=(cl_context *) AcquireQuantumMemory(clEnv->number_contexts,
2308 sizeof(cl_context));
2309 if (clEnv->contexts == (cl_context *) NULL)
2311 devices=(cl_device_id *) RelinquishMagickMemory(devices);
2312 platforms=(cl_platform_id *) RelinquishMagickMemory(platforms);
2313 RelinquishMagickCLDevices(clEnv);
2317 for (i = 0; i < number_platforms; i++)
2319 if (platforms[i] == (cl_platform_id) NULL)
2322 status=clEnv->library->clGetDeviceIDs(platforms[i],CL_DEVICE_TYPE_CPU |
2323 CL_DEVICE_TYPE_GPU,(cl_uint) clEnv->number_devices,devices,&number_devices);
2324 if (status != CL_SUCCESS)
2327 properties[0]=CL_CONTEXT_PLATFORM;
2328 properties[1]=(cl_context_properties) platforms[i];
2330 clEnv->contexts[i]=openCL_library->clCreateContext(properties,number_devices,
2331 devices,NULL,NULL,&status);
2332 if (status != CL_SUCCESS)
2335 for (j = 0; j < number_devices; j++,next++)
2340 device=AcquireMagickCLDevice();
2341 if (device == (MagickCLDevice) NULL)
2344 device->context=clEnv->contexts[i];
2345 device->deviceID=devices[j];
2347 openCL_library->clGetPlatformInfo(platforms[i],CL_PLATFORM_NAME,0,NULL,
2349 device->platform_name=AcquireCriticalMemory(length*
2350 sizeof(*device->platform_name));
2351 openCL_library->clGetPlatformInfo(platforms[i],CL_PLATFORM_NAME,length,
2352 device->platform_name,NULL);
2354 openCL_library->clGetPlatformInfo(platforms[i],CL_PLATFORM_VENDOR,0,NULL,
2356 device->vendor_name=AcquireCriticalMemory(length*
2357 sizeof(*device->vendor_name));
2358 openCL_library->clGetPlatformInfo(platforms[i],CL_PLATFORM_VENDOR,length,
2359 device->vendor_name,NULL);
2361 openCL_library->clGetDeviceInfo(devices[j],CL_DEVICE_NAME,0,NULL,
2363 device->name=AcquireCriticalMemory(length*sizeof(*device->name));
2364 openCL_library->clGetDeviceInfo(devices[j],CL_DEVICE_NAME,length,
2367 openCL_library->clGetDeviceInfo(devices[j],CL_DRIVER_VERSION,0,NULL,
2369 device->version=AcquireCriticalMemory(length*sizeof(*device->version));
2370 openCL_library->clGetDeviceInfo(devices[j],CL_DRIVER_VERSION,length,
2371 device->version,NULL);
2373 openCL_library->clGetDeviceInfo(devices[j],CL_DEVICE_MAX_CLOCK_FREQUENCY,
2374 sizeof(cl_uint),&device->max_clock_frequency,NULL);
2376 openCL_library->clGetDeviceInfo(devices[j],CL_DEVICE_MAX_COMPUTE_UNITS,
2377 sizeof(cl_uint),&device->max_compute_units,NULL);
2379 openCL_library->clGetDeviceInfo(devices[j],CL_DEVICE_TYPE,
2380 sizeof(cl_device_type),&device->type,NULL);
2382 openCL_library->clGetDeviceInfo(devices[j],CL_DEVICE_LOCAL_MEM_SIZE,
2383 sizeof(cl_ulong),&device->local_memory_size,NULL);
2385 clEnv->devices[next]=device;
2388 if (next != clEnv->number_devices)
2389 RelinquishMagickCLDevices(clEnv);
2390 platforms=(cl_platform_id *) RelinquishMagickMemory(platforms);
2391 devices=(cl_device_id *) RelinquishMagickMemory(devices);
2394 MagickPrivate MagickBooleanType InitializeOpenCL(MagickCLEnv clEnv,
2395 ExceptionInfo *exception)
2400 LockSemaphoreInfo(clEnv->lock);
2401 if (clEnv->initialized != MagickFalse)
2403 UnlockSemaphoreInfo(clEnv->lock);
2404 return(HasOpenCLDevices(clEnv,exception));
2406 if (LoadOpenCLLibrary() != MagickFalse)
2408 clEnv->library=openCL_library;
2409 LoadOpenCLDevices(clEnv);
2410 if (clEnv->number_devices > 0)
2411 AutoSelectOpenCLDevices(clEnv);
2413 clEnv->initialized=MagickTrue;
2414 /* NVIDIA is disabled by default due to reported access violation */
2415 for (i=0; i < (ssize_t) clEnv->number_devices; i++)
2417 if (strncmp(clEnv->devices[i]->platform_name,"NVIDIA",6) == 0)
2418 clEnv->devices[i]->enabled=MagickFalse;
2420 UnlockSemaphoreInfo(clEnv->lock);
2421 return(HasOpenCLDevices(clEnv,exception));
2425 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2429 % L o a d O p e n C L L i b r a r y %
2433 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2435 % LoadOpenCLLibrary() load and binds the OpenCL library.
2437 % The format of the LoadOpenCLLibrary method is:
2439 % MagickBooleanType LoadOpenCLLibrary(void)
2443 void *OsLibraryGetFunctionAddress(void *library,const char *functionName)
2445 if ((library == (void *) NULL) || (functionName == (const char *) NULL))
2446 return (void *) NULL;
2447 #ifdef MAGICKCORE_WINDOWS_SUPPORT
2448 return (void *) GetProcAddress((HMODULE)library,functionName);
2450 return (void *) dlsym(library,functionName);
2454 static MagickBooleanType BindOpenCLFunctions()
2456 #ifdef MAGICKCORE_OPENCL_MACOSX
2457 #define BIND(X) openCL_library->X= &X;
2459 (void) memset(openCL_library,0,sizeof(MagickLibrary));
2460 #ifdef MAGICKCORE_WINDOWS_SUPPORT
2461 openCL_library->library=(void *)LoadLibraryA("OpenCL.dll");
2463 openCL_library->library=(void *)dlopen("libOpenCL.so", RTLD_NOW);
2466 if ((openCL_library->X=(MAGICKpfn_##X)OsLibraryGetFunctionAddress(openCL_library->library,#X)) == NULL) \
2467 return(MagickFalse);
2470 if (openCL_library->library == (void*) NULL)
2471 return(MagickFalse);
2473 BIND(clGetPlatformIDs);
2474 BIND(clGetPlatformInfo);
2476 BIND(clGetDeviceIDs);
2477 BIND(clGetDeviceInfo);
2479 BIND(clCreateBuffer);
2480 BIND(clReleaseMemObject);
2481 BIND(clRetainMemObject);
2483 BIND(clCreateContext);
2484 BIND(clReleaseContext);
2486 BIND(clCreateCommandQueue);
2487 BIND(clReleaseCommandQueue);
2491 BIND(clCreateProgramWithSource);
2492 BIND(clCreateProgramWithBinary);
2493 BIND(clReleaseProgram);
2494 BIND(clBuildProgram);
2495 BIND(clGetProgramBuildInfo);
2496 BIND(clGetProgramInfo);
2498 BIND(clCreateKernel);
2499 BIND(clReleaseKernel);
2500 BIND(clSetKernelArg);
2501 BIND(clGetKernelInfo);
2503 BIND(clEnqueueReadBuffer);
2504 BIND(clEnqueueMapBuffer);
2505 BIND(clEnqueueUnmapMemObject);
2506 BIND(clEnqueueNDRangeKernel);
2508 BIND(clGetEventInfo);
2509 BIND(clWaitForEvents);
2510 BIND(clReleaseEvent);
2511 BIND(clRetainEvent);
2512 BIND(clSetEventCallback);
2514 BIND(clGetEventProfilingInfo);
2519 static MagickBooleanType LoadOpenCLLibrary(void)
2521 openCL_library=(MagickLibrary *) AcquireMagickMemory(sizeof(MagickLibrary));
2522 if (openCL_library == (MagickLibrary *) NULL)
2523 return(MagickFalse);
2525 if (BindOpenCLFunctions() == MagickFalse)
2527 openCL_library=(MagickLibrary *)RelinquishMagickMemory(openCL_library);
2528 return(MagickFalse);
2535 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2539 + O p e n C L T e r m i n u s %
2543 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2545 % OpenCLTerminus() destroys the OpenCL component.
2547 % The format of the OpenCLTerminus method is:
2549 % OpenCLTerminus(void)
2553 MagickPrivate void OpenCLTerminus()
2555 DumpOpenCLProfileData();
2556 if (cache_directory != (char *) NULL)
2557 cache_directory=DestroyString(cache_directory);
2558 if (cache_directory_lock != (SemaphoreInfo *) NULL)
2559 RelinquishSemaphoreInfo(&cache_directory_lock);
2560 if (default_CLEnv != (MagickCLEnv) NULL)
2561 default_CLEnv=RelinquishMagickCLEnv(default_CLEnv);
2562 if (openCL_lock != (SemaphoreInfo *) NULL)
2563 RelinquishSemaphoreInfo(&openCL_lock);
2564 if (openCL_library != (MagickLibrary *) NULL)
2566 if (openCL_library->library != (void *) NULL)
2567 (void) lt_dlclose(openCL_library->library);
2568 openCL_library=(MagickLibrary *) RelinquishMagickMemory(openCL_library);
2573 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2577 + O p e n C L T h r o w M a g i c k E x c e p t i o n %
2581 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2583 % OpenCLThrowMagickException logs an OpenCL exception as determined by the log
2584 % configuration file. If an error occurs, MagickFalse is returned
2585 % otherwise MagickTrue.
2587 % The format of the OpenCLThrowMagickException method is:
2589 % MagickBooleanType OpenCLThrowMagickException(ExceptionInfo *exception,
2590 % const char *module,const char *function,const size_t line,
2591 % const ExceptionType severity,const char *tag,const char *format,...)
2593 % A description of each parameter follows:
2595 % o exception: the exception info.
2597 % o filename: the source module filename.
2599 % o function: the function name.
2601 % o line: the line number of the source module.
2603 % o severity: Specifies the numeric error category.
2605 % o tag: the locale tag.
2607 % o format: the output format.
2611 MagickPrivate MagickBooleanType OpenCLThrowMagickException(
2612 MagickCLDevice device,ExceptionInfo *exception,const char *module,
2613 const char *function,const size_t line,const ExceptionType severity,
2614 const char *tag,const char *format,...)
2619 assert(device != (MagickCLDevice) NULL);
2620 assert(exception != (ExceptionInfo *) NULL);
2621 assert(exception->signature == MagickCoreSignature);
2626 if (device->type == CL_DEVICE_TYPE_CPU)
2628 /* Workaround for Intel OpenCL CPU runtime bug */
2629 /* Turn off OpenCL when a problem is detected! */
2630 if (strncmp(device->platform_name, "Intel",5) == 0)
2631 default_CLEnv->enabled=MagickFalse;
2635 #ifdef OPENCLLOG_ENABLED
2639 va_start(operands,format);
2640 status=ThrowMagickExceptionList(exception,module,function,line,severity,tag,
2645 magick_unreferenced(module);
2646 magick_unreferenced(function);
2647 magick_unreferenced(line);
2648 magick_unreferenced(tag);
2649 magick_unreferenced(format);
2656 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2660 + R e c o r d P r o f i l e D a t a %
2664 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2666 % RecordProfileData() records profile data.
2668 % The format of the RecordProfileData method is:
2670 % void RecordProfileData(MagickCLDevice device,ProfiledKernels kernel,
2673 % A description of each parameter follows:
2675 % o device: the OpenCL device that did the operation.
2677 % o event: the event that contains the profiling data.
2681 MagickPrivate MagickBooleanType RecordProfileData(MagickCLDevice device,
2682 cl_kernel kernel,cl_event event)
2702 if (device->profile_kernels == MagickFalse)
2703 return(MagickFalse);
2704 status=openCL_library->clWaitForEvents(1,&event);
2705 if (status != CL_SUCCESS)
2706 return(MagickFalse);
2707 status=openCL_library->clGetKernelInfo(kernel,CL_KERNEL_FUNCTION_NAME,0,NULL,
2709 if (status != CL_SUCCESS)
2711 name=AcquireQuantumMemory(length,sizeof(*name));
2712 if (name == (char *) NULL)
2714 start=end=elapsed=0;
2715 status=openCL_library->clGetKernelInfo(kernel,CL_KERNEL_FUNCTION_NAME,length,
2716 name,(size_t *) NULL);
2717 status|=openCL_library->clGetEventProfilingInfo(event,
2718 CL_PROFILING_COMMAND_START,sizeof(cl_ulong),&start,NULL);
2719 status|=openCL_library->clGetEventProfilingInfo(event,
2720 CL_PROFILING_COMMAND_END,sizeof(cl_ulong),&end,NULL);
2721 if (status != CL_SUCCESS)
2723 name=DestroyString(name);
2726 start/=1000; /* usecs */
2729 LockSemaphoreInfo(device->lock);
2731 profile_record=(KernelProfileRecord) NULL;
2732 if (device->profile_records != (KernelProfileRecord *) NULL)
2734 while (device->profile_records[i] != (KernelProfileRecord) NULL)
2736 if (LocaleCompare(device->profile_records[i]->kernel_name,name) == 0)
2738 profile_record=device->profile_records[i];
2744 if (profile_record != (KernelProfileRecord) NULL)
2745 name=DestroyString(name);
2748 profile_record=AcquireCriticalMemory(sizeof(*profile_record));
2749 (void) memset(profile_record,0,sizeof(*profile_record));
2750 profile_record->kernel_name=name;
2751 device->profile_records=ResizeMagickMemory(device->profile_records,(i+2)*
2752 sizeof(*device->profile_records));
2753 if (device->profile_records == (KernelProfileRecord *) NULL)
2754 ThrowFatalException(ResourceLimitFatalError,"MemoryAllocationFailed");
2755 device->profile_records[i]=profile_record;
2756 device->profile_records[i+1]=(KernelProfileRecord) NULL;
2758 if ((elapsed < profile_record->min) || (profile_record->count == 0))
2759 profile_record->min=elapsed;
2760 if (elapsed > profile_record->max)
2761 profile_record->max=elapsed;
2762 profile_record->total+=elapsed;
2763 profile_record->count+=1;
2764 UnlockSemaphoreInfo(device->lock);
2769 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2773 + R e l e a s e O p e n C L C o m m a n d Q u e u e %
2777 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2779 % ReleaseOpenCLCommandQueue() releases the OpenCL command queue
2781 % The format of the ReleaseOpenCLCommandQueue method is:
2783 % void ReleaseOpenCLCommandQueue(MagickCLDevice device,
2784 % cl_command_queue queue)
2786 % A description of each parameter follows:
2788 % o device: the OpenCL device.
2790 % o queue: the OpenCL queue to be released.
2793 MagickPrivate void ReleaseOpenCLCommandQueue(MagickCLDevice device,
2794 cl_command_queue queue)
2796 if (queue == (cl_command_queue) NULL)
2799 assert(device != (MagickCLDevice) NULL);
2800 LockSemaphoreInfo(device->lock);
2801 if ((device->profile_kernels != MagickFalse) ||
2802 (device->command_queues_index >= MAGICKCORE_OPENCL_COMMAND_QUEUES-1))
2804 UnlockSemaphoreInfo(device->lock);
2805 openCL_library->clFinish(queue);
2806 (void) openCL_library->clReleaseCommandQueue(queue);
2810 openCL_library->clFlush(queue);
2811 device->command_queues[++device->command_queues_index]=queue;
2812 UnlockSemaphoreInfo(device->lock);
2817 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2821 + R e l e a s e M a g i c k C L D e v i c e %
2825 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2827 % ReleaseOpenCLDevice() returns the OpenCL device to the environment
2829 % The format of the ReleaseOpenCLDevice method is:
2831 % void ReleaseOpenCLDevice(MagickCLDevice device)
2833 % A description of each parameter follows:
2835 % o device: the OpenCL device to be released.
2839 MagickPrivate void ReleaseOpenCLDevice(MagickCLDevice device)
2841 assert(device != (MagickCLDevice) NULL);
2842 LockSemaphoreInfo(openCL_lock);
2843 device->requested--;
2844 UnlockSemaphoreInfo(openCL_lock);
2848 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2852 + R e l i n q u i s h M a g i c k C L C a c h e I n f o %
2856 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2858 % RelinquishMagickCLCacheInfo() frees memory acquired with
2859 % AcquireMagickCLCacheInfo()
2861 % The format of the RelinquishMagickCLCacheInfo method is:
2863 % MagickCLCacheInfo RelinquishMagickCLCacheInfo(MagickCLCacheInfo info,
2864 % const MagickBooleanType relinquish_pixels)
2866 % A description of each parameter follows:
2868 % o info: the OpenCL cache info.
2870 % o relinquish_pixels: the pixels will be relinquish when set to true.
2874 static void CL_API_CALL DestroyMagickCLCacheInfoAndPixels(
2875 cl_event magick_unused(event),
2876 cl_int magick_unused(event_command_exec_status),void *user_data)
2887 magick_unreferenced(event);
2888 magick_unreferenced(event_command_exec_status);
2889 info=(MagickCLCacheInfo) user_data;
2890 for (i=(ssize_t)info->event_count-1; i >= 0; i--)
2898 status=openCL_library->clGetEventInfo(info->events[i],
2899 CL_EVENT_COMMAND_EXECUTION_STATUS,sizeof(event_status),&event_status,
2901 if ((status == CL_SUCCESS) && (event_status != CL_COMPLETE))
2903 openCL_library->clSetEventCallback(info->events[i],CL_COMPLETE,
2904 &DestroyMagickCLCacheInfoAndPixels,info);
2908 pixels=info->pixels;
2909 RelinquishMagickResource(MemoryResource,info->length);
2910 DestroyMagickCLCacheInfo(info);
2911 (void) RelinquishAlignedMemory(pixels);
2914 MagickPrivate MagickCLCacheInfo RelinquishMagickCLCacheInfo(
2915 MagickCLCacheInfo info,const MagickBooleanType relinquish_pixels)
2917 if (info == (MagickCLCacheInfo) NULL)
2918 return((MagickCLCacheInfo) NULL);
2919 if (relinquish_pixels != MagickFalse)
2920 DestroyMagickCLCacheInfoAndPixels((cl_event) NULL,0,info);
2922 DestroyMagickCLCacheInfo(info);
2923 return((MagickCLCacheInfo) NULL);
2927 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2931 % R e l i n q u i s h M a g i c k C L D e v i c e %
2935 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2937 % RelinquishMagickCLDevice() releases the OpenCL device
2939 % The format of the RelinquishMagickCLDevice method is:
2941 % MagickCLDevice RelinquishMagickCLDevice(MagickCLDevice device)
2943 % A description of each parameter follows:
2945 % o device: the OpenCL device to be released.
2949 static MagickCLDevice RelinquishMagickCLDevice(MagickCLDevice device)
2951 if (device == (MagickCLDevice) NULL)
2952 return((MagickCLDevice) NULL);
2954 device->platform_name=RelinquishMagickMemory(device->platform_name);
2955 device->vendor_name=RelinquishMagickMemory(device->vendor_name);
2956 device->name=RelinquishMagickMemory(device->name);
2957 device->version=RelinquishMagickMemory(device->version);
2958 if (device->program != (cl_program) NULL)
2959 (void) openCL_library->clReleaseProgram(device->program);
2960 while (device->command_queues_index >= 0)
2961 (void) openCL_library->clReleaseCommandQueue(
2962 device->command_queues[device->command_queues_index--]);
2963 RelinquishSemaphoreInfo(&device->lock);
2964 return((MagickCLDevice) RelinquishMagickMemory(device));
2968 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2972 % R e l i n q u i s h M a g i c k C L E n v %
2976 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
2978 % RelinquishMagickCLEnv() releases the OpenCL environment
2980 % The format of the RelinquishMagickCLEnv method is:
2982 % MagickCLEnv RelinquishMagickCLEnv(MagickCLEnv device)
2984 % A description of each parameter follows:
2986 % o clEnv: the OpenCL environment to be released.
2990 static MagickCLEnv RelinquishMagickCLEnv(MagickCLEnv clEnv)
2992 if (clEnv == (MagickCLEnv) NULL)
2993 return((MagickCLEnv) NULL);
2995 RelinquishSemaphoreInfo(&clEnv->lock);
2996 RelinquishMagickCLDevices(clEnv);
2997 if (clEnv->contexts != (cl_context *) NULL)
3002 for (i=0; i < clEnv->number_contexts; i++)
3003 if (clEnv->contexts[i] != (cl_context) NULL)
3004 (void) openCL_library->clReleaseContext(clEnv->contexts[i]);
3005 clEnv->contexts=(cl_context *) RelinquishMagickMemory(clEnv->contexts);
3007 return((MagickCLEnv) RelinquishMagickMemory(clEnv));
3011 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3015 + R e q u e s t O p e n C L D e v i c e %
3019 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3021 % RequestOpenCLDevice() returns one of the enabled OpenCL devices.
3023 % The format of the RequestOpenCLDevice method is:
3025 % MagickCLDevice RequestOpenCLDevice(MagickCLEnv clEnv)
3027 % A description of each parameter follows:
3029 % o clEnv: the OpenCL environment.
3032 MagickPrivate MagickCLDevice RequestOpenCLDevice(MagickCLEnv clEnv)
3044 if (clEnv == (MagickCLEnv) NULL)
3045 return((MagickCLDevice) NULL);
3047 if (clEnv->number_devices == 1)
3049 if (clEnv->devices[0]->enabled)
3050 return(clEnv->devices[0]);
3052 return((MagickCLDevice) NULL);
3055 device=(MagickCLDevice) NULL;
3057 LockSemaphoreInfo(openCL_lock);
3058 for (i = 0; i < clEnv->number_devices; i++)
3060 if (clEnv->devices[i]->enabled == MagickFalse)
3063 score=clEnv->devices[i]->score+(clEnv->devices[i]->score*
3064 clEnv->devices[i]->requested);
3065 if ((device == (MagickCLDevice) NULL) || (score < best_score))
3067 device=clEnv->devices[i];
3071 if (device != (MagickCLDevice)NULL)
3072 device->requested++;
3073 UnlockSemaphoreInfo(openCL_lock);
3079 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3083 % S e t O p e n C L D e v i c e E n a b l e d %
3087 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3089 % SetOpenCLDeviceEnabled() can be used to enable or disabled the device.
3091 % The format of the SetOpenCLDeviceEnabled method is:
3093 % void SetOpenCLDeviceEnabled(MagickCLDevice device,
3094 % MagickBooleanType value)
3096 % A description of each parameter follows:
3098 % o device: the OpenCL device.
3100 % o value: determines if the device should be enabled or disabled.
3103 MagickExport void SetOpenCLDeviceEnabled(MagickCLDevice device,
3104 const MagickBooleanType value)
3106 if (device == (MagickCLDevice) NULL)
3108 device->enabled=value;
3112 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3116 % S e t O p e n C L K e r n e l P r o f i l e E n a b l e d %
3120 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3122 % SetOpenCLKernelProfileEnabled() can be used to enable or disabled the
3123 % kernel profiling of a device.
3125 % The format of the SetOpenCLKernelProfileEnabled method is:
3127 % void SetOpenCLKernelProfileEnabled(MagickCLDevice device,
3128 % MagickBooleanType value)
3130 % A description of each parameter follows:
3132 % o device: the OpenCL device.
3134 % o value: determines if kernel profiling for the device should be enabled
3138 MagickExport void SetOpenCLKernelProfileEnabled(MagickCLDevice device,
3139 const MagickBooleanType value)
3141 if (device == (MagickCLDevice) NULL)
3143 device->profile_kernels=value;
3147 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3151 % S e t O p e n C L E n a b l e d %
3155 %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
3157 % SetOpenCLEnabled() can be used to enable or disable OpenCL acceleration.
3159 % The format of the SetOpenCLEnabled method is:
3161 % void SetOpenCLEnabled(MagickBooleanType)
3163 % A description of each parameter follows:
3165 % o value: specify true to enable OpenCL acceleration
3168 MagickExport MagickBooleanType SetOpenCLEnabled(const MagickBooleanType value)
3173 clEnv=GetCurrentOpenCLEnv();
3174 if (clEnv == (MagickCLEnv) NULL)
3175 return(MagickFalse);
3176 clEnv->enabled=value;
3177 return(clEnv->enabled);
3182 MagickExport double GetOpenCLDeviceBenchmarkScore(
3183 const MagickCLDevice magick_unused(device))
3185 magick_unreferenced(device);
3189 MagickExport MagickBooleanType GetOpenCLDeviceEnabled(
3190 const MagickCLDevice magick_unused(device))
3192 magick_unreferenced(device);
3193 return(MagickFalse);
3196 MagickExport const char *GetOpenCLDeviceName(
3197 const MagickCLDevice magick_unused(device))
3199 magick_unreferenced(device);
3200 return((const char *) NULL);
3203 MagickExport MagickCLDevice *GetOpenCLDevices(size_t *length,
3204 ExceptionInfo *magick_unused(exception))
3206 magick_unreferenced(exception);
3207 if (length != (size_t *) NULL)
3209 return((MagickCLDevice *) NULL);
3212 MagickExport MagickCLDeviceType GetOpenCLDeviceType(
3213 const MagickCLDevice magick_unused(device))
3215 magick_unreferenced(device);
3216 return(UndefinedCLDeviceType);
3219 MagickExport const KernelProfileRecord *GetOpenCLKernelProfileRecords(
3220 const MagickCLDevice magick_unused(device),size_t *length)
3222 magick_unreferenced(device);
3223 if (length != (size_t *) NULL)
3225 return((const KernelProfileRecord *) NULL);
3228 MagickExport const char *GetOpenCLDeviceVersion(
3229 const MagickCLDevice magick_unused(device))
3231 magick_unreferenced(device);
3232 return((const char *) NULL);
3235 MagickExport MagickBooleanType GetOpenCLEnabled(void)
3237 return(MagickFalse);
3240 MagickExport void SetOpenCLDeviceEnabled(
3241 MagickCLDevice magick_unused(device),
3242 const MagickBooleanType magick_unused(value))
3244 magick_unreferenced(device);
3245 magick_unreferenced(value);
3248 MagickExport MagickBooleanType SetOpenCLEnabled(
3249 const MagickBooleanType magick_unused(value))
3251 magick_unreferenced(value);
3252 return(MagickFalse);
3255 MagickExport void SetOpenCLKernelProfileEnabled(
3256 MagickCLDevice magick_unused(device),
3257 const MagickBooleanType magick_unused(value))
3259 magick_unreferenced(device);
3260 magick_unreferenced(value);