ViewVC Help
View File | Revision Log | Show Annotations | Download File
/cvs/OpenCL/OpenCL.xs
(Generate patch)

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.64 by root, Thu May 3 23:30:08 2012 UTC vs.
Revision 1.73 by root, Sat May 5 13:30:07 2012 UTC

40 #define PREFER_1_1 1 40 #define PREFER_1_1 1
41#endif 41#endif
42 42
43typedef cl_platform_id OpenCL__Platform; 43typedef cl_platform_id OpenCL__Platform;
44typedef cl_device_id OpenCL__Device; 44typedef cl_device_id OpenCL__Device;
45typedef cl_device_id OpenCL__SubDevice;
45typedef cl_context OpenCL__Context; 46typedef cl_context OpenCL__Context;
46typedef cl_command_queue OpenCL__Queue; 47typedef cl_command_queue OpenCL__Queue;
47typedef cl_mem OpenCL__Memory; 48typedef cl_mem OpenCL__Memory;
48typedef cl_mem OpenCL__Buffer; 49typedef cl_mem OpenCL__Buffer;
49typedef cl_mem OpenCL__BufferObj; 50typedef cl_mem OpenCL__BufferObj;
57typedef cl_event OpenCL__Event; 58typedef cl_event OpenCL__Event;
58typedef cl_event OpenCL__UserEvent; 59typedef cl_event OpenCL__UserEvent;
59 60
60typedef struct mapped * OpenCL__Mapped; 61typedef struct mapped * OpenCL__Mapped;
61 62
62typedef SV *FUTURE;
63
64static HV 63static HV
65 *stash_platform, 64 *stash_platform,
66 *stash_device, 65 *stash_device,
66 *stash_subdevice,
67 *stash_context, 67 *stash_context,
68 *stash_queue, 68 *stash_queue,
69 *stash_program, 69 *stash_program,
70 *stash_kernel, 70 *stash_kernel,
71 *stash_sampler, 71 *stash_sampler,
151 } 151 }
152 152
153 return buf [idx]; 153 return buf [idx];
154} 154}
155 155
156static const char * ecb_noinline
157cv_get_name (CV *cv)
158{
159 static char fullname [256];
160
161 GV *gv = CvGV (cv); // gv better be nonzero
162
163 HV *stash = GvSTASH (gv);
164 const char *hvname = HvNAME_get (stash); // stash also better be nonzero
165 const char *gvname = GvNAME (gv);
166
167 snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname);
168 return fullname;
169}
170
156/*****************************************************************************/ 171/*****************************************************************************/
157 172
158typedef struct 173typedef struct
159{ 174{
160 IV iv; 175 IV iv;
217 if (res) \ 232 if (res) \
218 FAIL (name); 233 FAIL (name);
219 234
220/*****************************************************************************/ 235/*****************************************************************************/
221 236
237static SV *
238new_clobj (HV *stash, IV id)
239{
240 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash));
241}
242
243#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id)))
244#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
245
246/* cl objects are either \$iv, or [$iv, ...] */
247/* they can be upgraded at runtime to the array form */
248static void * ecb_noinline
249SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg)
250{
251 // sv_derived_from is quite slow :(
252 if (SvROK (sv) && sv_derived_from (sv, pkg))
253 return (void *)SvIV (SvRV (sv));
254
255 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg);
256}
257
258// the "no-inherit" version of the above
259static void * ecb_noinline
260SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash)
261{
262 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash)
263 return (void *)SvIV (SvRV (sv));
264
265 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash));
266}
267
268/*****************************************************************************/
269
222static cl_context_properties * ecb_noinline 270static cl_context_properties * ecb_noinline
223SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) 271SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
224{ 272{
225 if (!sv || !SvOK (sv)) 273 if (!sv || !SvOK (sv))
226 if (extra) 274 if (extra)
227 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully 275 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
228 else 276 else
234 int i, len = av_len (av) + 1; 282 int i, len = av_len (av) + 1;
235 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); 283 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
236 cl_context_properties *l = p; 284 cl_context_properties *l = p;
237 285
238 if (len & 1) 286 if (len & 1)
239 croak ("%s: %s is not a property list (must be even number of elements)", func, svname); 287 croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname);
240 288
241 while (extracount--) 289 while (extracount--)
242 *l++ = *extra++; 290 *l++ = *extra++;
243 291
244 for (i = 0; i < len; i += 2) 292 for (i = 0; i < len; i += 2)
247 SV *p_sv = *av_fetch (av, i + 1, 0); 295 SV *p_sv = *av_fetch (av, i + 1, 0);
248 cl_context_properties v = SvIV (p_sv); // code below can override 296 cl_context_properties v = SvIV (p_sv); // code below can override
249 297
250 switch (t) 298 switch (t)
251 { 299 {
300 case CL_CONTEXT_PLATFORM:
301 if (SvROK (p_sv))
302 v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform");
303 break;
304
252 case CL_GLX_DISPLAY_KHR: 305 case CL_GLX_DISPLAY_KHR:
253 if (!SvOK (p_sv)) 306 if (!SvOK (p_sv))
254 { 307 {
255 void *func = glsym ("_glXGetCurrentDisplay"); 308 void *func = glsym ("_glXGetCurrentDisplay");
256 if (func) 309 if (func)
279 *l = 0; 332 *l = 0;
280 333
281 return p; 334 return p;
282 } 335 }
283 336
284 croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); 337 croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname);
285} 338}
286 339
287/*****************************************************************************/ 340// parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation
288 341// is a pointer (and only on well-behaved systems).
289static SV *
290new_clobj (HV *stash, IV id)
291{
292 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash));
293}
294
295#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id)))
296#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
297
298/* cl objects are either \$iv, or [$iv, ...] */
299/* they can be upgraded at runtime to the array form */
300static void * ecb_noinline 342static void * ecb_noinline
301SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) 343object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount)
302{ 344{
303 // sv_derived_from is quite slow :( 345 if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV)
304 if (SvROK (sv) && sv_derived_from (sv, pkg)) 346 croak ("%s: '%s' parameter must be %sa reference to an array of %s objects",
305 return (void *)SvIV (SvRV (sv)); 347 cv_get_name (cv), argname, or_undef ? "undef or " : "", klass);
306 348
307 croak ("%s: %s is not of type %s", func, svname, pkg); 349 AV *av = (AV *)SvRV (arg);
308} 350 void **list = 0;
351 cl_uint count = av_len (av) + 1;
309 352
310// the "no-inherit" version of the above 353 if (count)
311static void * ecb_noinline 354 {
312SvCLOBJ_ni (const char *func, const char *svname, SV *sv, HV *stash) 355 list = tmpbuf (sizeof (*list) * count);
313{ 356 int i;
314 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) 357 for (i = 0; i < count; ++i)
315 return (void *)SvIV (SvRV (sv)); 358 list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass);
359 }
316 360
317 croak ("%s: %s is not of type %s", func, svname, HvNAME (stash)); 361 if (!count && !or_undef)
362 croak ("%s: '%s' must contain at least one %s object",
363 cv_get_name (cv), argname, klass);
364
365 *rcount = count;
366 return (void *)list;
318} 367}
319 368
320/*****************************************************************************/ 369/*****************************************************************************/
321/* callback stuff */ 370/* callback stuff */
322 371
482static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; 531static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
483 532
484static void CL_CALLBACK 533static void CL_CALLBACK
485eq_program_notify (cl_program program, void *user_data) 534eq_program_notify (cl_program program, void *user_data)
486{ 535{
536 clRetainProgram (program);
537
487 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); 538 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
539}
540
541typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data);
542
543static program_callback ecb_noinline
544make_program_callback (SV *notify, void **ruser_data)
545{
546 if (SvOK (notify))
547 {
548 *ruser_data = SvREFCNT_inc (s_get_cv (notify));
549 return eq_program_notify;
550 }
551 else
552 {
553 *ruser_data = 0;
554 return 0;
555 }
488} 556}
489 557
490struct build_args 558struct build_args
491{ 559{
492 cl_program program; 560 cl_program program;
558 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 626 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
559 return res; 627 return res;
560} 628}
561 629
562static cl_event * ecb_noinline 630static cl_event * ecb_noinline
563event_list (SV **items, cl_uint *rcount, cl_event extra) 631event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
564{ 632{
565 cl_uint count = *rcount; 633 cl_uint count = *rcount;
566 634
567 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed 635 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed
568 *rcount = count = 0; 636 *rcount = count = 0;
573 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); 641 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
574 int i = 0; 642 int i = 0;
575 643
576 while (count--) 644 while (count--)
577 if (SvOK (items [count])) 645 if (SvOK (items [count]))
578 list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 646 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
579 647
580 if (extra) 648 if (extra)
581 list [i++] = extra; 649 list [i++] = extra;
582 650
583 *rcount = i; 651 *rcount = i;
585 return i ? list : 0; 653 return i ? list : 0;
586} 654}
587 655
588#define EVENT_LIST(skip) \ 656#define EVENT_LIST(skip) \
589 cl_uint event_list_count = items - (skip); \ 657 cl_uint event_list_count = items - (skip); \
590 cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) 658 cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0)
591 659
592#define INFO(class) \ 660#define INFO(class) \
593{ \ 661{ \
594 size_t size; \ 662 size_t size; \
595 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 663 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
677 745
678 mapped->ptr = 0; 746 mapped->ptr = 0;
679} 747}
680 748
681static void 749static void
682mapped_unmap (SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) 750mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count)
683{ 751{
684 cl_event *event_list_ptr = event_list (wait_list, &event_list_count, mapped->event); 752 cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event);
685 cl_event ev; 753 cl_event ev;
686 754
687 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); 755 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev));
688 756
689 clReleaseEvent (mapped->event); 757 clReleaseEvent (mapped->event);
722 { sizeof (cl_long ), "SIZEOF_LONG" }, 790 { sizeof (cl_long ), "SIZEOF_LONG" },
723 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 791 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
724 { sizeof (cl_half ), "SIZEOF_HALF" }, 792 { sizeof (cl_half ), "SIZEOF_HALF" },
725 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 793 { sizeof (cl_float ), "SIZEOF_FLOAT" },
726 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 794 { sizeof (cl_double), "SIZEOF_DOUBLE" },
795 { PREFER_1_1 , "PREFER_1_1" },
727#include "constiv.h" 796#include "constiv.h"
728 }; 797 };
729 798
730 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 799 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--)
731 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 800 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
732 801
733 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); 802 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
734 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); 803 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD);
804 stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD);
735 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); 805 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
736 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); 806 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD);
737 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); 807 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD);
738 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); 808 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD);
739 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); 809 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD);
780 list = tmpbuf (sizeof (*list) * count); 850 list = tmpbuf (sizeof (*list) * count);
781 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 851 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
782 852
783 EXTEND (SP, count); 853 EXTEND (SP, count);
784 for (i = 0; i < count; ++i) 854 for (i = 0; i < count; ++i)
785 PUSH_CLOBJ (stash_platform, (IV)list [i]); 855 PUSH_CLOBJ (stash_platform, list [i]);
786 856
787void 857void
788context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) 858context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
789 PPCODE: 859 PPCODE:
790 CONTEXT_NOTIFY_CALLBACK; 860 CONTEXT_NOTIFY_CALLBACK;
791 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); 861 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
792 XPUSH_CLOBJ_CONTEXT; 862 XPUSH_CLOBJ_CONTEXT;
793 863
794void 864void
795context (FUTURE properties, FUTURE devices, FUTURE notify) 865context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
796 PPCODE: 866 PPCODE:
797 /* der Gipfel der Kunst */ 867 cl_uint device_count;
868 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
869
870 CONTEXT_NOTIFY_CALLBACK;
871 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res));
872 XPUSH_CLOBJ_CONTEXT;
798 873
799void 874void
800wait_for_events (...) 875wait_for_events (...)
801 CODE: 876 CODE:
802 EVENT_LIST (0); 877 EVENT_LIST (0);
853 EXTEND (SP, count); 928 EXTEND (SP, count);
854 for (i = 0; i < count; ++i) 929 for (i = 0; i < count; ++i)
855 PUSH_CLOBJ (stash_device, list [i]); 930 PUSH_CLOBJ (stash_device, list [i]);
856 931
857void 932void
858context (OpenCL::Platform self, SV *properties = 0, SV *devices, SV *notify = &PL_sv_undef) 933context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
859 PPCODE: 934 PPCODE:
860 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
861 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
862
863 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; 935 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
864 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context", "properties", properties, extra, 2); 936 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
865 937
866 AV *av = (AV *)SvRV (devices); 938 cl_uint device_count;
867 cl_uint num_devices = av_len (av) + 1; 939 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
868 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
869
870 int i;
871 for (i = num_devices; i--; )
872 device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
873 940
874 CONTEXT_NOTIFY_CALLBACK; 941 CONTEXT_NOTIFY_CALLBACK;
875 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, num_devices, device_list, pfn_notify, user_data, &res)); 942 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
876 XPUSH_CLOBJ_CONTEXT; 943 XPUSH_CLOBJ_CONTEXT;
877 944
878void 945void
879context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) 946context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
880 PPCODE: 947 PPCODE:
881 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; 948 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
882 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); 949 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
883 950
884 CONTEXT_NOTIFY_CALLBACK; 951 CONTEXT_NOTIFY_CALLBACK;
885 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); 952 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res));
886 XPUSH_CLOBJ_CONTEXT; 953 XPUSH_CLOBJ_CONTEXT;
887 954
889 956
890void 957void
891info (OpenCL::Device self, cl_device_info name) 958info (OpenCL::Device self, cl_device_info name)
892 PPCODE: 959 PPCODE:
893 INFO (Device) 960 INFO (Device)
961
962#if CL_VERSION_1_2
963
964void
965sub_devices (OpenCL::Device self, SV *properties)
966 PPCODE:
967 if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV)
968 croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs");
969
970 properties = SvRV (properties);
971
972 cl_uint count = av_len ((AV *)properties) + 1;
973 cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1);
974
975 int i;
976 for (i = 0; i < count; ++i)
977 props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0));
978
979 props [count] = 0;
980
981 cl_uint num_devices;
982 NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices));
983 cl_device_id *list = tmpbuf (sizeof (*list) * num_devices);
984 NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0));
985
986 EXTEND (SP, num_devices);
987 for (i = 0; i < count; ++i)
988 PUSH_CLOBJ (stash_subdevice, list [i]);
989
990#endif
894 991
895#BEGIN:device 992#BEGIN:device
896 993
897void 994void
898type (OpenCL::Device self) 995type (OpenCL::Device self)
1114 for (i = 0; i < n; ++i) 1211 for (i = 0; i < n; ++i)
1115 PUSHs (sv_2mortal (newSVuv (value [i]))); 1212 PUSHs (sv_2mortal (newSVuv (value [i])));
1116 1213
1117#END:device 1214#END:device
1118 1215
1216MODULE = OpenCL PACKAGE = OpenCL::SubDevice
1217
1218#if CL_VERSION_1_2
1219
1220void
1221DESTROY (OpenCL::SubDevice self)
1222 CODE:
1223 clReleaseDevice (self);
1224
1225#endif
1226
1119MODULE = OpenCL PACKAGE = OpenCL::Context 1227MODULE = OpenCL PACKAGE = OpenCL::Context
1120 1228
1121void 1229void
1122DESTROY (OpenCL::Context context) 1230DESTROY (OpenCL::Context self)
1123 CODE: 1231 CODE:
1124 clReleaseContext (context); 1232 clReleaseContext (self);
1125 1233
1126void 1234void
1127info (OpenCL::Context self, cl_context_info name) 1235info (OpenCL::Context self, cl_context_info name)
1128 PPCODE: 1236 PPCODE:
1129 INFO (Context) 1237 INFO (Context)
1170 const cl_image_desc desc = { 1278 const cl_image_desc desc = {
1171 type, 1279 type,
1172 width, height, depth, 1280 width, height, depth,
1173 array_size, row_pitch, slice_pitch, 1281 array_size, row_pitch, slice_pitch,
1174 num_mip_level, num_samples, 1282 num_mip_level, num_samples,
1175 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 1283 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0
1176 }; 1284 };
1177 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); 1285 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1178 HV *stash = stash_image; 1286 HV *stash = stash_image;
1179 switch (type) 1287 switch (type)
1180 { 1288 {
1316 XPUSH_CLOBJ (stash_program, prog); 1424 XPUSH_CLOBJ (stash_program, prog);
1317 1425
1318void 1426void
1319program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) 1427program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1320 PPCODE: 1428 PPCODE:
1321 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 1429 cl_uint device_count;
1322 croak ("OpenCL::Context::program_with_binary: devices must be specified as reference to an array of device objects"); 1430 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1323
1324 devices = SvRV (devices);
1325 1431
1326 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) 1432 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1327 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); 1433 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1328 1434
1329 binaries = SvRV (binaries); 1435 binaries = SvRV (binaries);
1330 1436
1331 if (av_len ((AV *)devices) != av_len ((AV *)binaries)) 1437 if (device_count != av_len ((AV *)binaries) + 1)
1332 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); 1438 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1333 1439
1334 int count = av_len ((AV *)devices) + 1;
1335 cl_device_id *device_list = tmpbuf (sizeof (*device_list) * count);
1336 size_t *length_list = tmpbuf (sizeof (*length_list) * count); 1440 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1337 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * count); 1441 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1338 cl_int *status_list = tmpbuf (sizeof (*status_list) * count); 1442 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1339 1443
1340 int i; 1444 int i;
1341 for (i = 0; i < count; ++i) 1445 for (i = 0; i < device_count; ++i)
1342 { 1446 {
1343 device_list [i] = SvCLOBJ ("OpenCL::Context::program_with_binary", "devices", *av_fetch ((AV *)devices, i, 0), "OpenCL::Device");
1344 STRLEN len; 1447 STRLEN len;
1345 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); 1448 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1346 length_list [i] = len; 1449 length_list [i] = len;
1347 } 1450 }
1348 1451
1349 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, count, device_list, length_list, binary_list, GIMME_V == G_ARRAY ? status_list : 0, &res)); 1452 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list,
1453 length_list, binary_list,
1454 GIMME_V == G_ARRAY ? status_list : 0, &res));
1350 1455
1351 EXTEND (SP, 2); 1456 EXTEND (SP, 2);
1352 PUSH_CLOBJ (stash_program, prog); 1457 PUSH_CLOBJ (stash_program, prog);
1353 1458
1354 if (GIMME_V == G_ARRAY) 1459 if (GIMME_V == G_ARRAY)
1355 { 1460 {
1356 AV *av = newAV (); 1461 AV *av = newAV ();
1357 PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); 1462 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1358 1463
1359 for (i = count; i--; ) 1464 for (i = device_count; i--; )
1360 av_store (av, i, newSViv (status_list [i])); 1465 av_store (av, i, newSViv (status_list [i]));
1361 } 1466 }
1467
1468#if CL_VERSION_1_2
1469
1470void
1471program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names)
1472 PPCODE:
1473 cl_uint device_count;
1474 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1475
1476 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res));
1477
1478 XPUSH_CLOBJ (stash_program, prog);
1479
1480void
1481link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef)
1482 CODE:
1483 cl_uint device_count = 0;
1484 cl_device_id *device_list = 0;
1485
1486 if (SvOK (devices))
1487 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
1488
1489 cl_uint program_count;
1490 cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count);
1491
1492 void *user_data;
1493 program_callback pfn_notify = make_program_callback (notify, &user_data);
1494
1495 NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
1496 program_count, program_list, pfn_notify, user_data, &res));
1497
1498 XPUSH_CLOBJ (stash_program, prog);
1499
1500#endif
1362 1501
1363#BEGIN:context 1502#BEGIN:context
1364 1503
1365void 1504void
1366reference_count (OpenCL::Context self) 1505reference_count (OpenCL::Context self)
1412void 1551void
1413read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) 1552read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1414 ALIAS: 1553 ALIAS:
1415 enqueue_read_buffer = 0 1554 enqueue_read_buffer = 0
1416 PPCODE: 1555 PPCODE:
1417 cl_event ev = 0;
1418 EVENT_LIST (6); 1556 EVENT_LIST (6);
1419 1557
1420 SvUPGRADE (data, SVt_PV); 1558 SvUPGRADE (data, SVt_PV);
1421 SvGROW (data, len); 1559 SvGROW (data, len);
1422 SvPOK_only (data); 1560 SvPOK_only (data);
1423 SvCUR_set (data, len); 1561 SvCUR_set (data, len);
1562
1563 cl_event ev = 0;
1424 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1564 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1425 1565
1426 if (ev) 1566 if (ev)
1427 XPUSH_CLOBJ (stash_event, ev); 1567 XPUSH_CLOBJ (stash_event, ev);
1428 1568
1429void 1569void
1430write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) 1570write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1431 ALIAS: 1571 ALIAS:
1432 enqueue_write_buffer = 0 1572 enqueue_write_buffer = 0
1433 PPCODE: 1573 PPCODE:
1434 cl_event ev = 0; 1574 EVENT_LIST (5);
1575
1435 STRLEN len; 1576 STRLEN len;
1436 char *ptr = SvPVbyte (data, len); 1577 char *ptr = SvPVbyte (data, len);
1437 EVENT_LIST (5);
1438 1578
1579 cl_event ev = 0;
1439 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1580 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1440 1581
1441 if (ev) 1582 if (ev)
1442 XPUSH_CLOBJ (stash_event, ev); 1583 XPUSH_CLOBJ (stash_event, ev);
1443 1584
1446void 1587void
1447fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) 1588fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1448 ALIAS: 1589 ALIAS:
1449 enqueue_fill_buffer = 0 1590 enqueue_fill_buffer = 0
1450 PPCODE: 1591 PPCODE:
1451 cl_event ev = 0; 1592 EVENT_LIST (5);
1593
1452 STRLEN len; 1594 STRLEN len;
1453 char *ptr = SvPVbyte (data, len); 1595 char *ptr = SvPVbyte (data, len);
1454 EVENT_LIST (5);
1455 1596
1597 cl_event ev = 0;
1456 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1598 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1457 1599
1458 if (ev) 1600 if (ev)
1459 XPUSH_CLOBJ (stash_event, ev); 1601 XPUSH_CLOBJ (stash_event, ev);
1460 1602
1461void 1603void
1462fill_image (OpenCL::Queue self, OpenCL::Image img, NV r, NV g, NV b, NV a, size_t x, size_t y, size_t z, size_t width, size_t height, size_t depth, ...) 1604fill_image (OpenCL::Queue self, OpenCL::Image img, NV r, NV g, NV b, NV a, size_t x, size_t y, size_t z, size_t width, size_t height, size_t depth, ...)
1463 ALIAS: 1605 ALIAS:
1464 enqueue_fill_image = 0 1606 enqueue_fill_image = 0
1465 PPCODE: 1607 PPCODE:
1466 cl_event ev = 0; 1608 EVENT_LIST (12);
1609
1467 const size_t origin [3] = { x, y, z }; 1610 const size_t origin [3] = { x, y, z };
1468 const size_t region [3] = { width, height, depth }; 1611 const size_t region [3] = { width, height, depth };
1469 EVENT_LIST (12);
1470 1612
1471 const cl_float c_f [4] = { r, g, b, a }; 1613 const cl_float c_f [4] = { r, g, b, a };
1472 const cl_uint c_u [4] = { r, g, b, a }; 1614 const cl_uint c_u [4] = { r, g, b, a };
1473 const cl_int c_s [4] = { r, g, b, a }; 1615 const cl_int c_s [4] = { r, g, b, a };
1474 const void *c_fus [3] = { &c_f, &c_u, &c_s }; 1616 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1477 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); 1619 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1478 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); 1620 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1479 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) 1621 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1480 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); 1622 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1481 1623
1624 cl_event ev = 0;
1482 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], 1625 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1483 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1626 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1484 1627
1485 if (ev) 1628 if (ev)
1486 XPUSH_CLOBJ (stash_event, ev); 1629 XPUSH_CLOBJ (stash_event, ev);
1490void 1633void
1491copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1634copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1492 ALIAS: 1635 ALIAS:
1493 enqueue_copy_buffer = 0 1636 enqueue_copy_buffer = 0
1494 PPCODE: 1637 PPCODE:
1638 EVENT_LIST (6);
1639
1495 cl_event ev = 0; 1640 cl_event ev = 0;
1496 EVENT_LIST (6);
1497
1498 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1641 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1499 1642
1500 if (ev) 1643 if (ev)
1501 XPUSH_CLOBJ (stash_event, ev); 1644 XPUSH_CLOBJ (stash_event, ev);
1502 1645
1503void 1646void
1504read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) 1647read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...)
1505 ALIAS: 1648 ALIAS:
1506 enqueue_read_buffer_rect = 0 1649 enqueue_read_buffer_rect = 0
1507 PPCODE: 1650 PPCODE:
1508 cl_event ev = 0; 1651 EVENT_LIST (17);
1652
1509 const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; 1653 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1510 const size_t host_origin[3] = { host_x, host_y, host_z }; 1654 const size_t host_origin[3] = { host_x, host_y, host_z };
1511 const size_t region[3] = { width, height, depth }; 1655 const size_t region[3] = { width, height, depth };
1512 EVENT_LIST (17);
1513 1656
1514 if (!buf_row_pitch) 1657 if (!buf_row_pitch)
1515 buf_row_pitch = region [0]; 1658 buf_row_pitch = region [0];
1516 1659
1517 if (!buf_slice_pitch) 1660 if (!buf_slice_pitch)
1527 1670
1528 SvUPGRADE (data, SVt_PV); 1671 SvUPGRADE (data, SVt_PV);
1529 SvGROW (data, len); 1672 SvGROW (data, len);
1530 SvPOK_only (data); 1673 SvPOK_only (data);
1531 SvCUR_set (data, len); 1674 SvCUR_set (data, len);
1675
1676 cl_event ev = 0;
1532 NEED_SUCCESS (EnqueueReadBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1677 NEED_SUCCESS (EnqueueReadBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1533 1678
1534 if (ev) 1679 if (ev)
1535 XPUSH_CLOBJ (stash_event, ev); 1680 XPUSH_CLOBJ (stash_event, ev);
1536 1681
1537void 1682void
1538write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) 1683write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...)
1539 ALIAS: 1684 ALIAS:
1540 enqueue_write_buffer_rect = 0 1685 enqueue_write_buffer_rect = 0
1541 PPCODE: 1686 PPCODE:
1542 cl_event ev = 0; 1687 EVENT_LIST (17);
1688
1543 const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; 1689 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1544 const size_t host_origin[3] = { host_x, host_y, host_z }; 1690 const size_t host_origin[3] = { host_x, host_y, host_z };
1545 const size_t region[3] = { width, height, depth }; 1691 const size_t region[3] = { width, height, depth };
1546 STRLEN len; 1692 STRLEN len;
1547 char *ptr = SvPVbyte (data, len); 1693 char *ptr = SvPVbyte (data, len);
1548 EVENT_LIST (17);
1549 1694
1550 if (!buf_row_pitch) 1695 if (!buf_row_pitch)
1551 buf_row_pitch = region [0]; 1696 buf_row_pitch = region [0];
1552 1697
1553 if (!buf_slice_pitch) 1698 if (!buf_slice_pitch)
1562 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 1707 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
1563 1708
1564 if (len < min_len) 1709 if (len < min_len)
1565 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1710 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1566 1711
1712 cl_event ev = 0;
1567 NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1713 NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1568 1714
1569 if (ev) 1715 if (ev)
1570 XPUSH_CLOBJ (stash_event, ev); 1716 XPUSH_CLOBJ (stash_event, ev);
1571 1717
1572void 1718void
1573copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) 1719copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...)
1574 ALIAS: 1720 ALIAS:
1575 enqueue_copy_buffer_rect = 0 1721 enqueue_copy_buffer_rect = 0
1576 PPCODE: 1722 PPCODE:
1577 cl_event ev = 0; 1723 EVENT_LIST (16);
1724
1578 const size_t src_origin[3] = { src_x, src_y, src_z }; 1725 const size_t src_origin[3] = { src_x, src_y, src_z };
1579 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1726 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1580 const size_t region[3] = { width, height, depth }; 1727 const size_t region[3] = { width, height, depth };
1581 EVENT_LIST (16);
1582 1728
1729 cl_event ev = 0;
1583 NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1730 NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1584 1731
1585 if (ev) 1732 if (ev)
1586 XPUSH_CLOBJ (stash_event, ev); 1733 XPUSH_CLOBJ (stash_event, ev);
1587 1734
1588void 1735void
1589read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) 1736read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...)
1590 ALIAS: 1737 ALIAS:
1591 enqueue_read_image = 0 1738 enqueue_read_image = 0
1592 PPCODE: 1739 PPCODE:
1593 cl_event ev = 0; 1740 EVENT_LIST (12);
1741
1594 const size_t src_origin[3] = { src_x, src_y, src_z }; 1742 const size_t src_origin[3] = { src_x, src_y, src_z };
1595 const size_t region[3] = { width, height, depth }; 1743 const size_t region[3] = { width, height, depth };
1596 EVENT_LIST (12);
1597 1744
1598 if (!row_pitch) 1745 if (!row_pitch)
1599 row_pitch = img_row_pitch (src); 1746 row_pitch = img_row_pitch (src);
1600 1747
1601 if (depth > 1 && !slice_pitch) 1748 if (depth > 1 && !slice_pitch)
1605 1752
1606 SvUPGRADE (data, SVt_PV); 1753 SvUPGRADE (data, SVt_PV);
1607 SvGROW (data, len); 1754 SvGROW (data, len);
1608 SvPOK_only (data); 1755 SvPOK_only (data);
1609 SvCUR_set (data, len); 1756 SvCUR_set (data, len);
1757
1758 cl_event ev = 0;
1610 NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1759 NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1611 1760
1612 if (ev) 1761 if (ev)
1613 XPUSH_CLOBJ (stash_event, ev); 1762 XPUSH_CLOBJ (stash_event, ev);
1614 1763
1615void 1764void
1616write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) 1765write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...)
1617 ALIAS: 1766 ALIAS:
1618 enqueue_write_image = 0 1767 enqueue_write_image = 0
1619 PPCODE: 1768 PPCODE:
1620 cl_event ev = 0; 1769 EVENT_LIST (12);
1770
1621 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1771 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1622 const size_t region[3] = { width, height, depth }; 1772 const size_t region[3] = { width, height, depth };
1623 STRLEN len; 1773 STRLEN len;
1624 char *ptr = SvPVbyte (data, len); 1774 char *ptr = SvPVbyte (data, len);
1625 EVENT_LIST (12);
1626 1775
1627 if (!row_pitch) 1776 if (!row_pitch)
1628 row_pitch = img_row_pitch (dst); 1777 row_pitch = img_row_pitch (dst);
1629 1778
1630 if (depth > 1 && !slice_pitch) 1779 if (depth > 1 && !slice_pitch)
1633 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1782 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1634 1783
1635 if (len < min_len) 1784 if (len < min_len)
1636 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1785 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1637 1786
1787 cl_event ev = 0;
1638 NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1788 NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1639 1789
1640 if (ev) 1790 if (ev)
1641 XPUSH_CLOBJ (stash_event, ev); 1791 XPUSH_CLOBJ (stash_event, ev);
1642 1792
1643void 1793void
1644copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) 1794copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...)
1645 ALIAS: 1795 ALIAS:
1646 enqueue_copy_image = 0 1796 enqueue_copy_image = 0
1647 PPCODE: 1797 PPCODE:
1648 cl_event ev = 0; 1798 EVENT_LIST (12);
1799
1649 const size_t src_origin[3] = { src_x, src_y, src_z }; 1800 const size_t src_origin[3] = { src_x, src_y, src_z };
1650 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1801 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1651 const size_t region[3] = { width, height, depth }; 1802 const size_t region[3] = { width, height, depth };
1652 EVENT_LIST (12);
1653 1803
1804 cl_event ev = 0;
1654 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1805 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1655 1806
1656 if (ev) 1807 if (ev)
1657 XPUSH_CLOBJ (stash_event, ev); 1808 XPUSH_CLOBJ (stash_event, ev);
1658 1809
1659void 1810void
1660copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) 1811copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...)
1661 ALIAS: 1812 ALIAS:
1662 enqueue_copy_image_to_buffer = 0 1813 enqueue_copy_image_to_buffer = 0
1663 PPCODE: 1814 PPCODE:
1664 cl_event ev = 0; 1815 EVENT_LIST (10);
1816
1665 const size_t src_origin[3] = { src_x, src_y, src_z }; 1817 const size_t src_origin[3] = { src_x, src_y, src_z };
1666 const size_t region [3] = { width, height, depth }; 1818 const size_t region [3] = { width, height, depth };
1667 EVENT_LIST (10);
1668 1819
1820 cl_event ev = 0;
1669 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1821 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1670 1822
1671 if (ev) 1823 if (ev)
1672 XPUSH_CLOBJ (stash_event, ev); 1824 XPUSH_CLOBJ (stash_event, ev);
1673 1825
1674void 1826void
1675copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) 1827copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...)
1676 ALIAS: 1828 ALIAS:
1677 enqueue_copy_buffer_to_image = 0 1829 enqueue_copy_buffer_to_image = 0
1678 PPCODE: 1830 PPCODE:
1679 cl_event ev = 0; 1831 EVENT_LIST (10);
1832
1680 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1833 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1681 const size_t region [3] = { width, height, depth }; 1834 const size_t region [3] = { width, height, depth };
1682 EVENT_LIST (10);
1683 1835
1836 cl_event ev = 0;
1684 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1837 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1685 1838
1686 if (ev) 1839 if (ev)
1687 XPUSH_CLOBJ (stash_event, ev); 1840 XPUSH_CLOBJ (stash_event, ev);
1688 1841
1689void 1842void
1690map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, SV *cb_ = &PL_sv_undef, ...) 1843map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, SV *cb_ = &PL_sv_undef, ...)
1691 ALIAS: 1844 ALIAS:
1692 enqueue_map_buffer = 0 1845 enqueue_map_buffer = 0
1693 PPCODE: 1846 PPCODE:
1694 cl_event ev;
1695 EVENT_LIST (6); 1847 EVENT_LIST (6);
1848
1696 size_t cb = SvIV (cb_); 1849 size_t cb = SvIV (cb_);
1697 1850
1698 if (!SvOK (cb_)) 1851 if (!SvOK (cb_))
1699 { 1852 {
1700 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); 1853 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1701 cb -= offset; 1854 cb -= offset;
1702 } 1855 }
1703 1856
1857 cl_event ev;
1704 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); 1858 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
1705 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); 1859 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0));
1706 1860
1707void 1861void
1708map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...) 1862map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...)
1749 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); 1903 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch));
1750 1904
1751void 1905void
1752unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) 1906unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1753 PPCODE: 1907 PPCODE:
1754 mapped_unmap (ST (1), mapped, self, &ST (2), items - 2); 1908 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1755 if (GIMME_V != G_VOID) 1909 if (GIMME_V != G_VOID)
1756 { 1910 {
1757 clRetainEvent (mapped->event); 1911 clRetainEvent (mapped->event);
1758 XPUSH_CLOBJ (stash_event, mapped->event); 1912 XPUSH_CLOBJ (stash_event, mapped->event);
1759 } 1913 }
1761void 1915void
1762task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) 1916task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1763 ALIAS: 1917 ALIAS:
1764 enqueue_task = 0 1918 enqueue_task = 0
1765 PPCODE: 1919 PPCODE:
1920 EVENT_LIST (2);
1921
1766 cl_event ev = 0; 1922 cl_event ev = 0;
1767 EVENT_LIST (2);
1768
1769 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1923 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1770 1924
1771 if (ev) 1925 if (ev)
1772 XPUSH_CLOBJ (stash_event, ev); 1926 XPUSH_CLOBJ (stash_event, ev);
1773 1927
1774void 1928void
1775nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) 1929nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1776 ALIAS: 1930 ALIAS:
1777 enqueue_nd_range_kernel = 0 1931 enqueue_nd_range_kernel = 0
1778 PPCODE: 1932 PPCODE:
1779 cl_event ev = 0; 1933 EVENT_LIST (5);
1934
1780 size_t *gwo = 0, *gws, *lws = 0; 1935 size_t *gwo = 0, *gws, *lws = 0;
1781 int gws_len; 1936 int gws_len;
1782 size_t *lists; 1937 size_t *lists;
1783 int i; 1938 int i;
1784 EVENT_LIST (5);
1785 1939
1786 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) 1940 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
1787 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1941 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
1788 1942
1789 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1943 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
1828 if (!lws [i]) 1982 if (!lws [i])
1829 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); 1983 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
1830 } 1984 }
1831 } 1985 }
1832 1986
1987 cl_event ev = 0;
1833 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1988 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1834 1989
1835 if (ev) 1990 if (ev)
1836 XPUSH_CLOBJ (stash_event, ev); 1991 XPUSH_CLOBJ (stash_event, ev);
1992
1993#if CL_VERSION_1_2
1994
1995void
1996migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...)
1997 ALIAS:
1998 enqueue_migrate_mem_objects = 0
1999 PPCODE:
2000 EVENT_LIST (3);
2001
2002 cl_uint obj_count;
2003 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2004
2005 cl_event ev = 0;
2006 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2007
2008 if (ev)
2009 XPUSH_CLOBJ (stash_event, ev);
2010
2011#endif
1837 2012
1838#if cl_apple_gl_sharing || cl_khr_gl_sharing 2013#if cl_apple_gl_sharing || cl_khr_gl_sharing
1839 2014
1840void 2015void
1841acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) 2016acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1842 ALIAS: 2017 ALIAS:
1843 release_gl_objects = 1 2018 release_gl_objects = 1
1844 enqueue_acquire_gl_objects = 0 2019 enqueue_acquire_gl_objects = 0
1845 enqueue_release_gl_objects = 1 2020 enqueue_release_gl_objects = 1
1846 PPCODE: 2021 PPCODE:
1847 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 2022 EVENT_LIST (2);
1848 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); 2023
2024 cl_uint obj_count;
2025 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
1849 2026
1850 cl_event ev = 0; 2027 cl_event ev = 0;
1851 EVENT_LIST (2);
1852 AV *av = (AV *)SvRV (objects);
1853 cl_uint num_objects = av_len (av) + 1;
1854 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1855 int i;
1856
1857 for (i = num_objects; i--; )
1858 object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1859 2028
1860 if (ix) 2029 if (ix)
1861 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2030 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1862 else 2031 else
1863 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2032 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1864 2033
1865 if (ev) 2034 if (ev)
1866 XPUSH_CLOBJ (stash_event, ev); 2035 XPUSH_CLOBJ (stash_event, ev);
1867 2036
1868#endif 2037#endif
1882void 2051void
1883marker (OpenCL::Queue self, ...) 2052marker (OpenCL::Queue self, ...)
1884 ALIAS: 2053 ALIAS:
1885 enqueue_marker = 0 2054 enqueue_marker = 0
1886 PPCODE: 2055 PPCODE:
2056 EVENT_LIST (1);
1887 cl_event ev = 0; 2057 cl_event ev = 0;
1888 EVENT_LIST (1);
1889#if PREFER_1_1 2058#if PREFER_1_1
1890 if (!event_list_count) 2059 if (!event_list_count)
1891 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); 2060 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1892 else 2061 else
1893#if CL_VERSION_1_2 2062#if CL_VERSION_1_2
1907void 2076void
1908barrier (OpenCL::Queue self, ...) 2077barrier (OpenCL::Queue self, ...)
1909 ALIAS: 2078 ALIAS:
1910 enqueue_barrier = 0 2079 enqueue_barrier = 0
1911 PPCODE: 2080 PPCODE:
2081 EVENT_LIST (1);
1912 cl_event ev = 0; 2082 cl_event ev = 0;
1913 EVENT_LIST (1);
1914#if PREFER_1_1 2083#if PREFER_1_1
1915 if (!event_list_count && GIMME_V == G_VOID) 2084 if (!event_list_count && GIMME_V == G_VOID)
1916 NEED_SUCCESS (EnqueueBarrier, (self)); 2085 NEED_SUCCESS (EnqueueBarrier, (self));
1917 else 2086 else
1918#if CL_VERSION_1_2 2087#if CL_VERSION_1_2
2246void 2415void
2247build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) 2416build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2248 ALIAS: 2417 ALIAS:
2249 build_async = 1 2418 build_async = 1
2250 CODE: 2419 CODE:
2251 void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0;
2252 void *user_data = 0;
2253 cl_uint num_devices = 0; 2420 cl_uint device_count = 0;
2254 cl_device_id *device_list = 0; 2421 cl_device_id *device_list = 0;
2255 2422
2256 if (SvOK (devices)) 2423 if (SvOK (devices))
2257 { 2424 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2425
2426 void *user_data;
2427 program_callback pfn_notify = make_program_callback (notify, &user_data);
2428
2429 if (ix)
2430 build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data);
2431 else
2432 NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
2433
2434#if CL_VERSION_1_2
2435
2436void
2437compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef)
2438 CODE:
2439 cl_uint device_count = 0;
2440 cl_device_id *device_list = 0;
2441
2442 if (SvOK (devices))
2443 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2444
2445 cl_uint header_count = 0;
2446 cl_program *header_list = 0;
2447 const char **header_name = 0;
2448
2449 if (SvOK (headers))
2450 {
2258 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 2451 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2259 croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); 2452 croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs");
2260 2453
2261 AV *av = (AV *)SvRV (devices); 2454 HV *hv = (HV *)SvRV (devices);
2262 num_devices = av_len (av) + 1;
2263 2455
2264 if (num_devices) 2456 header_count = hv_iterinit (hv);
2457 header_list = tmpbuf (sizeof (*header_list) * header_count);
2458 header_name = tmpbuf (sizeof (*header_name) * header_count);
2459
2460 HE *he;
2461 int i = 0;
2462 while (he = hv_iternext (hv))
2265 { 2463 {
2266 device_list = tmpbuf (sizeof (*device_list) * num_devices); 2464 header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he));
2465 header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program");
2267 int count; 2466 ++i;
2268 for (count = 0; count < num_devices; ++count)
2269 device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device");
2270 } 2467 }
2271 } 2468 }
2272 2469
2273 if (SvOK (notify)) 2470 void *user_data;
2274 { 2471 program_callback pfn_notify = make_program_callback (notify, &user_data);
2275 NEED_SUCCESS (RetainProgram, (self));
2276 pfn_notify = eq_program_notify;
2277 user_data = SvREFCNT_inc (s_get_cv (notify));
2278 }
2279 2472
2280 if (ix)
2281 build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data);
2282 else
2283 NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); 2473 NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
2474 header_count, header_list, header_name, pfn_notify, user_data));
2475
2476#endif
2284 2477
2285void 2478void
2286build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) 2479build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
2287 PPCODE: 2480 PPCODE:
2288 size_t size; 2481 size_t size;
2316 char *value = tmpbuf (size); 2509 char *value = tmpbuf (size);
2317 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); 2510 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2318 EXTEND (SP, 1); 2511 EXTEND (SP, 1);
2319 const int i = 0; 2512 const int i = 0;
2320 PUSHs (sv_2mortal (newSVpv (value, 0))); 2513 PUSHs (sv_2mortal (newSVpv (value, 0)));
2514
2515void
2516binary_type (OpenCL::Program self, OpenCL::Device device)
2517 PPCODE:
2518 cl_program_binary_type value [1];
2519 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0));
2520 EXTEND (SP, 1);
2521 const int i = 0;
2522 PUSHs (sv_2mortal (newSVuv (UV)value [i])));
2321 2523
2322#END:program_build 2524#END:program_build
2323 2525
2324void 2526void
2325kernel (OpenCL::Program program, SV *function) 2527kernel (OpenCL::Program program, SV *function)
2492 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; 2694 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2493 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; 2695 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2494 2696
2495 case 'z': nullarg = 1; size = SvIV (sv); break; 2697 case 'z': nullarg = 1; size = SvIV (sv); break;
2496 2698
2497 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; 2699 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2498 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; 2700 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2499 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; 2701 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2500 2702
2501 default: 2703 default:
2502 croak ("OpenCL::Kernel::setf format character '%c' not supported", type); 2704 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2503 } 2705 }
2504 2706
2704 const int i = 0; 2906 const int i = 0;
2705 PUSHs (sv_2mortal (newSVuv (value [i]))); 2907 PUSHs (sv_2mortal (newSVuv (value [i])));
2706 2908
2707#END:kernel_work_group 2909#END:kernel_work_group
2708 2910
2911#if CL_VERSION_1_2
2912
2913void
2914arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name)
2915 PPCODE:
2916 size_t size;
2917 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size));
2918 SV *sv = sv_2mortal (newSV (size));
2919 SvUPGRADE (sv, SVt_PV);
2920 SvPOK_only (sv);
2921 SvCUR_set (sv, size);
2922 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0));
2923 XPUSHs (sv);
2924
2925#BEGIN:kernel_arg
2926
2927void
2928arg_address_qualifier (OpenCL::Kernel self, cl_uint idx)
2929 PPCODE:
2930 cl_kernel_arg_address_qualifier value [1];
2931 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0));
2932 EXTEND (SP, 1);
2933 const int i = 0;
2934 PUSHs (sv_2mortal (newSVuv (value [i])));
2935
2936void
2937arg_access_qualifier (OpenCL::Kernel self, cl_uint idx)
2938 PPCODE:
2939 cl_kernel_arg_access_qualifier value [1];
2940 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0));
2941 EXTEND (SP, 1);
2942 const int i = 0;
2943 PUSHs (sv_2mortal (newSVuv (value [i])));
2944
2945void
2946arg_type_name (OpenCL::Kernel self, cl_uint idx)
2947 ALIAS:
2948 arg_type_name = CL_KERNEL_ARG_TYPE_NAME
2949 arg_name = CL_KERNEL_ARG_NAME
2950 PPCODE:
2951 size_t size;
2952 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size));
2953 char *value = tmpbuf (size);
2954 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0));
2955 EXTEND (SP, 1);
2956 const int i = 0;
2957 PUSHs (sv_2mortal (newSVpv (value, 0)));
2958
2959void
2960arg_type_qualifier (OpenCL::Kernel self, cl_uint idx)
2961 PPCODE:
2962 cl_kernel_arg_type_qualifier value [1];
2963 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0));
2964 EXTEND (SP, 1);
2965 const int i = 0;
2966 PUSHs (sv_2mortal (newSVuv (value [i])));
2967
2968#END:kernel_arg
2969
2970#endif
2971
2709MODULE = OpenCL PACKAGE = OpenCL::Event 2972MODULE = OpenCL PACKAGE = OpenCL::Event
2710 2973
2711void 2974void
2712DESTROY (OpenCL::Event self) 2975DESTROY (OpenCL::Event self)
2713 CODE: 2976 CODE:
2822 Safefree (mapped); 3085 Safefree (mapped);
2823 3086
2824void 3087void
2825unmap (OpenCL::Mapped self, ...) 3088unmap (OpenCL::Mapped self, ...)
2826 CODE: 3089 CODE:
2827 mapped_unmap (ST (0), self, self->queue, &ST (1), items - 1); 3090 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
2828 3091
2829bool 3092bool
2830mapped (OpenCL::Mapped self) 3093mapped (OpenCL::Mapped self)
2831 CODE: 3094 CODE:
2832 RETVAL = !!self->ptr; 3095 RETVAL = !!self->ptr;

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines