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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.63 by root, Tue May 1 22:25:13 2012 UTC vs.
Revision 1.73 by root, Sat May 5 13:30:07 2012 UTC

1#include "EXTERN.h" 1#include "EXTERN.h"
2#include "perl.h" 2#include "perl.h"
3#include "XSUB.h" 3#include "XSUB.h"
4
5#include "ecb.h"//D
4 6
5#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really 7#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
6#include "xthread.h" 8#include "xthread.h"
7#include "schmorp.h" 9#include "schmorp.h"
8 10
38 #define PREFER_1_1 1 40 #define PREFER_1_1 1
39#endif 41#endif
40 42
41typedef cl_platform_id OpenCL__Platform; 43typedef cl_platform_id OpenCL__Platform;
42typedef cl_device_id OpenCL__Device; 44typedef cl_device_id OpenCL__Device;
45typedef cl_device_id OpenCL__SubDevice;
43typedef cl_context OpenCL__Context; 46typedef cl_context OpenCL__Context;
44typedef cl_command_queue OpenCL__Queue; 47typedef cl_command_queue OpenCL__Queue;
45typedef cl_mem OpenCL__Memory; 48typedef cl_mem OpenCL__Memory;
46typedef cl_mem OpenCL__Buffer; 49typedef cl_mem OpenCL__Buffer;
47typedef cl_mem OpenCL__BufferObj; 50typedef cl_mem OpenCL__BufferObj;
48typedef cl_mem OpenCL__Image; 51typedef cl_mem OpenCL__Image;
49typedef cl_mem OpenCL__Image2D;
50typedef cl_mem OpenCL__Image3D;
51typedef cl_mem OpenCL__Memory_ornull; 52typedef cl_mem OpenCL__Memory_ornull;
52typedef cl_mem OpenCL__Buffer_ornull; 53typedef cl_mem OpenCL__Buffer_ornull;
53typedef cl_mem OpenCL__Image_ornull; 54typedef cl_mem OpenCL__Image_ornull;
54typedef cl_mem OpenCL__Image2D_ornull;
55typedef cl_mem OpenCL__Image3D_ornull;
56typedef cl_sampler OpenCL__Sampler; 55typedef cl_sampler OpenCL__Sampler;
57typedef cl_program OpenCL__Program; 56typedef cl_program OpenCL__Program;
58typedef cl_kernel OpenCL__Kernel; 57typedef cl_kernel OpenCL__Kernel;
59typedef cl_event OpenCL__Event; 58typedef cl_event OpenCL__Event;
60typedef cl_event OpenCL__UserEvent; 59typedef cl_event OpenCL__UserEvent;
61 60
62typedef struct mapped * OpenCL__Mapped; 61typedef struct mapped * OpenCL__Mapped;
63 62
64typedef SV *FUTURE;
65
66static HV 63static HV
67 *stash_platform, 64 *stash_platform,
68 *stash_device, 65 *stash_device,
66 *stash_subdevice,
69 *stash_context, 67 *stash_context,
70 *stash_queue, 68 *stash_queue,
71 *stash_program, 69 *stash_program,
72 *stash_kernel, 70 *stash_kernel,
73 *stash_sampler, 71 *stash_sampler,
133} 131}
134 132
135/*****************************************************************************/ 133/*****************************************************************************/
136 134
137/* up to two temporary buffers */ 135/* up to two temporary buffers */
138static void * 136static void * ecb_noinline
139tmpbuf (size_t size) 137tmpbuf (size_t size)
140{ 138{
141 enum { buffers = 3 }; 139 enum { buffers = 4 };
142 static int idx; 140 static int idx;
143 static void *buf [buffers]; 141 static void *buf [buffers];
144 static size_t len [buffers]; 142 static size_t len [buffers];
145 143
146 idx = (idx + 1) % buffers; 144 idx = (idx + 1) % buffers;
151 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 149 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
152 buf [idx] = malloc (len [idx]); 150 buf [idx] = malloc (len [idx]);
153 } 151 }
154 152
155 return buf [idx]; 153 return buf [idx];
154}
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;
156} 169}
157 170
158/*****************************************************************************/ 171/*****************************************************************************/
159 172
160typedef struct 173typedef struct
219 if (res) \ 232 if (res) \
220 FAIL (name); 233 FAIL (name);
221 234
222/*****************************************************************************/ 235/*****************************************************************************/
223 236
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
224static cl_context_properties * 270static cl_context_properties * ecb_noinline
225SvCONTEXTPROPERTIES (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)
226{ 272{
227 if (!sv || !SvOK (sv)) 273 if (!sv || !SvOK (sv))
228 if (extra) 274 if (extra)
229 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
230 else 276 else
236 int i, len = av_len (av) + 1; 282 int i, len = av_len (av) + 1;
237 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));
238 cl_context_properties *l = p; 284 cl_context_properties *l = p;
239 285
240 if (len & 1) 286 if (len & 1)
241 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);
242 288
243 while (extracount--) 289 while (extracount--)
244 *l++ = *extra++; 290 *l++ = *extra++;
245 291
246 for (i = 0; i < len; i += 2) 292 for (i = 0; i < len; i += 2)
249 SV *p_sv = *av_fetch (av, i + 1, 0); 295 SV *p_sv = *av_fetch (av, i + 1, 0);
250 cl_context_properties v = SvIV (p_sv); // code below can override 296 cl_context_properties v = SvIV (p_sv); // code below can override
251 297
252 switch (t) 298 switch (t)
253 { 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
254 case CL_GLX_DISPLAY_KHR: 305 case CL_GLX_DISPLAY_KHR:
255 if (!SvOK (p_sv)) 306 if (!SvOK (p_sv))
256 { 307 {
257 void *func = glsym ("_glXGetCurrentDisplay"); 308 void *func = glsym ("_glXGetCurrentDisplay");
258 if (func) 309 if (func)
281 *l = 0; 332 *l = 0;
282 333
283 return p; 334 return p;
284 } 335 }
285 336
286 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);
287} 338}
288 339
289/*****************************************************************************/ 340// parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation
290 341// is a pointer (and only on well-behaved systems).
291static SV * 342static void * ecb_noinline
292new_clobj (HV *stash, IV id) 343object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount)
293{ 344{
294 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); 345 if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV)
295} 346 croak ("%s: '%s' parameter must be %sa reference to an array of %s objects",
347 cv_get_name (cv), argname, or_undef ? "undef or " : "", klass);
296 348
297#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) 349 AV *av = (AV *)SvRV (arg);
298#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) 350 void **list = 0;
351 cl_uint count = av_len (av) + 1;
299 352
300/* cl objects are either \$iv, or [$iv, ...] */ 353 if (count)
301/* they can be upgraded at runtime to the array form */ 354 {
302static void * 355 list = tmpbuf (sizeof (*list) * count);
303SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) 356 int i;
304{ 357 for (i = 0; i < count; ++i)
305 if (SvROK (sv) && sv_derived_from (sv, pkg)) 358 list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass);
306 return (void *)SvIV (SvRV (sv)); 359 }
307 360
308 croak ("%s: %s is not of type %s", func, svname, pkg); 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;
309} 367}
310 368
311/*****************************************************************************/ 369/*****************************************************************************/
312/* callback stuff */ 370/* callback stuff */
313 371
335static void (*eq_signal_func)(void *signal_arg, int value); 393static void (*eq_signal_func)(void *signal_arg, int value);
336static void *eq_signal_arg; 394static void *eq_signal_arg;
337static xmutex_t eq_lock = X_MUTEX_INIT; 395static xmutex_t eq_lock = X_MUTEX_INIT;
338static eq_item *eq_head, *eq_tail; 396static eq_item *eq_head, *eq_tail;
339 397
340static void 398static void ecb_noinline
341eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) 399eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3)
342{ 400{
343 eq_item *item = malloc (sizeof (eq_item)); 401 eq_item *item = malloc (sizeof (eq_item));
344 402
345 item->next = 0; 403 item->next = 0;
412} 470}
413 471
414/*****************************************************************************/ 472/*****************************************************************************/
415/* context notify */ 473/* context notify */
416 474
417static void 475static void ecb_noinline
418eq_context_push (void *data1, void *data2, void *data3) 476eq_context_push (void *data1, void *data2, void *data3)
419{ 477{
420 dSP; 478 dSP;
421 PUSHs (sv_2mortal (newSVpv (data1, 0))); 479 PUSHs (sv_2mortal (newSVpv (data1, 0)));
422 PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); 480 PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3)));
444 { \ 502 { \
445 pfn_notify = eq_context_notify; \ 503 pfn_notify = eq_context_notify; \
446 user_data = s_get_cv (notify); \ 504 user_data = s_get_cv (notify); \
447 } 505 }
448 506
449static SV * 507static SV * ecb_noinline
450new_clobj_context (cl_context ctx, void *user_data) 508new_clobj_context (cl_context ctx, void *user_data)
451{ 509{
452 SV *sv = new_clobj (stash_context, (IV)ctx); 510 SV *sv = new_clobj (stash_context, (IV)ctx);
453 511
454 if (user_data) 512 if (user_data)
473static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; 531static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
474 532
475static void CL_CALLBACK 533static void CL_CALLBACK
476eq_program_notify (cl_program program, void *user_data) 534eq_program_notify (cl_program program, void *user_data)
477{ 535{
536 clRetainProgram (program);
537
478 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 }
479} 556}
480 557
481struct build_args 558struct build_args
482{ 559{
483 cl_program program; 560 cl_program program;
497 else 574 else
498 clReleaseProgram (arg->program); 575 clReleaseProgram (arg->program);
499 576
500 free (arg->options); 577 free (arg->options);
501 free (arg); 578 free (arg);
579
580 return 0;
502} 581}
503 582
504static void 583static void
505build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) 584build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data)
506{ 585{
546 size_t res; 625 size_t res;
547 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 626 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
548 return res; 627 return res;
549} 628}
550 629
551static cl_event * 630static cl_event * ecb_noinline
552event_list (SV **items, cl_uint *rcount, cl_event extra) 631event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
553{ 632{
554 cl_uint count = *rcount; 633 cl_uint count = *rcount;
555 634
556 if (count > 0x7fffffff) // yeha, it's a hack - the caller might have underflowed 635 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed
557 count = 0; 636 *rcount = count = 0;
558 637
559 if (!count && !extra) 638 if (!count && !extra)
560 return 0; 639 return 0;
561 640
562 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); 641 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
563 int i = 0; 642 int i = 0;
564 643
565 while (count--) 644 while (count--)
566 if (SvOK (items [count])) 645 if (SvOK (items [count]))
567 list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 646 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
568 647
569 if (extra) 648 if (extra)
570 list [i++] = extra; 649 list [i++] = extra;
571 650
572 *rcount = i; 651 *rcount = i;
574 return i ? list : 0; 653 return i ? list : 0;
575} 654}
576 655
577#define EVENT_LIST(skip) \ 656#define EVENT_LIST(skip) \
578 cl_uint event_list_count = items - (skip); \ 657 cl_uint event_list_count = items - (skip); \
579 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)
580 659
581#define INFO(class) \ 660#define INFO(class) \
582{ \ 661{ \
583 size_t size; \ 662 size_t size; \
584 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 663 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
666 745
667 mapped->ptr = 0; 746 mapped->ptr = 0;
668} 747}
669 748
670static void 749static void
671mapped_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)
672{ 751{
673 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);
674 cl_event ev; 753 cl_event ev;
675 754
676 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));
677 756
678 clReleaseEvent (mapped->event); 757 clReleaseEvent (mapped->event);
711 { sizeof (cl_long ), "SIZEOF_LONG" }, 790 { sizeof (cl_long ), "SIZEOF_LONG" },
712 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 791 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
713 { sizeof (cl_half ), "SIZEOF_HALF" }, 792 { sizeof (cl_half ), "SIZEOF_HALF" },
714 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 793 { sizeof (cl_float ), "SIZEOF_FLOAT" },
715 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 794 { sizeof (cl_double), "SIZEOF_DOUBLE" },
795 { PREFER_1_1 , "PREFER_1_1" },
716#include "constiv.h" 796#include "constiv.h"
717 }; 797 };
718 798
719 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--)
720 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 800 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
721 801
722 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); 802 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
723 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);
724 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); 805 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
725 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); 806 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD);
726 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); 807 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD);
727 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); 808 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD);
728 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); 809 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD);
769 list = tmpbuf (sizeof (*list) * count); 850 list = tmpbuf (sizeof (*list) * count);
770 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 851 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
771 852
772 EXTEND (SP, count); 853 EXTEND (SP, count);
773 for (i = 0; i < count; ++i) 854 for (i = 0; i < count; ++i)
774 PUSH_CLOBJ (stash_platform, (IV)list [i]); 855 PUSH_CLOBJ (stash_platform, list [i]);
775 856
776void 857void
777context_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)
778 PPCODE: 859 PPCODE:
779 CONTEXT_NOTIFY_CALLBACK; 860 CONTEXT_NOTIFY_CALLBACK;
780 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); 861 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
781 XPUSH_CLOBJ_CONTEXT; 862 XPUSH_CLOBJ_CONTEXT;
782 863
783void 864void
784context (FUTURE properties, FUTURE devices, FUTURE notify) 865context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
785 PPCODE: 866 PPCODE:
786 /* 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;
787 873
788void 874void
789wait_for_events (...) 875wait_for_events (...)
790 CODE: 876 CODE:
791 EVENT_LIST (0); 877 EVENT_LIST (0);
842 EXTEND (SP, count); 928 EXTEND (SP, count);
843 for (i = 0; i < count; ++i) 929 for (i = 0; i < count; ++i)
844 PUSH_CLOBJ (stash_device, list [i]); 930 PUSH_CLOBJ (stash_device, list [i]);
845 931
846void 932void
847context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) 933context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
848 PPCODE: 934 PPCODE:
935 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
936 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
937
938 cl_uint device_count;
939 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
940
941 CONTEXT_NOTIFY_CALLBACK;
942 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
943 XPUSH_CLOBJ_CONTEXT;
944
945void
946context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
947 PPCODE:
948 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
949 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
950
951 CONTEXT_NOTIFY_CALLBACK;
952 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res));
953 XPUSH_CLOBJ_CONTEXT;
954
955MODULE = OpenCL PACKAGE = OpenCL::Device
956
957void
958info (OpenCL::Device self, cl_device_info name)
959 PPCODE:
960 INFO (Device)
961
962#if CL_VERSION_1_2
963
964void
965sub_devices (OpenCL::Device self, SV *properties)
966 PPCODE:
849 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 967 if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV)
850 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); 968 croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs");
851 969
852 AV *av = (AV *)SvRV (devices); 970 properties = SvRV (properties);
853 cl_uint num_devices = av_len (av) + 1; 971
854 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); 972 cl_uint count = av_len ((AV *)properties) + 1;
973 cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1);
855 974
856 int i; 975 int i;
857 for (i = num_devices; i--; ) 976 for (i = 0; i < count; ++i)
858 device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); 977 props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0));
859 978
860 CONTEXT_NOTIFY_CALLBACK; 979 props [count] = 0;
861 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res));
862 XPUSH_CLOBJ_CONTEXT;
863 980
864void 981 cl_uint num_devices;
865context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) 982 NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices));
866 PPCODE: 983 cl_device_id *list = tmpbuf (sizeof (*list) * num_devices);
867 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; 984 NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0));
868 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2);
869 985
870 CONTEXT_NOTIFY_CALLBACK; 986 EXTEND (SP, num_devices);
871 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); 987 for (i = 0; i < count; ++i)
872 XPUSH_CLOBJ_CONTEXT; 988 PUSH_CLOBJ (stash_subdevice, list [i]);
873 989
874MODULE = OpenCL PACKAGE = OpenCL::Device 990#endif
875
876void
877info (OpenCL::Device self, cl_device_info name)
878 PPCODE:
879 INFO (Device)
880 991
881#BEGIN:device 992#BEGIN:device
882 993
883void 994void
884type (OpenCL::Device self) 995type (OpenCL::Device self)
1100 for (i = 0; i < n; ++i) 1211 for (i = 0; i < n; ++i)
1101 PUSHs (sv_2mortal (newSVuv (value [i]))); 1212 PUSHs (sv_2mortal (newSVuv (value [i])));
1102 1213
1103#END:device 1214#END:device
1104 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
1105MODULE = OpenCL PACKAGE = OpenCL::Context 1227MODULE = OpenCL PACKAGE = OpenCL::Context
1106 1228
1107void 1229void
1108DESTROY (OpenCL::Context context) 1230DESTROY (OpenCL::Context self)
1109 CODE: 1231 CODE:
1110 clReleaseContext (context); 1232 clReleaseContext (self);
1111 1233
1112void 1234void
1113info (OpenCL::Context self, cl_context_info name) 1235info (OpenCL::Context self, cl_context_info name)
1114 PPCODE: 1236 PPCODE:
1115 INFO (Context) 1237 INFO (Context)
1156 const cl_image_desc desc = { 1278 const cl_image_desc desc = {
1157 type, 1279 type,
1158 width, height, depth, 1280 width, height, depth,
1159 array_size, row_pitch, slice_pitch, 1281 array_size, row_pitch, slice_pitch,
1160 num_mip_level, num_samples, 1282 num_mip_level, num_samples,
1161 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
1162 }; 1284 };
1163 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));
1164 HV *stash = stash_image; 1286 HV *stash = stash_image;
1165 switch (type) 1287 switch (type)
1166 { 1288 {
1299 1421
1300 len2 = len; 1422 len2 = len;
1301 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); 1423 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res));
1302 XPUSH_CLOBJ (stash_program, prog); 1424 XPUSH_CLOBJ (stash_program, prog);
1303 1425
1426void
1427program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1428 PPCODE:
1429 cl_uint device_count;
1430 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1431
1432 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1433 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1434
1435 binaries = SvRV (binaries);
1436
1437 if (device_count != av_len ((AV *)binaries) + 1)
1438 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1439
1440 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1441 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1442 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1443
1444 int i;
1445 for (i = 0; i < device_count; ++i)
1446 {
1447 STRLEN len;
1448 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1449 length_list [i] = len;
1450 }
1451
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));
1455
1456 EXTEND (SP, 2);
1457 PUSH_CLOBJ (stash_program, prog);
1458
1459 if (GIMME_V == G_ARRAY)
1460 {
1461 AV *av = newAV ();
1462 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1463
1464 for (i = device_count; i--; )
1465 av_store (av, i, newSViv (status_list [i]));
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
1501
1304#BEGIN:context 1502#BEGIN:context
1305 1503
1306void 1504void
1307reference_count (OpenCL::Context self) 1505reference_count (OpenCL::Context self)
1308 ALIAS: 1506 ALIAS:
1353void 1551void
1354read_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, ...)
1355 ALIAS: 1553 ALIAS:
1356 enqueue_read_buffer = 0 1554 enqueue_read_buffer = 0
1357 PPCODE: 1555 PPCODE:
1358 cl_event ev = 0;
1359 EVENT_LIST (6); 1556 EVENT_LIST (6);
1360 1557
1361 SvUPGRADE (data, SVt_PV); 1558 SvUPGRADE (data, SVt_PV);
1362 SvGROW (data, len); 1559 SvGROW (data, len);
1363 SvPOK_only (data); 1560 SvPOK_only (data);
1364 SvCUR_set (data, len); 1561 SvCUR_set (data, len);
1562
1563 cl_event ev = 0;
1365 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));
1366 1565
1367 if (ev) 1566 if (ev)
1368 XPUSH_CLOBJ (stash_event, ev); 1567 XPUSH_CLOBJ (stash_event, ev);
1369 1568
1370void 1569void
1371write_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, ...)
1372 ALIAS: 1571 ALIAS:
1373 enqueue_write_buffer = 0 1572 enqueue_write_buffer = 0
1374 PPCODE: 1573 PPCODE:
1375 cl_event ev = 0; 1574 EVENT_LIST (5);
1575
1376 STRLEN len; 1576 STRLEN len;
1377 char *ptr = SvPVbyte (data, len); 1577 char *ptr = SvPVbyte (data, len);
1378 EVENT_LIST (5);
1379 1578
1579 cl_event ev = 0;
1380 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));
1381 1581
1382 if (ev) 1582 if (ev)
1383 XPUSH_CLOBJ (stash_event, ev); 1583 XPUSH_CLOBJ (stash_event, ev);
1384 1584
1387void 1587void
1388fill_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, ...)
1389 ALIAS: 1589 ALIAS:
1390 enqueue_fill_buffer = 0 1590 enqueue_fill_buffer = 0
1391 PPCODE: 1591 PPCODE:
1392 cl_event ev = 0; 1592 EVENT_LIST (5);
1593
1393 STRLEN len; 1594 STRLEN len;
1394 char *ptr = SvPVbyte (data, len); 1595 char *ptr = SvPVbyte (data, len);
1395 EVENT_LIST (5);
1396 1596
1597 cl_event ev = 0;
1397 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));
1398 1599
1399 if (ev) 1600 if (ev)
1400 XPUSH_CLOBJ (stash_event, ev); 1601 XPUSH_CLOBJ (stash_event, ev);
1401 1602
1402void 1603void
1403fill_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, ...)
1404 ALIAS: 1605 ALIAS:
1405 enqueue_fill_image = 0 1606 enqueue_fill_image = 0
1406 PPCODE: 1607 PPCODE:
1407 cl_event ev = 0; 1608 EVENT_LIST (12);
1408 STRLEN len; 1609
1409 const size_t origin [3] = { x, y, z }; 1610 const size_t origin [3] = { x, y, z };
1410 const size_t region [3] = { width, height, depth }; 1611 const size_t region [3] = { width, height, depth };
1411 EVENT_LIST (12);
1412 1612
1413 const cl_float c_f [4] = { r, g, b, a }; 1613 const cl_float c_f [4] = { r, g, b, a };
1414 const cl_uint c_u [4] = { r, g, b, a }; 1614 const cl_uint c_u [4] = { r, g, b, a };
1415 const cl_int c_s [4] = { r, g, b, a }; 1615 const cl_int c_s [4] = { r, g, b, a };
1416 const void *c_fus [3] = { &c_f, &c_u, &c_s }; 1616 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1419 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); 1619 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1420 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); 1620 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1421 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)
1422 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.");
1423 1623
1624 cl_event ev = 0;
1424 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]],
1425 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));
1426 1627
1427 if (ev) 1628 if (ev)
1428 XPUSH_CLOBJ (stash_event, ev); 1629 XPUSH_CLOBJ (stash_event, ev);
1432void 1633void
1433copy_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, ...)
1434 ALIAS: 1635 ALIAS:
1435 enqueue_copy_buffer = 0 1636 enqueue_copy_buffer = 0
1436 PPCODE: 1637 PPCODE:
1638 EVENT_LIST (6);
1639
1437 cl_event ev = 0; 1640 cl_event ev = 0;
1438 EVENT_LIST (6);
1439
1440 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));
1441 1642
1442 if (ev) 1643 if (ev)
1443 XPUSH_CLOBJ (stash_event, ev); 1644 XPUSH_CLOBJ (stash_event, ev);
1444 1645
1445void 1646void
1446read_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, ...)
1447 ALIAS: 1648 ALIAS:
1448 enqueue_read_buffer_rect = 0 1649 enqueue_read_buffer_rect = 0
1449 PPCODE: 1650 PPCODE:
1450 cl_event ev = 0; 1651 EVENT_LIST (17);
1652
1451 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 };
1452 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 };
1453 const size_t region[3] = { width, height, depth }; 1655 const size_t region[3] = { width, height, depth };
1454 EVENT_LIST (17);
1455 1656
1456 if (!buf_row_pitch) 1657 if (!buf_row_pitch)
1457 buf_row_pitch = region [0]; 1658 buf_row_pitch = region [0];
1458 1659
1459 if (!buf_slice_pitch) 1660 if (!buf_slice_pitch)
1469 1670
1470 SvUPGRADE (data, SVt_PV); 1671 SvUPGRADE (data, SVt_PV);
1471 SvGROW (data, len); 1672 SvGROW (data, len);
1472 SvPOK_only (data); 1673 SvPOK_only (data);
1473 SvCUR_set (data, len); 1674 SvCUR_set (data, len);
1675
1676 cl_event ev = 0;
1474 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));
1475 1678
1476 if (ev) 1679 if (ev)
1477 XPUSH_CLOBJ (stash_event, ev); 1680 XPUSH_CLOBJ (stash_event, ev);
1478 1681
1479void 1682void
1480write_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, ...)
1481 ALIAS: 1684 ALIAS:
1482 enqueue_write_buffer_rect = 0 1685 enqueue_write_buffer_rect = 0
1483 PPCODE: 1686 PPCODE:
1484 cl_event ev = 0; 1687 EVENT_LIST (17);
1688
1485 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 };
1486 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 };
1487 const size_t region[3] = { width, height, depth }; 1691 const size_t region[3] = { width, height, depth };
1488 STRLEN len; 1692 STRLEN len;
1489 char *ptr = SvPVbyte (data, len); 1693 char *ptr = SvPVbyte (data, len);
1490 EVENT_LIST (17);
1491 1694
1492 if (!buf_row_pitch) 1695 if (!buf_row_pitch)
1493 buf_row_pitch = region [0]; 1696 buf_row_pitch = region [0];
1494 1697
1495 if (!buf_slice_pitch) 1698 if (!buf_slice_pitch)
1504 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];
1505 1708
1506 if (len < min_len) 1709 if (len < min_len)
1507 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1710 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1508 1711
1712 cl_event ev = 0;
1509 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));
1510 1714
1511 if (ev) 1715 if (ev)
1512 XPUSH_CLOBJ (stash_event, ev); 1716 XPUSH_CLOBJ (stash_event, ev);
1513 1717
1514void 1718void
1515copy_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, ...)
1516 ALIAS: 1720 ALIAS:
1517 enqueue_copy_buffer_rect = 0 1721 enqueue_copy_buffer_rect = 0
1518 PPCODE: 1722 PPCODE:
1519 cl_event ev = 0; 1723 EVENT_LIST (16);
1724
1520 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 };
1521 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 };
1522 const size_t region[3] = { width, height, depth }; 1727 const size_t region[3] = { width, height, depth };
1523 EVENT_LIST (16);
1524 1728
1729 cl_event ev = 0;
1525 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));
1526 1731
1527 if (ev) 1732 if (ev)
1528 XPUSH_CLOBJ (stash_event, ev); 1733 XPUSH_CLOBJ (stash_event, ev);
1529 1734
1530void 1735void
1531read_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, ...)
1532 ALIAS: 1737 ALIAS:
1533 enqueue_read_image = 0 1738 enqueue_read_image = 0
1534 PPCODE: 1739 PPCODE:
1535 cl_event ev = 0; 1740 EVENT_LIST (12);
1741
1536 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 };
1537 const size_t region[3] = { width, height, depth }; 1743 const size_t region[3] = { width, height, depth };
1538 EVENT_LIST (12);
1539 1744
1540 if (!row_pitch) 1745 if (!row_pitch)
1541 row_pitch = img_row_pitch (src); 1746 row_pitch = img_row_pitch (src);
1542 1747
1543 if (depth > 1 && !slice_pitch) 1748 if (depth > 1 && !slice_pitch)
1547 1752
1548 SvUPGRADE (data, SVt_PV); 1753 SvUPGRADE (data, SVt_PV);
1549 SvGROW (data, len); 1754 SvGROW (data, len);
1550 SvPOK_only (data); 1755 SvPOK_only (data);
1551 SvCUR_set (data, len); 1756 SvCUR_set (data, len);
1757
1758 cl_event ev = 0;
1552 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));
1553 1760
1554 if (ev) 1761 if (ev)
1555 XPUSH_CLOBJ (stash_event, ev); 1762 XPUSH_CLOBJ (stash_event, ev);
1556 1763
1557void 1764void
1558write_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, ...)
1559 ALIAS: 1766 ALIAS:
1560 enqueue_write_image = 0 1767 enqueue_write_image = 0
1561 PPCODE: 1768 PPCODE:
1562 cl_event ev = 0; 1769 EVENT_LIST (12);
1770
1563 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 };
1564 const size_t region[3] = { width, height, depth }; 1772 const size_t region[3] = { width, height, depth };
1565 STRLEN len; 1773 STRLEN len;
1566 char *ptr = SvPVbyte (data, len); 1774 char *ptr = SvPVbyte (data, len);
1567 EVENT_LIST (12);
1568 1775
1569 if (!row_pitch) 1776 if (!row_pitch)
1570 row_pitch = img_row_pitch (dst); 1777 row_pitch = img_row_pitch (dst);
1571 1778
1572 if (depth > 1 && !slice_pitch) 1779 if (depth > 1 && !slice_pitch)
1575 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;
1576 1783
1577 if (len < min_len) 1784 if (len < min_len)
1578 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1785 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1579 1786
1787 cl_event ev = 0;
1580 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));
1581 1789
1582 if (ev) 1790 if (ev)
1583 XPUSH_CLOBJ (stash_event, ev); 1791 XPUSH_CLOBJ (stash_event, ev);
1584 1792
1585void 1793void
1586copy_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, ...)
1587 ALIAS: 1795 ALIAS:
1588 enqueue_copy_image = 0 1796 enqueue_copy_image = 0
1589 PPCODE: 1797 PPCODE:
1590 cl_event ev = 0; 1798 EVENT_LIST (12);
1799
1591 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 };
1592 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 };
1593 const size_t region[3] = { width, height, depth }; 1802 const size_t region[3] = { width, height, depth };
1594 EVENT_LIST (12);
1595 1803
1804 cl_event ev = 0;
1596 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));
1597 1806
1598 if (ev) 1807 if (ev)
1599 XPUSH_CLOBJ (stash_event, ev); 1808 XPUSH_CLOBJ (stash_event, ev);
1600 1809
1601void 1810void
1602copy_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, ...)
1603 ALIAS: 1812 ALIAS:
1604 enqueue_copy_image_to_buffer = 0 1813 enqueue_copy_image_to_buffer = 0
1605 PPCODE: 1814 PPCODE:
1606 cl_event ev = 0; 1815 EVENT_LIST (10);
1816
1607 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 };
1608 const size_t region [3] = { width, height, depth }; 1818 const size_t region [3] = { width, height, depth };
1609 EVENT_LIST (10);
1610 1819
1820 cl_event ev = 0;
1611 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));
1612 1822
1613 if (ev) 1823 if (ev)
1614 XPUSH_CLOBJ (stash_event, ev); 1824 XPUSH_CLOBJ (stash_event, ev);
1615 1825
1616void 1826void
1617copy_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, ...)
1618 ALIAS: 1828 ALIAS:
1619 enqueue_copy_buffer_to_image = 0 1829 enqueue_copy_buffer_to_image = 0
1620 PPCODE: 1830 PPCODE:
1621 cl_event ev = 0; 1831 EVENT_LIST (10);
1832
1622 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 };
1623 const size_t region [3] = { width, height, depth }; 1834 const size_t region [3] = { width, height, depth };
1624 EVENT_LIST (10);
1625 1835
1836 cl_event ev = 0;
1626 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));
1627 1838
1628 if (ev) 1839 if (ev)
1629 XPUSH_CLOBJ (stash_event, ev); 1840 XPUSH_CLOBJ (stash_event, ev);
1630 1841
1631void 1842void
1632map_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, size_t cb = 0, ...) 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, ...)
1633 ALIAS: 1844 ALIAS:
1634 enqueue_map_buffer = 0 1845 enqueue_map_buffer = 0
1635 PPCODE: 1846 PPCODE:
1847 EVENT_LIST (6);
1848
1849 size_t cb = SvIV (cb_);
1850
1851 if (!SvOK (cb_))
1852 {
1853 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1854 cb -= offset;
1855 }
1856
1636 cl_event ev; 1857 cl_event ev;
1637 EVENT_LIST (6);
1638
1639 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));
1640 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));
1641 1860
1642void 1861void
1643map_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, size_t width = 0, size_t height = 0, size_t depth = 0, ...) 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, ...)
1644 ALIAS: 1863 ALIAS:
1645 enqueue_map_image = 0 1864 enqueue_map_image = 0
1646 PPCODE: 1865 PPCODE:
1647 cl_event ev; 1866 size_t width = SvIV (width_);
1867 if (!SvOK (width_))
1868 {
1869 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0));
1870 width -= x;
1871 }
1872
1873 size_t height = SvIV (width_);
1874 if (!SvOK (height_))
1875 {
1876 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0));
1877 height -= y;
1878 }
1879
1880 size_t depth = SvIV (width_);
1881 if (!SvOK (depth_))
1882 {
1883 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0));
1884 depth -= z;
1885
1886 // stupid opencl returns 0 for depth, but requires 1 for 2d images
1887 if (!depth)
1888 depth = 1;
1889 }
1890
1648 const size_t origin[3] = { x, y, z }; 1891 const size_t origin[3] = { x, y, z };
1649 const size_t region[3] = { width, height, depth }; 1892 const size_t region[3] = { width, height, depth };
1650 size_t row_pitch, slice_pitch; 1893 size_t row_pitch, slice_pitch;
1651 EVENT_LIST (10); 1894 EVENT_LIST (10);
1652 1895
1896 cl_event ev;
1653 NEED_SUCCESS_ARG (void *ptr, EnqueueMapImage, (self, img, blocking, map_flags, origin, region, &row_pitch, &slice_pitch, event_list_count, event_list_ptr, &ev, &res)); 1897 NEED_SUCCESS_ARG (void *ptr, EnqueueMapImage, (self, img, blocking, map_flags, origin, region, &row_pitch, &slice_pitch, event_list_count, event_list_ptr, &ev, &res));
1654 1898
1655 size_t cb = slice_pitch ? slice_pitch * region [2] 1899 size_t cb = slice_pitch ? slice_pitch * region [2]
1656 : row_pitch ? row_pitch * region [1] 1900 : row_pitch ? row_pitch * region [1]
1657 : region [0]; 1901 : region [0];
1659 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));
1660 1904
1661void 1905void
1662unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) 1906unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1663 PPCODE: 1907 PPCODE:
1664 mapped_unmap (ST (1), mapped, self, &ST (2), items - 2); 1908 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1665 if (GIMME_V != G_VOID) 1909 if (GIMME_V != G_VOID)
1666 { 1910 {
1667 clRetainEvent (mapped->event); 1911 clRetainEvent (mapped->event);
1668 XPUSH_CLOBJ (stash_event, mapped->event); 1912 XPUSH_CLOBJ (stash_event, mapped->event);
1669 } 1913 }
1671void 1915void
1672task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) 1916task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1673 ALIAS: 1917 ALIAS:
1674 enqueue_task = 0 1918 enqueue_task = 0
1675 PPCODE: 1919 PPCODE:
1920 EVENT_LIST (2);
1921
1676 cl_event ev = 0; 1922 cl_event ev = 0;
1677 EVENT_LIST (2);
1678
1679 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));
1680 1924
1681 if (ev) 1925 if (ev)
1682 XPUSH_CLOBJ (stash_event, ev); 1926 XPUSH_CLOBJ (stash_event, ev);
1683 1927
1684void 1928void
1685nd_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, ...)
1686 ALIAS: 1930 ALIAS:
1687 enqueue_nd_range_kernel = 0 1931 enqueue_nd_range_kernel = 0
1688 PPCODE: 1932 PPCODE:
1689 cl_event ev = 0; 1933 EVENT_LIST (5);
1934
1690 size_t *gwo = 0, *gws, *lws = 0; 1935 size_t *gwo = 0, *gws, *lws = 0;
1691 int gws_len; 1936 int gws_len;
1692 size_t *lists; 1937 size_t *lists;
1693 int i; 1938 int i;
1694 EVENT_LIST (5);
1695 1939
1696 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)
1697 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1941 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
1698 1942
1699 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1943 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
1738 if (!lws [i]) 1982 if (!lws [i])
1739 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);
1740 } 1984 }
1741 } 1985 }
1742 1986
1987 cl_event ev = 0;
1743 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));
1744 1989
1745 if (ev) 1990 if (ev)
1746 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
1747 2012
1748#if cl_apple_gl_sharing || cl_khr_gl_sharing 2013#if cl_apple_gl_sharing || cl_khr_gl_sharing
1749 2014
1750void 2015void
1751acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) 2016acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1752 ALIAS: 2017 ALIAS:
1753 release_gl_objects = 1 2018 release_gl_objects = 1
1754 enqueue_acquire_gl_objects = 0 2019 enqueue_acquire_gl_objects = 0
1755 enqueue_release_gl_objects = 1 2020 enqueue_release_gl_objects = 1
1756 PPCODE: 2021 PPCODE:
1757 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) 2022 EVENT_LIST (2);
1758 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);
1759 2026
1760 cl_event ev = 0; 2027 cl_event ev = 0;
1761 EVENT_LIST (2);
1762 AV *av = (AV *)SvRV (objects);
1763 cl_uint num_objects = av_len (av) + 1;
1764 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1765 int i;
1766
1767 for (i = num_objects; i--; )
1768 object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1769 2028
1770 if (ix) 2029 if (ix)
1771 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));
1772 else 2031 else
1773 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));
1774 2033
1775 if (ev) 2034 if (ev)
1776 XPUSH_CLOBJ (stash_event, ev); 2035 XPUSH_CLOBJ (stash_event, ev);
1777 2036
1778#endif 2037#endif
1792void 2051void
1793marker (OpenCL::Queue self, ...) 2052marker (OpenCL::Queue self, ...)
1794 ALIAS: 2053 ALIAS:
1795 enqueue_marker = 0 2054 enqueue_marker = 0
1796 PPCODE: 2055 PPCODE:
2056 EVENT_LIST (1);
1797 cl_event ev = 0; 2057 cl_event ev = 0;
1798 EVENT_LIST (1);
1799#if PREFER_1_1 2058#if PREFER_1_1
1800 if (!event_list_count) 2059 if (!event_list_count)
1801 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); 2060 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1802 else 2061 else
1803#if CL_VERSION_1_2 2062#if CL_VERSION_1_2
1817void 2076void
1818barrier (OpenCL::Queue self, ...) 2077barrier (OpenCL::Queue self, ...)
1819 ALIAS: 2078 ALIAS:
1820 enqueue_barrier = 0 2079 enqueue_barrier = 0
1821 PPCODE: 2080 PPCODE:
2081 EVENT_LIST (1);
1822 cl_event ev = 0; 2082 cl_event ev = 0;
1823 EVENT_LIST (1);
1824#if PREFER_1_1 2083#if PREFER_1_1
1825 if (!event_list_count && GIMME_V == G_VOID) 2084 if (!event_list_count && GIMME_V == G_VOID)
1826 NEED_SUCCESS (EnqueueBarrier, (self)); 2085 NEED_SUCCESS (EnqueueBarrier, (self));
1827 else 2086 else
1828#if CL_VERSION_1_2 2087#if CL_VERSION_1_2
2156void 2415void
2157build (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)
2158 ALIAS: 2417 ALIAS:
2159 build_async = 1 2418 build_async = 1
2160 CODE: 2419 CODE:
2161 void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0;
2162 void *user_data = 0;
2163 cl_uint num_devices = 0; 2420 cl_uint device_count = 0;
2164 cl_device_id *device_list = 0; 2421 cl_device_id *device_list = 0;
2165 2422
2166 if (SvOK (devices)) 2423 if (SvOK (devices))
2167 { 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 {
2168 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 2451 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2169 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");
2170 2453
2171 AV *av = (AV *)SvRV (devices); 2454 HV *hv = (HV *)SvRV (devices);
2172 num_devices = av_len (av) + 1;
2173 2455
2174 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))
2175 { 2463 {
2176 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");
2177 int count; 2466 ++i;
2178 for (count = 0; count < num_devices; ++count)
2179 device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device");
2180 } 2467 }
2181 } 2468 }
2182 2469
2183 if (SvOK (notify)) 2470 void *user_data;
2184 { 2471 program_callback pfn_notify = make_program_callback (notify, &user_data);
2185 NEED_SUCCESS (RetainProgram, (self));
2186 pfn_notify = eq_program_notify;
2187 user_data = SvREFCNT_inc (s_get_cv (notify));
2188 }
2189 2472
2190 if (ix)
2191 build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data);
2192 else
2193 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
2194 2477
2195void 2478void
2196build_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)
2197 PPCODE: 2480 PPCODE:
2198 size_t size; 2481 size_t size;
2226 char *value = tmpbuf (size); 2509 char *value = tmpbuf (size);
2227 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); 2510 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2228 EXTEND (SP, 1); 2511 EXTEND (SP, 1);
2229 const int i = 0; 2512 const int i = 0;
2230 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])));
2231 2523
2232#END:program_build 2524#END:program_build
2233 2525
2234void 2526void
2235kernel (OpenCL::Program program, SV *function) 2527kernel (OpenCL::Program program, SV *function)
2402 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; 2694 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2403 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; 2695 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2404 2696
2405 case 'z': nullarg = 1; size = SvIV (sv); break; 2697 case 'z': nullarg = 1; size = SvIV (sv); break;
2406 2698
2407 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;
2408 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;
2409 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;
2410 2702
2411 default: 2703 default:
2412 croak ("OpenCL::Kernel::setf format character '%c' not supported", type); 2704 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2413 } 2705 }
2414 2706
2614 const int i = 0; 2906 const int i = 0;
2615 PUSHs (sv_2mortal (newSVuv (value [i]))); 2907 PUSHs (sv_2mortal (newSVuv (value [i])));
2616 2908
2617#END:kernel_work_group 2909#END:kernel_work_group
2618 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
2619MODULE = OpenCL PACKAGE = OpenCL::Event 2972MODULE = OpenCL PACKAGE = OpenCL::Event
2620 2973
2621void 2974void
2622DESTROY (OpenCL::Event self) 2975DESTROY (OpenCL::Event self)
2623 CODE: 2976 CODE:
2732 Safefree (mapped); 3085 Safefree (mapped);
2733 3086
2734void 3087void
2735unmap (OpenCL::Mapped self, ...) 3088unmap (OpenCL::Mapped self, ...)
2736 CODE: 3089 CODE:
2737 mapped_unmap (ST (0), self, self->queue, &ST (1), items - 1); 3090 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
2738 3091
2739bool 3092bool
2740mapped (OpenCL::Mapped self) 3093mapped (OpenCL::Mapped self)
2741 CODE: 3094 CODE:
2742 RETVAL = !!self->ptr; 3095 RETVAL = !!self->ptr;
2776set (OpenCL::Mapped self, size_t offset, SV *data) 3129set (OpenCL::Mapped self, size_t offset, SV *data)
2777 CODE: 3130 CODE:
2778 STRLEN len; 3131 STRLEN len;
2779 const char *ptr = SvPVbyte (data, len); 3132 const char *ptr = SvPVbyte (data, len);
2780 3133
2781 if (offset < 0 || offset + len > self->cb) 3134 if (offset + len > self->cb)
2782 croak ("OpenCL::Mapped::set out of bound condition detected"); 3135 croak ("OpenCL::Mapped::set out of bound condition detected");
2783 3136
2784 memcpy (offset + (char *)self->ptr, ptr, len); 3137 memcpy (offset + (char *)self->ptr, ptr, len);
2785 3138
2786MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer 3139MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines