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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.6 by root, Thu Nov 17 01:36:52 2011 UTC vs.
Revision 1.75 by root, Sat May 5 14:18:35 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 4
5#include "ecb.h"//D
6
7#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
8#include "xthread.h"
9#include "schmorp.h"
10
11#ifdef I_DLFCN
12 #include <dlfcn.h>
13#endif
14
15// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
16// but then fail to define the api functions unless you ALSO define
17// this. This breaks 100% of the opencl 1.1 apps, for what reason?
18// after all, the functions are deprecated, not removed.
19// in addition, you cannot test for this in any future-proof way.
20// each time a new opencl version comes out, you need to make a new
21// release.
22#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
23
24#ifndef PREFER_1_1
25 #define PREFER_1_1 1
26#endif
27
28#if PREFER_1_1
29 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
30#endif
31
32#ifdef __APPLE__
33 #define CLHDR(name) <OpenCL/name>
34#else
35 #define CLHDR(name) <CL/name>
36#endif
37
5#include <CL/opencl.h> 38#include CLHDR(opencl.h)
6 39
40#ifndef CL_VERSION_1_2
41 #include CLHDR(cl_d3d9.h)
42#endif
43
44#if _WIN32
45 #include CLHDR(cl_d3d10.h)
46 #if CL_VERSION_1_2
47 #include CLHDR<cl_d3d11.h>
48 #endif
49 #include CLHDR<cl_dx9_media_sharing.h.h>
50#endif
51
52#ifndef CL_VERSION_1_2
53 #undef PREFER_1_1
54 #define PREFER_1_1 1
55#endif
56
57// make sure all constants we might use are actually defined
58#include "default.h"
59
7typedef cl_platform_id OpenCL__Platform; 60typedef cl_platform_id OpenCL__Platform;
8typedef cl_device_id OpenCL__Device; 61typedef cl_device_id OpenCL__Device;
62typedef cl_device_id OpenCL__SubDevice;
9typedef cl_context OpenCL__Context; 63typedef cl_context OpenCL__Context;
10typedef cl_command_queue OpenCL__Queue; 64typedef cl_command_queue OpenCL__Queue;
11typedef cl_mem OpenCL__Memory; 65typedef cl_mem OpenCL__Memory;
12typedef cl_mem OpenCL__Buffer; 66typedef cl_mem OpenCL__Buffer;
67typedef cl_mem OpenCL__BufferObj;
13typedef cl_mem OpenCL__Image; 68typedef cl_mem OpenCL__Image;
14typedef cl_mem OpenCL__Image2D;
15typedef cl_mem OpenCL__Image3D;
16typedef cl_mem OpenCL__Memory_ornull; 69typedef cl_mem OpenCL__Memory_ornull;
17typedef cl_mem OpenCL__Buffer_ornull; 70typedef cl_mem OpenCL__Buffer_ornull;
18typedef cl_mem OpenCL__Image_ornull; 71typedef cl_mem OpenCL__Image_ornull;
19typedef cl_mem OpenCL__Image2D_ornull;
20typedef cl_mem OpenCL__Image3D_ornull;
21typedef cl_sampler OpenCL__Sampler; 72typedef cl_sampler OpenCL__Sampler;
22typedef cl_program OpenCL__Program; 73typedef cl_program OpenCL__Program;
23typedef cl_kernel OpenCL__Kernel; 74typedef cl_kernel OpenCL__Kernel;
24typedef cl_event OpenCL__Event; 75typedef cl_event OpenCL__Event;
25typedef cl_event OpenCL__UserEvent; 76typedef cl_event OpenCL__UserEvent;
77
78typedef struct mapped * OpenCL__Mapped;
79
80static HV
81 *stash_platform,
82 *stash_device,
83 *stash_subdevice,
84 *stash_context,
85 *stash_queue,
86 *stash_program,
87 *stash_kernel,
88 *stash_sampler,
89 *stash_event,
90 *stash_userevent,
91 *stash_memory,
92 *stash_buffer,
93 *stash_bufferobj,
94 *stash_image,
95 *stash_image1d,
96 *stash_image1darray,
97 *stash_image1dbuffer,
98 *stash_image2d,
99 *stash_image2darray,
100 *stash_image3d,
101 *stash_mapped,
102 *stash_mappedbuffer,
103 *stash_mappedimage;
26 104
27/*****************************************************************************/ 105/*****************************************************************************/
28 106
107// name must include a leading underscore
108// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
109// for perl. doh.
110static void *
111glsym (const char *name)
112{
113 void *fun = 0;
114
115 #if defined I_DLFCN && defined RTLD_DEFAULT
116 fun = dlsym (RTLD_DEFAULT, name + 1);
117 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
118
119 if (!fun)
120 {
121 static void *libgl;
122 static const char *glso[] = {
123 "libGL.so.1",
124 "libGL.so.3",
125 "libGL.so.4.0",
126 "libGL.so",
127 "/usr/lib/libGL.so",
128 "/usr/X11R6/lib/libGL.1.dylib"
129 };
130 int i;
131
132 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
133 {
134 libgl = dlopen (glso [i], RTLD_LAZY);
135 if (libgl)
136 break;
137 }
138
139 if (libgl)
140 {
141 fun = dlsym (libgl, name + 1);
142 if (!fun) fun = dlsym (libgl, name);
143 }
144 }
145 #endif
146
147 return fun;
148}
149
150/*****************************************************************************/
151
29/* up to two temporary buffers */ 152/* up to two temporary buffers */
30static void * 153static void * ecb_noinline
31tmpbuf (size_t size) 154tmpbuf (size_t size)
32{ 155{
156 enum { buffers = 4 };
33 static int idx; 157 static int idx;
34 static void *buf [2]; 158 static void *buf [buffers];
35 static size_t len [2]; 159 static size_t len [buffers];
36 160
37 idx ^= 1; 161 idx = (idx + 1) % buffers;
38 162
39 if (len [idx] < size) 163 if (len [idx] < size)
40 { 164 {
41 free (buf [idx]); 165 free (buf [idx]);
42 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 166 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
44 } 168 }
45 169
46 return buf [idx]; 170 return buf [idx];
47} 171}
48 172
173static const char * ecb_noinline
174cv_get_name (CV *cv)
175{
176 static char fullname [256];
177
178 GV *gv = CvGV (cv); // gv better be nonzero
179
180 HV *stash = GvSTASH (gv);
181 const char *hvname = HvNAME_get (stash); // stash also better be nonzero
182 const char *gvname = GvNAME (gv);
183
184 snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname);
185 return fullname;
186}
187
49/*****************************************************************************/ 188/*****************************************************************************/
50 189
51typedef struct 190typedef struct
52{ 191{
53 IV iv; 192 IV iv;
90 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); 229 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)");
91} 230}
92 231
93/*****************************************************************************/ 232/*****************************************************************************/
94 233
95static cl_int last_error; 234static cl_int res;
96 235
97#define FAIL(name,err) \ 236#define FAIL(name) \
98 croak ("cl" # name ": %s", err2str (last_error = err)); 237 croak ("cl" # name ": %s", err2str (res));
99 238
100#define NEED_SUCCESS(name,args) \ 239#define NEED_SUCCESS(name,args) \
101 do { \ 240 do { \
102 cl_int res = cl ## name args; \ 241 res = cl ## name args; \
103 \ 242 \
104 if (res) \ 243 if (res) \
105 FAIL (name, res); \ 244 FAIL (name); \
106 } while (0) 245 } while (0)
107 246
247#define NEED_SUCCESS_ARG(retdecl, name, args) \
248 retdecl = cl ## name args; \
249 if (res) \
250 FAIL (name);
251
108/*****************************************************************************/ 252/*****************************************************************************/
109 253
110#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr)
111#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr))
112
113static void * 254static SV *
255new_clobj (HV *stash, IV id)
256{
257 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash));
258}
259
260#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id)))
261#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
262
263/* cl objects are either \$iv, or [$iv, ...] */
264/* they can be upgraded at runtime to the array form */
265static void * ecb_noinline
114SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) 266SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg)
115{ 267{
268 // sv_derived_from is quite slow :(
116 if (SvROK (sv) && sv_derived_from (sv, pkg)) 269 if (SvROK (sv) && sv_derived_from (sv, pkg))
117 return (void *)SvIV (SvRV (sv)); 270 return (void *)SvIV (SvRV (sv));
118 271
119 croak ("%s: %s is not of type %s", func, svname, pkg); 272 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg);
273}
274
275// the "no-inherit" version of the above
276static void * ecb_noinline
277SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash)
278{
279 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash)
280 return (void *)SvIV (SvRV (sv));
281
282 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash));
120} 283}
121 284
122/*****************************************************************************/ 285/*****************************************************************************/
123 286
124static cl_event * 287static cl_context_properties * ecb_noinline
125event_list (SV **items, int count) 288SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
126{ 289{
290 if (!sv || !SvOK (sv))
291 if (extra)
292 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
293 else
294 return 0;
295
296 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
297 {
298 AV *av = (AV *)SvRV (sv);
299 int i, len = av_len (av) + 1;
300 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
301 cl_context_properties *l = p;
302
303 if (len & 1)
304 croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname);
305
306 while (extracount--)
307 *l++ = *extra++;
308
309 for (i = 0; i < len; i += 2)
310 {
311 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
312 SV *p_sv = *av_fetch (av, i + 1, 0);
313 cl_context_properties v = SvIV (p_sv); // code below can override
314
315 switch (t)
316 {
317 case CL_CONTEXT_PLATFORM:
318 if (SvROK (p_sv))
319 v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform");
320 break;
321
322 case CL_GLX_DISPLAY_KHR:
323 if (!SvOK (p_sv))
324 {
325 void *func = glsym ("_glXGetCurrentDisplay");
326 if (func)
327 v = (cl_context_properties)((void *(*)(void))func)();
328 }
329 break;
330
331 case CL_GL_CONTEXT_KHR:
332 if (!SvOK (p_sv))
333 {
334 void *func = glsym ("_glXGetCurrentContext");
335 if (func)
336 v = (cl_context_properties)((void *(*)(void))func)();
337 }
338 break;
339
340 default:
341 /* unknown property, treat as int */
342 break;
343 }
344
345 *l++ = t;
346 *l++ = v;
347 }
348
349 *l = 0;
350
351 return p;
352 }
353
354 croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname);
355}
356
357// parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation
358// is a pointer (and only on well-behaved systems).
359static void * ecb_noinline
360object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount)
361{
362 if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV)
363 croak ("%s: '%s' parameter must be %sa reference to an array of %s objects",
364 cv_get_name (cv), argname, or_undef ? "undef or " : "", klass);
365
366 AV *av = (AV *)SvRV (arg);
367 void **list = 0;
368 cl_uint count = av_len (av) + 1;
369
370 if (count)
371 {
372 list = tmpbuf (sizeof (*list) * count);
373 int i;
374 for (i = 0; i < count; ++i)
375 list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass);
376 }
377
378 if (!count && !or_undef)
379 croak ("%s: '%s' must contain at least one %s object",
380 cv_get_name (cv), argname, klass);
381
382 *rcount = count;
383 return (void *)list;
384}
385
386/*****************************************************************************/
387/* callback stuff */
388
389/* default context callback, log to stderr */
390static void CL_CALLBACK
391context_default_notify (const char *msg, const void *info, size_t cb, void *data)
392{
393 fprintf (stderr, "OpenCL Context Notify: %s\n", msg);
394}
395
396typedef struct
397{
398 int free_cb;
399 void (*push)(void *data1, void *data2, void *data3);
400} eq_vtbl;
401
402typedef struct eq_item
403{
404 struct eq_item *next;
405 eq_vtbl *vtbl;
406 SV *cb;
407 void *data1, *data2, *data3;
408} eq_item;
409
410static void (*eq_signal_func)(void *signal_arg, int value);
411static void *eq_signal_arg;
412static xmutex_t eq_lock = X_MUTEX_INIT;
413static eq_item *eq_head, *eq_tail;
414
415static void ecb_noinline
416eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3)
417{
418 eq_item *item = malloc (sizeof (eq_item));
419
420 item->next = 0;
421 item->vtbl = vtbl;
422 item->cb = cb;
423 item->data1 = data1;
424 item->data2 = data2;
425 item->data3 = data3;
426
427 X_LOCK (eq_lock);
428
429 *(eq_head ? &eq_tail->next : &eq_head) = item;
430 eq_tail = item;
431
432 X_UNLOCK (eq_lock);
433
434 eq_signal_func (eq_signal_arg, 0);
435}
436
437static eq_item *
438eq_dec (void)
439{
440 eq_item *res;
441
442 X_LOCK (eq_lock);
443
444 res = eq_head;
445 if (res)
446 eq_head = res->next;
447
448 X_UNLOCK (eq_lock);
449
450 return res;
451}
452
453static void
454eq_poll (void)
455{
456 eq_item *item;
457
458 while ((item = eq_dec ()))
459 {
460 ENTER;
461 SAVETMPS;
462
463 dSP;
464 PUSHMARK (SP);
465 EXTEND (SP, 2);
466
467 if (item->vtbl->free_cb)
468 sv_2mortal (item->cb);
469
470 PUTBACK;
471 item->vtbl->push (item->data1, item->data2, item->data3);
472
473 SV *cb = item->cb;
474 free (item);
475
476 call_sv (cb, G_DISCARD | G_VOID);
477
478 FREETMPS;
479 LEAVE;
480 }
481}
482
483static void
484eq_poll_interrupt (pTHX_ void *c_arg, int value)
485{
486 eq_poll ();
487}
488
489/*****************************************************************************/
490/* context notify */
491
492static void ecb_noinline
493eq_context_push (void *data1, void *data2, void *data3)
494{
495 dSP;
496 PUSHs (sv_2mortal (newSVpv (data1, 0)));
497 PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3)));
498 PUTBACK;
499
500 free (data1);
501 free (data2);
502}
503
504static eq_vtbl eq_context_vtbl = { 0, eq_context_push };
505
506static void CL_CALLBACK
507eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data)
508{
509 void *pvt_copy = malloc (cb);
510 memcpy (pvt_copy, pvt, cb);
511 eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb);
512}
513
514#define CONTEXT_NOTIFY_CALLBACK \
515 void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \
516 void *user_data = 0; \
517 \
518 if (SvOK (notify)) \
519 { \
520 pfn_notify = eq_context_notify; \
521 user_data = s_get_cv (notify); \
522 }
523
524static SV * ecb_noinline
525new_clobj_context (cl_context ctx, void *user_data)
526{
527 SV *sv = new_clobj (stash_context, (IV)ctx);
528
529 if (user_data)
530 sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0);
531
532 return sv;
533}
534
535#define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data));
536
537/*****************************************************************************/
538/* build/compile/link notify */
539
540static void
541eq_program_push (void *data1, void *data2, void *data3)
542{
543 dSP;
544 PUSH_CLOBJ (stash_program, data1);
545 PUTBACK;
546}
547
548static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
549
550static void CL_CALLBACK
551eq_program_notify (cl_program program, void *user_data)
552{
553 clRetainProgram (program);
554
555 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
556}
557
558typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data);
559
560static program_callback ecb_noinline
561make_program_callback (SV *notify, void **ruser_data)
562{
563 if (SvOK (notify))
564 {
565 *ruser_data = SvREFCNT_inc (s_get_cv (notify));
566 return eq_program_notify;
567 }
568 else
569 {
570 *ruser_data = 0;
571 return 0;
572 }
573}
574
575struct build_args
576{
577 cl_program program;
578 char *options;
579 void *user_data;
580 cl_uint num_devices;
581};
582
583X_THREAD_PROC (build_program_thread)
584{
585 struct build_args *arg = thr_arg;
586
587 clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0);
588
589 if (arg->user_data)
590 eq_program_notify (arg->program, arg->user_data);
591 else
592 clReleaseProgram (arg->program);
593
594 free (arg->options);
595 free (arg);
596
597 return 0;
598}
599
600static void
601build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data)
602{
603 struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices);
604
605 arg->program = program;
606 arg->options = strdup (options);
607 arg->user_data = user_data;
608 arg->num_devices = num_devices;
609 memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices);
610
611 xthread_t id;
612 thread_create (&id, build_program_thread, arg);
613}
614
615/*****************************************************************************/
616/* event objects */
617
618static void
619eq_event_push (void *data1, void *data2, void *data3)
620{
621 dSP;
622 PUSH_CLOBJ (stash_event, data1);
623 PUSHs (sv_2mortal (newSViv ((IV)data2)));
624 PUTBACK;
625}
626
627static eq_vtbl eq_event_vtbl = { 1, eq_event_push };
628
629static void CL_CALLBACK
630eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data)
631{
632 clRetainEvent (event);
633 eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0);
634}
635
636/*****************************************************************************/
637/* utilities for XS code */
638
639static size_t
640img_row_pitch (cl_mem img)
641{
642 size_t res;
643 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
644 return res;
645}
646
647static cl_event * ecb_noinline
648event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
649{
650 cl_uint count = *rcount;
651
652 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed
653 *rcount = count = 0;
654
655 if (!count && !extra)
656 return 0;
657
127 cl_event *list = tmpbuf (sizeof (cl_event) * count); 658 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
659 int i = 0;
128 660
129 while (count--) 661 while (count--)
662 if (SvOK (items [count]))
130 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 663 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
131 664
665 if (extra)
666 list [i++] = extra;
667
668 *rcount = i;
669
132 return list; 670 return i ? list : 0;
133} 671}
134 672
135#define EVENT_LIST(items,count) \ 673#define EVENT_LIST(skip) \
136 cl_uint event_list_count = (count); \ 674 cl_uint event_list_count = items - (skip); \
137 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 675 cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0)
138 676
139#define INFO(class) \ 677#define INFO(class) \
140{ \ 678{ \
141 size_t size; \ 679 size_t size; \
142 SV *sv; \
143 \
144 NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ 680 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
145 sv = sv_2mortal (newSV (size)); \ 681 SV *sv = sv_2mortal (newSV (size)); \
146 SvUPGRADE (sv, SVt_PV); \ 682 SvUPGRADE (sv, SVt_PV); \
147 SvPOK_only (sv); \ 683 SvPOK_only (sv); \
148 SvCUR_set (sv, size); \ 684 SvCUR_set (sv, size); \
149 NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ 685 NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \
150 XPUSHs (sv); \ 686 XPUSHs (sv); \
151} 687}
152 688
689/*****************************************************************************/
690/* mapped_xxx */
691
692static OpenCL__Mapped
693SvMAPPED (SV *sv)
694{
695 // no typechecking atm., keep your fingers crossed
696 return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr;
697}
698
699struct mapped
700{
701 cl_command_queue queue;
702 cl_mem memobj;
703 void *ptr;
704 size_t cb;
705 cl_event event;
706 size_t row_pitch;
707 size_t slice_pitch;
708};
709
710static SV *
711mapped_new (HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags, void *ptr, size_t cb, cl_event ev, size_t row_pitch, size_t slice_pitch)
712{
713 SV *data = newSV (0);
714 SvUPGRADE (data, SVt_PVMG);
715
716 OpenCL__Mapped mapped;
717 New (0, mapped, 1, struct mapped);
718
719 clRetainCommandQueue (queue);
720
721 mapped->queue = queue;
722 mapped->memobj = memobj;
723 mapped->ptr = ptr;
724 mapped->cb = cb;
725 mapped->event = ev;
726 mapped->row_pitch = row_pitch;
727 mapped->slice_pitch = slice_pitch;
728
729 sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0);
730
731 if (SvLEN (data))
732 Safefree (data);
733
734 SvPVX (data) = (char *)ptr;
735 SvCUR_set (data, cb);
736 SvLEN_set (data, 0);
737 SvPOK_only (data);
738
739 SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash));
740
741 if (!(flags & CL_MAP_WRITE))
742 SvREADONLY_on (data);
743
744 return obj;
745}
746
747static void
748mapped_detach (SV *sv, OpenCL__Mapped mapped)
749{
750 SV *data = SvRV (sv);
751
752 // the next check checks both before AND after detach, where SvPVX should STILL be 0
753 if (SvPVX (data) != (char *)mapped->ptr)
754 warn ("FATAL: OpenCL memory mapped scalar changed location, detected");
755 else
756 {
757 SvREADONLY_off (data);
758 SvCUR_set (data, 0);
759 SvPVX (data) = 0;
760 SvOK_off (data);
761 }
762
763 mapped->ptr = 0;
764}
765
766static void
767mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count)
768{
769 cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event);
770 cl_event ev;
771
772 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev));
773
774 clReleaseEvent (mapped->event);
775 mapped->event = ev;
776
777 mapped_detach (self, mapped);
778}
779
780/*****************************************************************************/
781
153MODULE = OpenCL PACKAGE = OpenCL 782MODULE = OpenCL PACKAGE = OpenCL
154 783
155PROTOTYPES: ENABLE 784PROTOTYPES: ENABLE
156 785
786void
787poll ()
788 CODE:
789 eq_poll ();
790
791void
792_eq_initialise (IV func, IV arg)
793 CODE:
794 eq_signal_func = (void (*)(void *, int))func;
795 eq_signal_arg = (void*)arg;
796
157BOOT: 797BOOT:
158{ 798{
159 HV *stash = gv_stashpv ("OpenCL", 1); 799 HV *stash = gv_stashpv ("OpenCL", 1);
160 static const ivstr *civ, const_iv[] = { 800 static const ivstr *civ, const_iv[] = {
161 { sizeof (cl_char ), "SIZEOF_CHAR" }, 801 { sizeof (cl_char ), "SIZEOF_CHAR" },
162 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 802 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
163 { sizeof (cl_short ), "SIZEOF_SHORT" }, 803 { sizeof (cl_short ), "SIZEOF_SHORT" },
164 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 804 { sizeof (cl_ushort), "SIZEOF_USHORT" },
165 { sizeof (cl_int ), "SIZEOF_INT" }, 805 { sizeof (cl_int ), "SIZEOF_INT" },
166 { sizeof (cl_uint ), "SIZEOF_UINT" }, 806 { sizeof (cl_uint ), "SIZEOF_UINT" },
167 { sizeof (cl_long ), "SIZEOF_LONG" }, 807 { sizeof (cl_long ), "SIZEOF_LONG" },
168 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 808 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
169 { sizeof (cl_half ), "SIZEOF_HALF" }, 809 { sizeof (cl_half ), "SIZEOF_HALF" },
170 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 810 { sizeof (cl_float ), "SIZEOF_FLOAT" },
171 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 811 { sizeof (cl_double), "SIZEOF_DOUBLE" },
812 { PREFER_1_1 , "PREFER_1_1" },
172#include "constiv.h" 813#include "constiv.h"
173 }; 814 };
815
174 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 816 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--)
175 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 817 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
818
819 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
820 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD);
821 stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD);
822 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
823 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD);
824 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD);
825 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD);
826 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD);
827 stash_event = gv_stashpv ("OpenCL::Event", GV_ADD);
828 stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD);
829 stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD);
830 stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD);
831 stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD);
832 stash_image = gv_stashpv ("OpenCL::Image", GV_ADD);
833 stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD);
834 stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD);
835 stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD);
836 stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD);
837 stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD);
838 stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD);
839 stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD);
840 stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD);
841 stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD);
842
843 sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt);
176} 844}
177 845
178cl_int 846cl_int
179errno () 847errno ()
180 CODE: 848 CODE:
181 errno = last_error; 849 RETVAL = res;
850 OUTPUT:
851 RETVAL
182 852
183const char * 853const char *
184err2str (cl_int err) 854err2str (cl_int err = res)
185 855
186const char * 856const char *
187enum2str (cl_uint value) 857enum2str (cl_uint value)
188 858
189void 859void
190platforms () 860platforms ()
191 PPCODE: 861 PPCODE:
192{
193 cl_platform_id *list; 862 cl_platform_id *list;
194 cl_uint count; 863 cl_uint count;
195 int i; 864 int i;
196 865
197 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); 866 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count));
198 list = tmpbuf (sizeof (*list) * count); 867 list = tmpbuf (sizeof (*list) * count);
199 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 868 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
200 869
201 EXTEND (SP, count); 870 EXTEND (SP, count);
202 for (i = 0; i < count; ++i) 871 for (i = 0; i < count; ++i)
203 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 872 PUSH_CLOBJ (stash_platform, list [i]);
204}
205 873
206void 874void
207context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 875context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
208 PPCODE: 876 PPCODE:
209{ 877 CONTEXT_NOTIFY_CALLBACK;
210 cl_int res; 878 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
211 cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); 879 XPUSH_CLOBJ_CONTEXT;
212 880
213 if (res) 881void
214 FAIL (CreateContextFromType, res); 882context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
883 PPCODE:
884 cl_uint device_count;
885 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
215 886
216 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 887 CONTEXT_NOTIFY_CALLBACK;
217} 888 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res));
889 XPUSH_CLOBJ_CONTEXT;
218 890
219void 891void
220wait_for_events (...) 892wait_for_events (...)
221 CODE: 893 CODE:
222{
223 EVENT_LIST (0, items); 894 EVENT_LIST (0);
224 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); 895 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr));
225}
226 896
227PROTOTYPES: DISABLE 897PROTOTYPES: DISABLE
228 898
229MODULE = OpenCL PACKAGE = OpenCL::Platform 899MODULE = OpenCL PACKAGE = OpenCL::Platform
230 900
231void 901void
232info (OpenCL::Platform this, cl_platform_info name) 902info (OpenCL::Platform self, cl_platform_info name)
233 PPCODE: 903 PPCODE:
234 INFO (Platform) 904 INFO (Platform)
235 905
236void 906void
907unload_compiler (OpenCL::Platform self)
908 CODE:
909#if CL_VERSION_1_2
910 clUnloadPlatformCompiler (self);
911#endif
912
913#BEGIN:platform
914
915void
916profile (OpenCL::Platform self)
917 ALIAS:
918 profile = CL_PLATFORM_PROFILE
919 version = CL_PLATFORM_VERSION
920 name = CL_PLATFORM_NAME
921 vendor = CL_PLATFORM_VENDOR
922 extensions = CL_PLATFORM_EXTENSIONS
923 PPCODE:
924 size_t size;
925 NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size));
926 char *value = tmpbuf (size);
927 NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0));
928 EXTEND (SP, 1);
929 const int i = 0;
930 PUSHs (sv_2mortal (newSVpv (value, 0)));
931
932#END:platform
933
934void
237devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) 935devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL)
238 PPCODE: 936 PPCODE:
239{
240 cl_device_id *list; 937 cl_device_id *list;
241 cl_uint count; 938 cl_uint count;
242 int i; 939 int i;
243 940
244 NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); 941 NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count));
245 list = tmpbuf (sizeof (*list) * count); 942 list = tmpbuf (sizeof (*list) * count);
246 NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); 943 NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0));
247 944
248 EXTEND (SP, count); 945 EXTEND (SP, count);
249 for (i = 0; i < count; ++i) 946 for (i = 0; i < count; ++i)
250 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 947 PUSH_CLOBJ (stash_device, list [i]);
251}
252 948
253void 949void
254context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 950context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
255 PPCODE: 951 PPCODE:
256{
257 cl_int res;
258 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; 952 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
259 cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); 953 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
260 954
261 if (res) 955 cl_uint device_count;
262 FAIL (CreateContextFromType, res); 956 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
263 957
264 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 958 CONTEXT_NOTIFY_CALLBACK;
265} 959 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
960 XPUSH_CLOBJ_CONTEXT;
961
962void
963context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
964 PPCODE:
965 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
966 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
967
968 CONTEXT_NOTIFY_CALLBACK;
969 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res));
970 XPUSH_CLOBJ_CONTEXT;
266 971
267MODULE = OpenCL PACKAGE = OpenCL::Device 972MODULE = OpenCL PACKAGE = OpenCL::Device
268 973
269void 974void
270info (OpenCL::Device this, cl_device_info name) 975info (OpenCL::Device self, cl_device_info name)
271 PPCODE: 976 PPCODE:
272 INFO (Device) 977 INFO (Device)
273 978
979#if CL_VERSION_1_2
980
274void 981void
275context_simple (OpenCL::Device this) 982sub_devices (OpenCL::Device self, SV *properties)
983 PPCODE:
984 if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV)
985 croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs");
986
987 properties = SvRV (properties);
988
989 cl_uint count = av_len ((AV *)properties) + 1;
990 cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1);
991
992 int i;
993 for (i = 0; i < count; ++i)
994 props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0));
995
996 props [count] = 0;
997
998 cl_uint num_devices;
999 NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices));
1000 cl_device_id *list = tmpbuf (sizeof (*list) * num_devices);
1001 NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0));
1002
1003 EXTEND (SP, num_devices);
1004 for (i = 0; i < count; ++i)
1005 PUSH_CLOBJ (stash_subdevice, list [i]);
1006
1007#endif
1008
1009#BEGIN:device
1010
1011void
1012type (OpenCL::Device self)
276 PPCODE: 1013 PPCODE:
277{ 1014 cl_device_type value [1];
278 cl_int res; 1015 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0));
279 cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); 1016 EXTEND (SP, 1);
1017 const int i = 0;
1018 PUSHs (sv_2mortal (newSVuv (value [i])));
280 1019
281 if (res) 1020void
282 FAIL (CreateContext, res); 1021vendor_id (OpenCL::Device self)
1022 ALIAS:
1023 vendor_id = CL_DEVICE_VENDOR_ID
1024 max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS
1025 max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS
1026 preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR
1027 preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT
1028 preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT
1029 preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG
1030 preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT
1031 preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE
1032 max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY
1033 max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS
1034 max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS
1035 image_support = CL_DEVICE_IMAGE_SUPPORT
1036 max_samplers = CL_DEVICE_MAX_SAMPLERS
1037 mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN
1038 min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE
1039 global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE
1040 max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS
1041 preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF
1042 native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR
1043 native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT
1044 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
1045 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
1046 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
1047 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
1048 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
1049 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
1050 PPCODE:
1051 cl_uint value [1];
1052 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1053 EXTEND (SP, 1);
1054 const int i = 0;
1055 PUSHs (sv_2mortal (newSVuv (value [i])));
283 1056
284 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 1057void
285} 1058max_work_group_size (OpenCL::Device self)
1059 ALIAS:
1060 max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE
1061 image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH
1062 image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT
1063 image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH
1064 image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT
1065 image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH
1066 max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE
1067 profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION
1068 PPCODE:
1069 size_t value [1];
1070 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1071 EXTEND (SP, 1);
1072 const int i = 0;
1073 PUSHs (sv_2mortal (newSVuv (value [i])));
1074
1075void
1076max_work_item_sizes (OpenCL::Device self)
1077 PPCODE:
1078 size_t size;
1079 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size));
1080 size_t *value = tmpbuf (size);
1081 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0));
1082 int i, n = size / sizeof (*value);
1083 EXTEND (SP, n);
1084 for (i = 0; i < n; ++i)
1085 PUSHs (sv_2mortal (newSVuv (value [i])));
1086
1087void
1088address_bits (OpenCL::Device self)
1089 PPCODE:
1090 cl_bitfield value [1];
1091 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_ADDRESS_BITS, sizeof (value), value, 0));
1092 EXTEND (SP, 1);
1093 const int i = 0;
1094 PUSHs (sv_2mortal (newSVuv (value [i])));
1095
1096void
1097max_mem_alloc_size (OpenCL::Device self)
1098 ALIAS:
1099 max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE
1100 global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE
1101 global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE
1102 max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE
1103 local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE
1104 PPCODE:
1105 cl_ulong value [1];
1106 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1107 EXTEND (SP, 1);
1108 const int i = 0;
1109 PUSHs (sv_2mortal (newSVuv (value [i])));
1110
1111void
1112single_fp_config (OpenCL::Device self)
1113 ALIAS:
1114 single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG
1115 double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG
1116 half_fp_config = CL_DEVICE_HALF_FP_CONFIG
1117 PPCODE:
1118 cl_device_fp_config value [1];
1119 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1120 EXTEND (SP, 1);
1121 const int i = 0;
1122 PUSHs (sv_2mortal (newSVuv (value [i])));
1123
1124void
1125global_mem_cache_type (OpenCL::Device self)
1126 PPCODE:
1127 cl_device_mem_cache_type value [1];
1128 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_GLOBAL_MEM_CACHE_TYPE, sizeof (value), value, 0));
1129 EXTEND (SP, 1);
1130 const int i = 0;
1131 PUSHs (sv_2mortal (newSVuv (value [i])));
1132
1133void
1134local_mem_type (OpenCL::Device self)
1135 PPCODE:
1136 cl_device_local_mem_type value [1];
1137 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_LOCAL_MEM_TYPE, sizeof (value), value, 0));
1138 EXTEND (SP, 1);
1139 const int i = 0;
1140 PUSHs (sv_2mortal (newSVuv (value [i])));
1141
1142void
1143error_correction_support (OpenCL::Device self)
1144 ALIAS:
1145 error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT
1146 endian_little = CL_DEVICE_ENDIAN_LITTLE
1147 available = CL_DEVICE_AVAILABLE
1148 compiler_available = CL_DEVICE_COMPILER_AVAILABLE
1149 host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY
1150 PPCODE:
1151 cl_bool value [1];
1152 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1153 EXTEND (SP, 1);
1154 const int i = 0;
1155 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
1156
1157void
1158execution_capabilities (OpenCL::Device self)
1159 PPCODE:
1160 cl_device_exec_capabilities value [1];
1161 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_EXECUTION_CAPABILITIES, sizeof (value), value, 0));
1162 EXTEND (SP, 1);
1163 const int i = 0;
1164 PUSHs (sv_2mortal (newSVuv (value [i])));
1165
1166void
1167properties (OpenCL::Device self)
1168 PPCODE:
1169 cl_command_queue_properties value [1];
1170 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0));
1171 EXTEND (SP, 1);
1172 const int i = 0;
1173 PUSHs (sv_2mortal (newSVuv (value [i])));
1174
1175void
1176platform (OpenCL::Device self)
1177 PPCODE:
1178 cl_platform_id value [1];
1179 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0));
1180 EXTEND (SP, 1);
1181 const int i = 0;
1182 {
1183 PUSH_CLOBJ (stash_platform, value [i]);
1184 }
1185
1186void
1187name (OpenCL::Device self)
1188 ALIAS:
1189 name = CL_DEVICE_NAME
1190 vendor = CL_DEVICE_VENDOR
1191 driver_version = CL_DRIVER_VERSION
1192 profile = CL_DEVICE_PROFILE
1193 version = CL_DEVICE_VERSION
1194 extensions = CL_DEVICE_EXTENSIONS
1195 PPCODE:
1196 size_t size;
1197 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1198 char *value = tmpbuf (size);
1199 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1200 EXTEND (SP, 1);
1201 const int i = 0;
1202 PUSHs (sv_2mortal (newSVpv (value, 0)));
1203
1204void
1205parent_device_ext (OpenCL::Device self)
1206 PPCODE:
1207 cl_device_id value [1];
1208 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0));
1209 EXTEND (SP, 1);
1210 const int i = 0;
1211 {
1212 PUSH_CLOBJ (stash_device, value [i]);
1213 }
1214
1215void
1216partition_types_ext (OpenCL::Device self)
1217 ALIAS:
1218 partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT
1219 affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT
1220 partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT
1221 PPCODE:
1222 size_t size;
1223 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1224 cl_device_partition_property_ext *value = tmpbuf (size);
1225 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1226 int i, n = size / sizeof (*value);
1227 EXTEND (SP, n);
1228 for (i = 0; i < n; ++i)
1229 PUSHs (sv_2mortal (newSVuv (value [i])));
1230
1231#END:device
1232
1233MODULE = OpenCL PACKAGE = OpenCL::SubDevice
1234
1235#if CL_VERSION_1_2
1236
1237void
1238DESTROY (OpenCL::SubDevice self)
1239 CODE:
1240 clReleaseDevice (self);
1241
1242#endif
286 1243
287MODULE = OpenCL PACKAGE = OpenCL::Context 1244MODULE = OpenCL PACKAGE = OpenCL::Context
288 1245
289void 1246void
290DESTROY (OpenCL::Context context) 1247DESTROY (OpenCL::Context self)
291 CODE: 1248 CODE:
292 clReleaseContext (context); 1249 clReleaseContext (self);
293 1250
294void 1251void
295info (OpenCL::Context this, cl_context_info name) 1252info (OpenCL::Context self, cl_context_info name)
296 PPCODE: 1253 PPCODE:
297 INFO (Context) 1254 INFO (Context)
298 1255
299void 1256void
300queue (OpenCL::Context this, OpenCL::Device device, cl_command_queue_properties properties) 1257queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0)
301 PPCODE: 1258 PPCODE:
302{
303 cl_int res;
304 cl_command_queue queue = clCreateCommandQueue (this, device, properties, &res); 1259 NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res));
1260 XPUSH_CLOBJ (stash_queue, queue);
305 1261
306 if (res)
307 FAIL (CreateCommandQueue, res);
308
309 XPUSH_NEW_OBJ ("OpenCL::Queue", queue);
310}
311
312void 1262void
313user_event (OpenCL::Context this) 1263user_event (OpenCL::Context self)
314 PPCODE: 1264 PPCODE:
315{ 1265 NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res));
316 cl_int res; 1266 XPUSH_CLOBJ (stash_userevent, ev);
317 cl_event ev = clCreateUserEvent (this, &res);
318 1267
319 if (res)
320 FAIL (CreateUserevent, res);
321
322 XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev);
323}
324
325void 1268void
326buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) 1269buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
327 PPCODE: 1270 PPCODE:
328{
329 cl_int res;
330 cl_mem mem;
331
332 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 1271 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))
333 croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); 1272 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
334 1273
335 mem = clCreateBuffer (this, flags, len, 0, &res); 1274 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
1275 XPUSH_CLOBJ (stash_bufferobj, mem);
336 1276
337 if (res)
338 FAIL (CreateBuffer, res);
339
340 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
341}
342
343void 1277void
344buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) 1278buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
345 PPCODE: 1279 PPCODE:
346{
347 STRLEN len; 1280 STRLEN len;
348 char *ptr = SvPVbyte (data, len); 1281 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
349 cl_int res;
350 cl_mem mem;
351
352 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 1282 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
353 croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); 1283 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
354 1284 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
355 mem = clCreateBuffer (this, flags, len, ptr, &res); 1285 XPUSH_CLOBJ (stash_bufferobj, mem);
356 1286
357 if (res) 1287#if CL_VERSION_1_2
358 FAIL (CreateBuffer, res);
359 1288
360 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
361}
362
363void 1289void
364image2d (OpenCL::Context this, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, SV *data) 1290image (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, cl_mem_object_type type, size_t width, size_t height, size_t depth = 0, size_t array_size = 0, size_t row_pitch = 0, size_t slice_pitch = 0, cl_uint num_mip_level = 0, cl_uint num_samples = 0, SV *data = &PL_sv_undef)
365 PPCODE: 1291 PPCODE:
366{
367 STRLEN len; 1292 STRLEN len;
368 char *ptr = SvPVbyte (data, len); 1293 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
369 const cl_image_format format = { channel_order, channel_type }; 1294 const cl_image_format format = { channel_order, channel_type };
370 cl_int res; 1295 const cl_image_desc desc = {
371 cl_mem mem = clCreateImage2D (this, flags, &format, width, height, len / height, ptr, &res); 1296 type,
1297 width, height, depth,
1298 array_size, row_pitch, slice_pitch,
1299 num_mip_level, num_samples,
1300 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0
1301 };
1302 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1303 HV *stash = stash_image;
1304 switch (type)
1305 {
1306 case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break;
1307 case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break;
1308 case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break;
1309 case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break;
1310 case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break;
1311 case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break;
1312 }
1313 XPUSH_CLOBJ (stash, mem);
372 1314
373 if (res) 1315#endif
374 FAIL (CreateImage2D, res);
375 1316
376 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
377}
378
379void 1317void
380image3d (OpenCL::Context this, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t depth, size_t slice_pitch, SV *data) 1318image2d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t row_pitch = 0, SV *data = &PL_sv_undef)
381 PPCODE: 1319 PPCODE:
382{
383 STRLEN len; 1320 STRLEN len;
384 char *ptr = SvPVbyte (data, len); 1321 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
385 const cl_image_format format = { channel_order, channel_type }; 1322 const cl_image_format format = { channel_order, channel_type };
386 cl_int res; 1323#if PREFER_1_1
387 cl_mem mem = clCreateImage3D (this, flags, &format, width, height, 1324 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
388 depth, len / (height * slice_pitch), slice_pitch, ptr, &res); 1325#else
1326 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
1327 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1328#endif
1329 XPUSH_CLOBJ (stash_image2d, mem);
389 1330
390 if (res)
391 FAIL (CreateImage3D, res);
392
393 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
394}
395
396void 1331void
1332image3d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t depth, size_t row_pitch = 0, size_t slice_pitch = 0, SV *data = &PL_sv_undef)
1333 PPCODE:
1334 STRLEN len;
1335 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1336 const cl_image_format format = { channel_order, channel_type };
1337#if PREFER_1_1
1338 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
1339#else
1340 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
1341 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1342#endif
1343 XPUSH_CLOBJ (stash_image3d, mem);
1344
1345#if cl_apple_gl_sharing || cl_khr_gl_sharing
1346
1347void
1348gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
1349 PPCODE:
1350 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
1351 XPUSH_CLOBJ (stash_bufferobj, mem);
1352
1353void
1354gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
1355 PPCODE:
1356 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
1357 XPUSH_CLOBJ (stash_image2d, mem);
1358
1359#if CL_VERSION_1_2
1360
1361void
1362gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1363 ALIAS:
1364 PPCODE:
1365 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
1366 cl_gl_object_type type;
1367 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
1368 HV *stash = stash_memory;
1369 switch (type)
1370 {
1371 case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break;
1372 case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break;
1373 case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break;
1374 case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break;
1375 case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break;
1376 case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break;
1377 }
1378 XPUSH_CLOBJ (stash, mem);
1379
1380#endif
1381
1382void
1383gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1384 PPCODE:
1385#if PREFER_1_1
1386 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
1387#else
1388 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1389#endif
1390 XPUSH_CLOBJ (stash_image2d, mem);
1391
1392void
1393gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1394 PPCODE:
1395#if PREFER_1_1
1396 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
1397#else
1398 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1399#endif
1400 XPUSH_CLOBJ (stash_image3d, mem);
1401
1402#endif
1403
1404void
397supported_image_formats (OpenCL::Context this, cl_mem_flags flags, cl_mem_object_type image_type) 1405supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
398 PPCODE: 1406 PPCODE:
399{ 1407{
400 cl_uint count; 1408 cl_uint count;
401 cl_image_format *list; 1409 cl_image_format *list;
402 int i; 1410 int i;
403 1411
404 NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, 0, 0, &count)); 1412 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count));
405 Newx (list, count, cl_image_format); 1413 Newx (list, count, cl_image_format);
406 NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, count, list, 0)); 1414 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0));
407 1415
408 EXTEND (SP, count); 1416 EXTEND (SP, count);
409 for (i = 0; i < count; ++i) 1417 for (i = 0; i < count; ++i)
410 { 1418 {
411 AV *av = newAV (); 1419 AV *av = newAV ();
414 PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); 1422 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
415 } 1423 }
416} 1424}
417 1425
418void 1426void
419sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) 1427sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode)
420 PPCODE: 1428 PPCODE:
421{
422 cl_int res;
423 cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); 1429 NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res));
424
425 if (res)
426 FAIL (CreateSampler, res);
427
428 XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); 1430 XPUSH_CLOBJ (stash_sampler, sampler);
429}
430 1431
431void 1432void
432program_with_source (OpenCL::Context this, SV *program) 1433program_with_source (OpenCL::Context self, SV *program)
433 PPCODE: 1434 PPCODE:
434{
435 STRLEN len; 1435 STRLEN len;
436 size_t len2; 1436 size_t len2;
437 const char *ptr = SvPVbyte (program, len); 1437 const char *ptr = SvPVbyte (program, len);
438 cl_int res;
439 cl_program prog;
440 1438
441 len2 = len; 1439 len2 = len;
442 prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); 1440 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res));
443
444 if (res)
445 FAIL (CreateProgramWithSource, res);
446
447 XPUSH_NEW_OBJ ("OpenCL::Program", prog); 1441 XPUSH_CLOBJ (stash_program, prog);
448} 1442
1443void
1444program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1445 PPCODE:
1446 cl_uint device_count;
1447 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1448
1449 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1450 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1451
1452 binaries = SvRV (binaries);
1453
1454 if (device_count != av_len ((AV *)binaries) + 1)
1455 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1456
1457 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1458 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1459 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1460
1461 int i;
1462 for (i = 0; i < device_count; ++i)
1463 {
1464 STRLEN len;
1465 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1466 length_list [i] = len;
1467 }
1468
1469 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list,
1470 length_list, binary_list,
1471 GIMME_V == G_ARRAY ? status_list : 0, &res));
1472
1473 EXTEND (SP, 2);
1474 PUSH_CLOBJ (stash_program, prog);
1475
1476 if (GIMME_V == G_ARRAY)
1477 {
1478 AV *av = newAV ();
1479 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1480
1481 for (i = device_count; i--; )
1482 av_store (av, i, newSViv (status_list [i]));
1483 }
1484
1485#if CL_VERSION_1_2
1486
1487void
1488program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names)
1489 PPCODE:
1490 cl_uint device_count;
1491 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1492
1493 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res));
1494
1495 XPUSH_CLOBJ (stash_program, prog);
1496
1497void
1498link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef)
1499 CODE:
1500 cl_uint device_count = 0;
1501 cl_device_id *device_list = 0;
1502
1503 if (SvOK (devices))
1504 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
1505
1506 cl_uint program_count;
1507 cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count);
1508
1509 void *user_data;
1510 program_callback pfn_notify = make_program_callback (notify, &user_data);
1511
1512 NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
1513 program_count, program_list, pfn_notify, user_data, &res));
1514
1515 XPUSH_CLOBJ (stash_program, prog);
1516
1517#endif
1518
1519#BEGIN:context
1520
1521void
1522reference_count (OpenCL::Context self)
1523 ALIAS:
1524 reference_count = CL_CONTEXT_REFERENCE_COUNT
1525 num_devices = CL_CONTEXT_NUM_DEVICES
1526 PPCODE:
1527 cl_uint value [1];
1528 NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0));
1529 EXTEND (SP, 1);
1530 const int i = 0;
1531 PUSHs (sv_2mortal (newSVuv (value [i])));
1532
1533void
1534devices (OpenCL::Context self)
1535 PPCODE:
1536 size_t size;
1537 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size));
1538 cl_device_id *value = tmpbuf (size);
1539 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0));
1540 int i, n = size / sizeof (*value);
1541 EXTEND (SP, n);
1542 for (i = 0; i < n; ++i)
1543 {
1544 PUSH_CLOBJ (stash_device, value [i]);
1545 }
1546
1547void
1548properties (OpenCL::Context self)
1549 PPCODE:
1550 size_t size;
1551 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size));
1552 cl_context_properties *value = tmpbuf (size);
1553 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0));
1554 int i, n = size / sizeof (*value);
1555 EXTEND (SP, n);
1556 for (i = 0; i < n; ++i)
1557 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
1558
1559#END:context
449 1560
450MODULE = OpenCL PACKAGE = OpenCL::Queue 1561MODULE = OpenCL PACKAGE = OpenCL::Queue
451 1562
452void 1563void
453DESTROY (OpenCL::Queue this) 1564DESTROY (OpenCL::Queue self)
454 CODE: 1565 CODE:
455 clReleaseCommandQueue (this); 1566 clReleaseCommandQueue (self);
456 1567
457void 1568void
458info (OpenCL::Queue this, cl_command_queue_info name)
459 PPCODE:
460 INFO (CommandQueue)
461
462void
463enqueue_read_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) 1569read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1570 ALIAS:
1571 enqueue_read_buffer = 0
464 PPCODE: 1572 PPCODE:
465{
466 cl_event ev = 0;
467 EVENT_LIST (6, items - 6); 1573 EVENT_LIST (6);
468 1574
469 SvUPGRADE (data, SVt_PV); 1575 SvUPGRADE (data, SVt_PV);
470 SvGROW (data, len); 1576 SvGROW (data, len);
471 SvPOK_only (data); 1577 SvPOK_only (data);
472 SvCUR_set (data, len); 1578 SvCUR_set (data, len);
1579
1580 cl_event ev = 0;
473 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1581 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
474 1582
475 if (ev) 1583 if (ev)
476 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1584 XPUSH_CLOBJ (stash_event, ev);
477}
478 1585
479void 1586void
480enqueue_write_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) 1587write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1588 ALIAS:
1589 enqueue_write_buffer = 0
481 PPCODE: 1590 PPCODE:
482{ 1591 EVENT_LIST (5);
483 cl_event ev = 0; 1592
484 STRLEN len; 1593 STRLEN len;
485 char *ptr = SvPVbyte (data, len); 1594 char *ptr = SvPVbyte (data, len);
486 EVENT_LIST (5, items - 5);
487 1595
1596 cl_event ev = 0;
488 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1597 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
489 1598
490 if (ev) 1599 if (ev)
491 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1600 XPUSH_CLOBJ (stash_event, ev);
492}
493 1601
1602#if CL_VERSION_1_2
1603
494void 1604void
495enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1605fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1606 ALIAS:
1607 enqueue_fill_buffer = 0
496 PPCODE: 1608 PPCODE:
497{ 1609 EVENT_LIST (5);
1610
1611 STRLEN len;
1612 char *ptr = SvPVbyte (data, len);
1613
498 cl_event ev = 0; 1614 cl_event ev = 0;
499 EVENT_LIST (6, items - 6);
500
501 NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1615 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
502 1616
503 if (ev) 1617 if (ev)
504 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1618 XPUSH_CLOBJ (stash_event, ev);
505}
506 1619
507 /*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueReadBufferRect.html */
508 /*TODO http://www.khronos.org/registry/cl/sdk/1.1/docs/man/xhtml/clEnqueueWriteBufferRect.html */
509
510void 1620void
511enqueue_read_image (OpenCL::Queue this, 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, ...) 1621fill_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, ...)
1622 ALIAS:
1623 enqueue_fill_image = 0
512 PPCODE: 1624 PPCODE:
513{ 1625 EVENT_LIST (12);
1626
1627 const size_t origin [3] = { x, y, z };
1628 const size_t region [3] = { width, height, depth };
1629
1630 const cl_float c_f [4] = { r, g, b, a };
1631 const cl_uint c_u [4] = { r, g, b, a };
1632 const cl_int c_s [4] = { r, g, b, a };
1633 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1634 static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 };
1635 cl_image_format format;
1636 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1637 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1638 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1639 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1640
514 cl_event ev = 0; 1641 cl_event ev = 0;
1642 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1643 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1644
1645 if (ev)
1646 XPUSH_CLOBJ (stash_event, ev);
1647
1648#endif
1649
1650void
1651copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1652 ALIAS:
1653 enqueue_copy_buffer = 0
1654 PPCODE:
1655 EVENT_LIST (6);
1656
1657 cl_event ev = 0;
1658 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1659
1660 if (ev)
1661 XPUSH_CLOBJ (stash_event, ev);
1662
1663void
1664read_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, ...)
1665 ALIAS:
1666 enqueue_read_buffer_rect = 0
1667 PPCODE:
1668 EVENT_LIST (17);
1669
1670 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
515 const size_t src_origin[3] = { src_x, src_y, src_z }; 1671 const size_t host_origin[3] = { host_x, host_y, host_z };
516 const size_t region[3] = { width, height, depth }; 1672 const size_t region[3] = { width, height, depth };
1673
1674 if (!buf_row_pitch)
1675 buf_row_pitch = region [0];
1676
1677 if (!buf_slice_pitch)
1678 buf_slice_pitch = region [1] * buf_row_pitch;
1679
1680 if (!host_row_pitch)
1681 host_row_pitch = region [0];
1682
1683 if (!host_slice_pitch)
1684 host_slice_pitch = region [1] * host_row_pitch;
1685
517 size_t len = row_pitch * slice_pitch * depth; 1686 size_t len = host_row_pitch * host_slice_pitch * region [2];
518 EVENT_LIST (11, items - 11);
519 1687
520 SvUPGRADE (data, SVt_PV); 1688 SvUPGRADE (data, SVt_PV);
521 SvGROW (data, len); 1689 SvGROW (data, len);
522 SvPOK_only (data); 1690 SvPOK_only (data);
523 SvCUR_set (data, len); 1691 SvCUR_set (data, len);
524 NEED_SUCCESS (EnqueueReadImage, (this, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1692
1693 cl_event ev = 0;
1694 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));
525 1695
526 if (ev) 1696 if (ev)
527 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1697 XPUSH_CLOBJ (stash_event, ev);
528}
529 1698
530void 1699void
531enqueue_write_image (OpenCL::Queue this, 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, SV *data, ...) 1700write_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, ...)
1701 ALIAS:
1702 enqueue_write_buffer_rect = 0
532 PPCODE: 1703 PPCODE:
533{ 1704 EVENT_LIST (17);
1705
1706 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1707 const size_t host_origin[3] = { host_x, host_y, host_z };
1708 const size_t region[3] = { width, height, depth };
1709 STRLEN len;
1710 char *ptr = SvPVbyte (data, len);
1711
1712 if (!buf_row_pitch)
1713 buf_row_pitch = region [0];
1714
1715 if (!buf_slice_pitch)
1716 buf_slice_pitch = region [1] * buf_row_pitch;
1717
1718 if (!host_row_pitch)
1719 host_row_pitch = region [0];
1720
1721 if (!host_slice_pitch)
1722 host_slice_pitch = region [1] * host_row_pitch;
1723
1724 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
1725
1726 if (len < min_len)
1727 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1728
534 cl_event ev = 0; 1729 cl_event ev = 0;
1730 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));
1731
1732 if (ev)
1733 XPUSH_CLOBJ (stash_event, ev);
1734
1735void
1736copy_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, ...)
1737 ALIAS:
1738 enqueue_copy_buffer_rect = 0
1739 PPCODE:
1740 EVENT_LIST (16);
1741
1742 const size_t src_origin[3] = { src_x, src_y, src_z };
1743 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1744 const size_t region[3] = { width, height, depth };
1745
1746 cl_event ev = 0;
1747 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));
1748
1749 if (ev)
1750 XPUSH_CLOBJ (stash_event, ev);
1751
1752void
1753read_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, ...)
1754 ALIAS:
1755 enqueue_read_image = 0
1756 PPCODE:
1757 EVENT_LIST (12);
1758
1759 const size_t src_origin[3] = { src_x, src_y, src_z };
1760 const size_t region[3] = { width, height, depth };
1761
1762 if (!row_pitch)
1763 row_pitch = img_row_pitch (src);
1764
1765 if (depth > 1 && !slice_pitch)
1766 slice_pitch = row_pitch * height;
1767
1768 size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1769
1770 SvUPGRADE (data, SVt_PV);
1771 SvGROW (data, len);
1772 SvPOK_only (data);
1773 SvCUR_set (data, len);
1774
1775 cl_event ev = 0;
1776 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));
1777
1778 if (ev)
1779 XPUSH_CLOBJ (stash_event, ev);
1780
1781void
1782write_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, ...)
1783 ALIAS:
1784 enqueue_write_image = 0
1785 PPCODE:
1786 EVENT_LIST (12);
1787
535 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1788 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
536 const size_t region[3] = { width, height, depth }; 1789 const size_t region[3] = { width, height, depth };
537 STRLEN len; 1790 STRLEN len;
538 char *ptr = SvPVbyte (data, len); 1791 char *ptr = SvPVbyte (data, len);
1792
1793 if (!row_pitch)
1794 row_pitch = img_row_pitch (dst);
1795
1796 if (depth > 1 && !slice_pitch)
539 size_t slice_pitch = len / (row_pitch * height); 1797 slice_pitch = row_pitch * height;
540 EVENT_LIST (11, items - 11);
541 1798
1799 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1800
1801 if (len < min_len)
1802 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1803
1804 cl_event ev = 0;
542 NEED_SUCCESS (EnqueueWriteImage, (this, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1805 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));
543 1806
544 if (ev) 1807 if (ev)
545 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1808 XPUSH_CLOBJ (stash_event, ev);
546}
547 1809
548void 1810void
549enqueue_copy_buffer_rect (OpenCL::Queue this, 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, ...) 1811copy_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, ...)
1812 ALIAS:
1813 enqueue_copy_image = 0
550 PPCODE: 1814 PPCODE:
551{ 1815 EVENT_LIST (12);
552 cl_event ev = 0; 1816
553 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 };
554 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1818 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
555 const size_t region[3] = { width, height, depth }; 1819 const size_t region[3] = { width, height, depth };
556 EVENT_LIST (16, items - 16);
557 1820
1821 cl_event ev = 0;
558 NEED_SUCCESS (EnqueueCopyBufferRect, (this, 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)); 1822 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
559 1823
560 if (ev) 1824 if (ev)
561 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1825 XPUSH_CLOBJ (stash_event, ev);
562}
563 1826
564void 1827void
565enqueue_copy_buffer_to_image (OpenCL::Queue this, 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, ...) 1828copy_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, ...)
1829 ALIAS:
1830 enqueue_copy_image_to_buffer = 0
566 PPCODE: 1831 PPCODE:
567{ 1832 EVENT_LIST (10);
1833
1834 const size_t src_origin[3] = { src_x, src_y, src_z };
1835 const size_t region [3] = { width, height, depth };
1836
568 cl_event ev = 0; 1837 cl_event ev = 0;
1838 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1839
1840 if (ev)
1841 XPUSH_CLOBJ (stash_event, ev);
1842
1843void
1844copy_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, ...)
1845 ALIAS:
1846 enqueue_copy_buffer_to_image = 0
1847 PPCODE:
1848 EVENT_LIST (10);
1849
569 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1850 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1851 const size_t region [3] = { width, height, depth };
1852
1853 cl_event ev = 0;
1854 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1855
1856 if (ev)
1857 XPUSH_CLOBJ (stash_event, ev);
1858
1859void
1860map_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, ...)
1861 ALIAS:
1862 enqueue_map_buffer = 0
1863 PPCODE:
1864 EVENT_LIST (6);
1865
1866 size_t cb = SvIV (cb_);
1867
1868 if (!SvOK (cb_))
1869 {
1870 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1871 cb -= offset;
1872 }
1873
1874 cl_event ev;
1875 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
1876 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0));
1877
1878void
1879map_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, ...)
1880 ALIAS:
1881 enqueue_map_image = 0
1882 PPCODE:
1883 size_t width = SvIV (width_);
1884 if (!SvOK (width_))
1885 {
1886 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0));
1887 width -= x;
1888 }
1889
1890 size_t height = SvIV (width_);
1891 if (!SvOK (height_))
1892 {
1893 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0));
1894 height -= y;
1895 }
1896
1897 size_t depth = SvIV (width_);
1898 if (!SvOK (depth_))
1899 {
1900 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0));
1901 depth -= z;
1902
1903 // stupid opencl returns 0 for depth, but requires 1 for 2d images
1904 if (!depth)
1905 depth = 1;
1906 }
1907
1908 const size_t origin[3] = { x, y, z };
570 const size_t region[3] = { width, height, depth }; 1909 const size_t region[3] = { width, height, depth };
1910 size_t row_pitch, slice_pitch;
571 EVENT_LIST (10, items - 10); 1911 EVENT_LIST (10);
572 1912
573 NEED_SUCCESS (EnqueueCopyBufferToImage, (this, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1913 cl_event ev;
1914 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));
1915
1916 size_t cb = slice_pitch ? slice_pitch * region [2]
1917 : row_pitch ? row_pitch * region [1]
1918 : region [0];
1919
1920 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch));
1921
1922void
1923unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1924 PPCODE:
1925 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1926 if (GIMME_V != G_VOID)
1927 {
1928 clRetainEvent (mapped->event);
1929 XPUSH_CLOBJ (stash_event, mapped->event);
1930 }
1931
1932void
1933task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1934 ALIAS:
1935 enqueue_task = 0
1936 PPCODE:
1937 EVENT_LIST (2);
1938
1939 cl_event ev = 0;
1940 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
574 1941
575 if (ev) 1942 if (ev)
576 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1943 XPUSH_CLOBJ (stash_event, ev);
577}
578 1944
579void 1945void
580enqueue_copy_image (OpenCL::Queue this, OpenCL::Image 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, ...)
581 PPCODE:
582{
583 cl_event ev = 0;
584 const size_t src_origin[3] = { src_x, src_y, src_z };
585 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
586 const size_t region[3] = { width, height, depth };
587 EVENT_LIST (12, items - 12);
588
589 NEED_SUCCESS (EnqueueCopyImage, (this, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
590
591 if (ev)
592 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
593}
594
595void
596enqueue_copy_image_to_buffer (OpenCL::Queue this, 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, ...)
597 PPCODE:
598{
599 cl_event ev = 0;
600 const size_t src_origin[3] = { src_x, src_y, src_z };
601 const size_t region[3] = { width, height, depth };
602 EVENT_LIST (10, items - 10);
603
604 NEED_SUCCESS (EnqueueCopyImageToBuffer, (this, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
605
606 if (ev)
607 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
608}
609
610void
611enqueue_task (OpenCL::Queue this, OpenCL::Kernel kernel, ...)
612 PPCODE:
613{
614 cl_event ev = 0;
615 EVENT_LIST (2, items - 2);
616
617 NEED_SUCCESS (EnqueueTask, (this, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
618
619 if (ev)
620 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
621}
622
623void
624enqueue_nd_range_kernel (OpenCL::Queue this, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) 1946nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1947 ALIAS:
1948 enqueue_nd_range_kernel = 0
625 PPCODE: 1949 PPCODE:
626{ 1950 EVENT_LIST (5);
627 cl_event ev = 0; 1951
628 size_t *gwo = 0, *gws, *lws = 0; 1952 size_t *gwo = 0, *gws, *lws = 0;
629 int gws_len; 1953 int gws_len;
630 size_t *lists; 1954 size_t *lists;
631 int i; 1955 int i;
632 EVENT_LIST (5, items - 5);
633 1956
634 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) 1957 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
635 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1958 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
636 1959
637 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1960 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
638 1961
639 lists = tmpbuf (sizeof (size_t) * 3 * gws_len); 1962 lists = tmpbuf (sizeof (size_t) * 3 * gws_len);
640 1963
641 gws = lists + gws_len * 0; 1964 gws = lists + gws_len * 0;
642 for (i = 0; i < gws_len; ++i) 1965 for (i = 0; i < gws_len; ++i)
1966 {
643 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); 1967 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]);
1968 // at least nvidia crashes for 0-sized work group sizes, work around
1969 if (!gws [i])
1970 croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i);
1971 }
644 1972
645 if (SvOK (global_work_offset)) 1973 if (SvOK (global_work_offset))
646 { 1974 {
647 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) 1975 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV)
648 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); 1976 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference");
655 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1983 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
656 } 1984 }
657 1985
658 if (SvOK (local_work_size)) 1986 if (SvOK (local_work_size))
659 { 1987 {
660 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1988 if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV)
661 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); 1989 croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference");
662 1990
663 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1991 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
664 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); 1992 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
665 1993
666 lws = lists + gws_len * 2; 1994 lws = lists + gws_len * 2;
667 for (i = 0; i < gws_len; ++i) 1995 for (i = 0; i < gws_len; ++i)
1996 {
668 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); 1997 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]);
1998 // at least nvidia crashes for 0-sized work group sizes, work around
1999 if (!lws [i])
2000 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
2001 }
669 } 2002 }
670 2003
2004 cl_event ev = 0;
671 NEED_SUCCESS (EnqueueNDRangeKernel, (this, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2005 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
672 2006
673 if (ev) 2007 if (ev)
674 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 2008 XPUSH_CLOBJ (stash_event, ev);
675}
676 2009
2010#if CL_VERSION_1_2
2011
677void 2012void
678enqueue_marker (OpenCL::Queue this) 2013migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...)
2014 ALIAS:
2015 enqueue_migrate_mem_objects = 0
679 PPCODE: 2016 PPCODE:
680{ 2017 EVENT_LIST (3);
2018
2019 cl_uint obj_count;
2020 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2021
2022 cl_event ev = 0;
2023 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2024
2025 if (ev)
2026 XPUSH_CLOBJ (stash_event, ev);
2027
2028#endif
2029
2030#if cl_apple_gl_sharing || cl_khr_gl_sharing
2031
2032void
2033acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
2034 ALIAS:
2035 release_gl_objects = 1
2036 enqueue_acquire_gl_objects = 0
2037 enqueue_release_gl_objects = 1
2038 PPCODE:
2039 EVENT_LIST (2);
2040
2041 cl_uint obj_count;
2042 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2043
681 cl_event ev; 2044 cl_event ev = 0;
682 NEED_SUCCESS (EnqueueMarker, (this, &ev));
683 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
684}
685 2045
2046 if (ix)
2047 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2048 else
2049 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2050
2051 if (ev)
2052 XPUSH_CLOBJ (stash_event, ev);
2053
2054#endif
2055
686void 2056void
687enqueue_wait_for_events (OpenCL::Queue this, ...) 2057wait_for_events (OpenCL::Queue self, ...)
2058 ALIAS:
2059 enqueue_wait_for_events = 0
688 CODE: 2060 CODE:
689{
690 EVENT_LIST (1, items - 1); 2061 EVENT_LIST (1);
2062#if PREFER_1_1
691 NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); 2063 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
692} 2064#else
2065 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0));
2066#endif
693 2067
694void 2068void
695enqueue_barrier (OpenCL::Queue this) 2069marker (OpenCL::Queue self, ...)
2070 ALIAS:
2071 enqueue_marker = 0
696 CODE: 2072 PPCODE:
2073 EVENT_LIST (1);
2074 cl_event ev = 0;
2075#if PREFER_1_1
2076 if (!event_list_count)
2077 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2078 else
2079#if CL_VERSION_1_2
2080 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2081#else
2082 {
2083 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier
2084 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2085 }
2086#endif
2087#else
2088 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2089#endif
2090 if (ev)
2091 XPUSH_CLOBJ (stash_event, ev);
2092
2093void
2094barrier (OpenCL::Queue self, ...)
2095 ALIAS:
2096 enqueue_barrier = 0
2097 PPCODE:
2098 EVENT_LIST (1);
2099 cl_event ev = 0;
2100#if PREFER_1_1
2101 if (!event_list_count && GIMME_V == G_VOID)
697 NEED_SUCCESS (EnqueueBarrier, (this)); 2102 NEED_SUCCESS (EnqueueBarrier, (self));
2103 else
2104#if CL_VERSION_1_2
2105 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2106#else
2107 {
2108 if (event_list_count)
2109 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
698 2110
2111 if (GIMME_V != G_VOID)
2112 NEED_SUCCESS (EnqueueMarker, (self, &ev));
2113 }
2114#endif
2115#else
2116 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2117#endif
2118 if (ev)
2119 XPUSH_CLOBJ (stash_event, ev);
2120
699void 2121void
700flush (OpenCL::Queue this) 2122flush (OpenCL::Queue self)
701 CODE: 2123 CODE:
702 NEED_SUCCESS (Flush, (this)); 2124 NEED_SUCCESS (Flush, (self));
703 2125
704void 2126void
705finish (OpenCL::Queue this) 2127finish (OpenCL::Queue self)
706 CODE: 2128 CODE:
707 NEED_SUCCESS (Finish, (this)); 2129 NEED_SUCCESS (Finish, (self));
2130
2131void
2132info (OpenCL::Queue self, cl_command_queue_info name)
2133 PPCODE:
2134 INFO (CommandQueue)
2135
2136#BEGIN:command_queue
2137
2138void
2139context (OpenCL::Queue self)
2140 PPCODE:
2141 cl_context value [1];
2142 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0));
2143 EXTEND (SP, 1);
2144 const int i = 0;
2145 {
2146 NEED_SUCCESS (RetainContext, (value [i]));
2147 PUSH_CLOBJ (stash_context, value [i]);
2148 }
2149
2150void
2151device (OpenCL::Queue self)
2152 PPCODE:
2153 cl_device_id value [1];
2154 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0));
2155 EXTEND (SP, 1);
2156 const int i = 0;
2157 {
2158 PUSH_CLOBJ (stash_device, value [i]);
2159 }
2160
2161void
2162reference_count (OpenCL::Queue self)
2163 PPCODE:
2164 cl_uint value [1];
2165 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0));
2166 EXTEND (SP, 1);
2167 const int i = 0;
2168 PUSHs (sv_2mortal (newSVuv (value [i])));
2169
2170void
2171properties (OpenCL::Queue self)
2172 PPCODE:
2173 cl_command_queue_properties value [1];
2174 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0));
2175 EXTEND (SP, 1);
2176 const int i = 0;
2177 PUSHs (sv_2mortal (newSVuv (value [i])));
2178
2179#END:command_queue
708 2180
709MODULE = OpenCL PACKAGE = OpenCL::Memory 2181MODULE = OpenCL PACKAGE = OpenCL::Memory
710 2182
711void 2183void
712DESTROY (OpenCL::Memory this) 2184DESTROY (OpenCL::Memory self)
713 CODE: 2185 CODE:
714 clReleaseMemObject (this); 2186 clReleaseMemObject (self);
715 2187
716void 2188void
717info (OpenCL::Memory this, cl_mem_info name) 2189info (OpenCL::Memory self, cl_mem_info name)
718 PPCODE: 2190 PPCODE:
719 INFO (MemObject) 2191 INFO (MemObject)
720 2192
2193#BEGIN:mem
2194
2195void
2196type (OpenCL::Memory self)
2197 PPCODE:
2198 cl_mem_object_type value [1];
2199 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0));
2200 EXTEND (SP, 1);
2201 const int i = 0;
2202 PUSHs (sv_2mortal (newSVuv (value [i])));
2203
2204void
2205flags (OpenCL::Memory self)
2206 PPCODE:
2207 cl_mem_flags value [1];
2208 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0));
2209 EXTEND (SP, 1);
2210 const int i = 0;
2211 PUSHs (sv_2mortal (newSVuv (value [i])));
2212
2213void
2214size (OpenCL::Memory self)
2215 ALIAS:
2216 size = CL_MEM_SIZE
2217 offset = CL_MEM_OFFSET
2218 PPCODE:
2219 size_t value [1];
2220 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
2221 EXTEND (SP, 1);
2222 const int i = 0;
2223 PUSHs (sv_2mortal (newSVuv (value [i])));
2224
2225void
2226host_ptr (OpenCL::Memory self)
2227 PPCODE:
2228 void * value [1];
2229 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0));
2230 EXTEND (SP, 1);
2231 const int i = 0;
2232 PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i])));
2233
2234void
2235map_count (OpenCL::Memory self)
2236 ALIAS:
2237 map_count = CL_MEM_MAP_COUNT
2238 reference_count = CL_MEM_REFERENCE_COUNT
2239 PPCODE:
2240 cl_uint value [1];
2241 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
2242 EXTEND (SP, 1);
2243 const int i = 0;
2244 PUSHs (sv_2mortal (newSVuv (value [i])));
2245
2246void
2247context (OpenCL::Memory self)
2248 PPCODE:
2249 cl_context value [1];
2250 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0));
2251 EXTEND (SP, 1);
2252 const int i = 0;
2253 {
2254 NEED_SUCCESS (RetainContext, (value [i]));
2255 PUSH_CLOBJ (stash_context, value [i]);
2256 }
2257
2258void
2259associated_memobject (OpenCL::Memory self)
2260 PPCODE:
2261 cl_mem value [1];
2262 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0));
2263 EXTEND (SP, 1);
2264 const int i = 0;
2265 {
2266 NEED_SUCCESS (RetainMemObject, (value [i]));
2267 PUSH_CLOBJ (stash_memory, value [i]);
2268 }
2269
2270#END:mem
2271
2272#if cl_apple_gl_sharing || cl_khr_gl_sharing
2273
2274void
2275gl_object_info (OpenCL::Memory self)
2276 PPCODE:
2277 cl_gl_object_type type;
2278 cl_GLuint name;
2279 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
2280 EXTEND (SP, 2);
2281 PUSHs (sv_2mortal (newSVuv (type)));
2282 PUSHs (sv_2mortal (newSVuv (name)));
2283
2284#endif
2285
2286MODULE = OpenCL PACKAGE = OpenCL::BufferObj
2287
2288void
2289sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
2290 PPCODE:
2291 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR))
2292 croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!");
2293
2294 cl_buffer_region crdata = { origin, size };
2295
2296 NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res));
2297 XPUSH_CLOBJ (stash_buffer, mem);
2298
2299MODULE = OpenCL PACKAGE = OpenCL::Image
2300
2301void
2302image_info (OpenCL::Image self, cl_image_info name)
2303 PPCODE:
2304 INFO (Image)
2305
2306void
2307format (OpenCL::Image self)
2308 PPCODE:
2309 cl_image_format format;
2310 NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
2311 EXTEND (SP, 2);
2312 PUSHs (sv_2mortal (newSVuv (format.image_channel_order)));
2313 PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type)));
2314
2315#BEGIN:image
2316
2317void
2318element_size (OpenCL::Image self)
2319 ALIAS:
2320 element_size = CL_IMAGE_ELEMENT_SIZE
2321 row_pitch = CL_IMAGE_ROW_PITCH
2322 slice_pitch = CL_IMAGE_SLICE_PITCH
2323 width = CL_IMAGE_WIDTH
2324 height = CL_IMAGE_HEIGHT
2325 depth = CL_IMAGE_DEPTH
2326 PPCODE:
2327 size_t value [1];
2328 NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0));
2329 EXTEND (SP, 1);
2330 const int i = 0;
2331 PUSHs (sv_2mortal (newSVuv (value [i])));
2332
2333#END:image
2334
2335#if cl_apple_gl_sharing || cl_khr_gl_sharing
2336
2337#BEGIN:gl_texture
2338
2339void
2340target (OpenCL::Image self)
2341 PPCODE:
2342 cl_GLenum value [1];
2343 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
2344 EXTEND (SP, 1);
2345 const int i = 0;
2346 PUSHs (sv_2mortal (newSVuv (value [i])));
2347
2348void
2349gl_mipmap_level (OpenCL::Image self)
2350 PPCODE:
2351 cl_GLint value [1];
2352 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
2353 EXTEND (SP, 1);
2354 const int i = 0;
2355 PUSHs (sv_2mortal (newSViv (value [i])));
2356
2357#END:gl_texture
2358
2359#endif
2360
721MODULE = OpenCL PACKAGE = OpenCL::Sampler 2361MODULE = OpenCL PACKAGE = OpenCL::Sampler
722 2362
723void 2363void
724DESTROY (OpenCL::Sampler this) 2364DESTROY (OpenCL::Sampler self)
725 CODE: 2365 CODE:
726 clReleaseSampler (this); 2366 clReleaseSampler (self);
727 2367
728void 2368void
729info (OpenCL::Sampler this, cl_sampler_info name) 2369info (OpenCL::Sampler self, cl_sampler_info name)
730 PPCODE: 2370 PPCODE:
731 INFO (Sampler) 2371 INFO (Sampler)
732 2372
2373#BEGIN:sampler
2374
2375void
2376reference_count (OpenCL::Sampler self)
2377 PPCODE:
2378 cl_uint value [1];
2379 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_REFERENCE_COUNT, sizeof (value), value, 0));
2380 EXTEND (SP, 1);
2381 const int i = 0;
2382 PUSHs (sv_2mortal (newSVuv (value [i])));
2383
2384void
2385context (OpenCL::Sampler self)
2386 PPCODE:
2387 cl_context value [1];
2388 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0));
2389 EXTEND (SP, 1);
2390 const int i = 0;
2391 {
2392 NEED_SUCCESS (RetainContext, (value [i]));
2393 PUSH_CLOBJ (stash_context, value [i]);
2394 }
2395
2396void
2397normalized_coords (OpenCL::Sampler self)
2398 PPCODE:
2399 cl_addressing_mode value [1];
2400 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0));
2401 EXTEND (SP, 1);
2402 const int i = 0;
2403 PUSHs (sv_2mortal (newSVuv (value [i])));
2404
2405void
2406addressing_mode (OpenCL::Sampler self)
2407 PPCODE:
2408 cl_filter_mode value [1];
2409 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0));
2410 EXTEND (SP, 1);
2411 const int i = 0;
2412 PUSHs (sv_2mortal (newSVuv (value [i])));
2413
2414void
2415filter_mode (OpenCL::Sampler self)
2416 PPCODE:
2417 cl_bool value [1];
2418 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0));
2419 EXTEND (SP, 1);
2420 const int i = 0;
2421 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
2422
2423#END:sampler
2424
733MODULE = OpenCL PACKAGE = OpenCL::Program 2425MODULE = OpenCL PACKAGE = OpenCL::Program
734 2426
735void 2427void
736DESTROY (OpenCL::Program this) 2428DESTROY (OpenCL::Program self)
737 CODE: 2429 CODE:
738 clReleaseProgram (this); 2430 clReleaseProgram (self);
739 2431
740void 2432void
741info (OpenCL::Program this, cl_program_info name) 2433build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2434 ALIAS:
2435 build_async = 1
742 PPCODE: 2436 CODE:
743 INFO (Program) 2437 cl_uint device_count = 0;
2438 cl_device_id *device_list = 0;
744 2439
2440 if (SvOK (devices))
2441 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2442
2443 void *user_data;
2444 program_callback pfn_notify = make_program_callback (notify, &user_data);
2445
2446 if (ix)
2447 build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data);
2448 else
2449 NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
2450
2451#if CL_VERSION_1_2
2452
745void 2453void
746build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) 2454compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef)
747 CODE: 2455 CODE:
2456 cl_uint device_count = 0;
2457 cl_device_id *device_list = 0;
2458
2459 if (SvOK (devices))
2460 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2461
2462 cl_uint header_count = 0;
2463 cl_program *header_list = 0;
2464 const char **header_name = 0;
2465
2466 if (SvOK (headers))
2467 {
2468 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2469 croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs");
2470
2471 HV *hv = (HV *)SvRV (devices);
2472
2473 header_count = hv_iterinit (hv);
2474 header_list = tmpbuf (sizeof (*header_list) * header_count);
2475 header_name = tmpbuf (sizeof (*header_name) * header_count);
2476
2477 HE *he;
2478 int i = 0;
2479 while (he = hv_iternext (hv))
2480 {
2481 header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he));
2482 header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program");
2483 ++i;
2484 }
2485 }
2486
2487 void *user_data;
2488 program_callback pfn_notify = make_program_callback (notify, &user_data);
2489
748 NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); 2490 NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
2491 header_count, header_list, header_name, pfn_notify, user_data));
749 2492
2493#endif
2494
750void 2495void
751build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) 2496build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
752 PPCODE: 2497 PPCODE:
753{
754 size_t size; 2498 size_t size;
755 SV *sv;
756
757 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); 2499 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size));
758 sv = sv_2mortal (newSV (size)); 2500 SV *sv = sv_2mortal (newSV (size));
759 SvUPGRADE (sv, SVt_PV); 2501 SvUPGRADE (sv, SVt_PV);
760 SvPOK_only (sv); 2502 SvPOK_only (sv);
761 SvCUR_set (sv, size); 2503 SvCUR_set (sv, size);
762 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); 2504 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0));
763 XPUSHs (sv); 2505 XPUSHs (sv);
764} 2506
2507#BEGIN:program_build
2508
2509void
2510build_status (OpenCL::Program self, OpenCL::Device device)
2511 PPCODE:
2512 cl_build_status value [1];
2513 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0));
2514 EXTEND (SP, 1);
2515 const int i = 0;
2516 PUSHs (sv_2mortal (newSViv (value [i])));
2517
2518void
2519build_options (OpenCL::Program self, OpenCL::Device device)
2520 ALIAS:
2521 build_options = CL_PROGRAM_BUILD_OPTIONS
2522 build_log = CL_PROGRAM_BUILD_LOG
2523 PPCODE:
2524 size_t size;
2525 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size));
2526 char *value = tmpbuf (size);
2527 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2528 EXTEND (SP, 1);
2529 const int i = 0;
2530 PUSHs (sv_2mortal (newSVpv (value, 0)));
2531
2532void
2533binary_type (OpenCL::Program self, OpenCL::Device device)
2534 PPCODE:
2535 cl_program_binary_type value [1];
2536 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0));
2537 EXTEND (SP, 1);
2538 const int i = 0;
2539 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
2540
2541#END:program_build
765 2542
766void 2543void
767kernel (OpenCL::Program program, SV *function) 2544kernel (OpenCL::Program program, SV *function)
768 PPCODE: 2545 PPCODE:
769{
770 cl_int res;
771 cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); 2546 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
2547 XPUSH_CLOBJ (stash_kernel, kernel);
772 2548
773 if (res) 2549void
774 FAIL (CreateKernel, res); 2550kernels_in_program (OpenCL::Program program)
2551 PPCODE:
2552 cl_uint num_kernels;
2553 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
2554 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
2555 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
775 2556
776 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 2557 int i;
777} 2558 EXTEND (SP, num_kernels);
2559 for (i = 0; i < num_kernels; ++i)
2560 PUSH_CLOBJ (stash_kernel, kernels [i]);
2561
2562void
2563info (OpenCL::Program self, cl_program_info name)
2564 PPCODE:
2565 INFO (Program)
2566
2567void
2568binaries (OpenCL::Program self)
2569 PPCODE:
2570 cl_uint n, i;
2571 size_t size;
2572
2573 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0));
2574 if (!n) XSRETURN_EMPTY;
2575
2576 size_t *sizes = tmpbuf (sizeof (*sizes) * n);
2577 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size));
2578 if (size != sizeof (*sizes) * n) XSRETURN_EMPTY;
2579 unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n);
2580
2581 EXTEND (SP, n);
2582 for (i = 0; i < n; ++i)
2583 {
2584 SV *sv = sv_2mortal (newSV (sizes [i]));
2585 SvUPGRADE (sv, SVt_PV);
2586 SvPOK_only (sv);
2587 SvCUR_set (sv, sizes [i]);
2588 ptrs [i] = (void *)SvPVX (sv);
2589 PUSHs (sv);
2590 }
2591
2592 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
2593 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
2594
2595#BEGIN:program
2596
2597void
2598reference_count (OpenCL::Program self)
2599 ALIAS:
2600 reference_count = CL_PROGRAM_REFERENCE_COUNT
2601 num_devices = CL_PROGRAM_NUM_DEVICES
2602 PPCODE:
2603 cl_uint value [1];
2604 NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0));
2605 EXTEND (SP, 1);
2606 const int i = 0;
2607 PUSHs (sv_2mortal (newSVuv (value [i])));
2608
2609void
2610context (OpenCL::Program self)
2611 PPCODE:
2612 cl_context value [1];
2613 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0));
2614 EXTEND (SP, 1);
2615 const int i = 0;
2616 {
2617 NEED_SUCCESS (RetainContext, (value [i]));
2618 PUSH_CLOBJ (stash_context, value [i]);
2619 }
2620
2621void
2622devices (OpenCL::Program self)
2623 PPCODE:
2624 size_t size;
2625 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size));
2626 cl_device_id *value = tmpbuf (size);
2627 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0));
2628 int i, n = size / sizeof (*value);
2629 EXTEND (SP, n);
2630 for (i = 0; i < n; ++i)
2631 {
2632 PUSH_CLOBJ (stash_device, value [i]);
2633 }
2634
2635void
2636source (OpenCL::Program self)
2637 PPCODE:
2638 size_t size;
2639 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size));
2640 char *value = tmpbuf (size);
2641 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0));
2642 EXTEND (SP, 1);
2643 const int i = 0;
2644 PUSHs (sv_2mortal (newSVpv (value, 0)));
2645
2646void
2647binary_sizes (OpenCL::Program self)
2648 PPCODE:
2649 size_t size;
2650 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size));
2651 size_t *value = tmpbuf (size);
2652 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0));
2653 int i, n = size / sizeof (*value);
2654 EXTEND (SP, n);
2655 for (i = 0; i < n; ++i)
2656 PUSHs (sv_2mortal (newSVuv (value [i])));
2657
2658#END:program
778 2659
779MODULE = OpenCL PACKAGE = OpenCL::Kernel 2660MODULE = OpenCL PACKAGE = OpenCL::Kernel
780 2661
781void 2662void
782DESTROY (OpenCL::Kernel this) 2663DESTROY (OpenCL::Kernel self)
783 CODE: 2664 CODE:
784 clReleaseKernel (this); 2665 clReleaseKernel (self);
785 2666
786void 2667void
2668setf (OpenCL::Kernel self, const char *format, ...)
2669 CODE:
2670 int i;
2671 for (i = 2; ; ++i)
2672 {
2673 while (*format == ' ')
2674 ++format;
2675
2676 char type = *format++;
2677
2678 if (!type)
2679 break;
2680
2681 if (i >= items)
2682 croak ("OpenCL::Kernel::setf format string too long (not enough arguments)");
2683
2684 SV *sv = ST (i);
2685
2686 union
2687 {
2688 cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS;
2689 cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL;
2690 cl_half ch; cl_float cf; cl_double cd;
2691 cl_mem cm;
2692 cl_sampler ca;
2693 size_t cz;
2694 cl_event ce;
2695 } arg;
2696 size_t size;
2697 int nullarg = 0;
2698
2699 switch (type)
2700 {
2701 case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break;
2702 case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break;
2703 case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break;
2704 case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break;
2705 case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break;
2706 case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break;
2707 case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break;
2708 case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break;
2709
2710 case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break;
2711 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2712 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2713
2714 case 'z': nullarg = 1; size = SvIV (sv); break;
2715
2716 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2717 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2718 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2719
2720 default:
2721 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2722 }
2723
2724 res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg);
2725 if (res)
2726 croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res));
2727 }
2728
2729 if (i != items)
2730 croak ("OpenCL::Kernel::setf format string too short (too many arguments)");
2731
2732void
2733set_char (OpenCL::Kernel self, cl_uint idx, cl_char value)
2734 CODE:
2735 clSetKernelArg (self, idx, sizeof (value), &value);
2736
2737void
2738set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value)
2739 CODE:
2740 clSetKernelArg (self, idx, sizeof (value), &value);
2741
2742void
2743set_short (OpenCL::Kernel self, cl_uint idx, cl_short value)
2744 CODE:
2745 clSetKernelArg (self, idx, sizeof (value), &value);
2746
2747void
2748set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value)
2749 CODE:
2750 clSetKernelArg (self, idx, sizeof (value), &value);
2751
2752void
2753set_int (OpenCL::Kernel self, cl_uint idx, cl_int value)
2754 CODE:
2755 clSetKernelArg (self, idx, sizeof (value), &value);
2756
2757void
2758set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value)
2759 CODE:
2760 clSetKernelArg (self, idx, sizeof (value), &value);
2761
2762void
2763set_long (OpenCL::Kernel self, cl_uint idx, cl_long value)
2764 CODE:
2765 clSetKernelArg (self, idx, sizeof (value), &value);
2766
2767void
2768set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value)
2769 CODE:
2770 clSetKernelArg (self, idx, sizeof (value), &value);
2771
2772void
2773set_half (OpenCL::Kernel self, cl_uint idx, cl_half value)
2774 CODE:
2775 clSetKernelArg (self, idx, sizeof (value), &value);
2776
2777void
2778set_float (OpenCL::Kernel self, cl_uint idx, cl_float value)
2779 CODE:
2780 clSetKernelArg (self, idx, sizeof (value), &value);
2781
2782void
2783set_double (OpenCL::Kernel self, cl_uint idx, cl_double value)
2784 CODE:
2785 clSetKernelArg (self, idx, sizeof (value), &value);
2786
2787void
2788set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value)
2789 CODE:
2790 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2791
2792void
2793set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value)
2794 CODE:
2795 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2796
2797void
2798set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value)
2799 CODE:
2800 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2801
2802void
2803set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
2804 CODE:
2805 clSetKernelArg (self, idx, sizeof (value), &value);
2806
2807void
2808set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
2809 CODE:
2810 clSetKernelArg (self, idx, size, 0);
2811
2812void
2813set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
2814 CODE:
2815 clSetKernelArg (self, idx, sizeof (value), &value);
2816
2817void
787info (OpenCL::Kernel this, cl_kernel_info name) 2818info (OpenCL::Kernel self, cl_kernel_info name)
788 PPCODE: 2819 PPCODE:
789 INFO (Kernel) 2820 INFO (Kernel)
790 2821
2822#BEGIN:kernel
2823
791void 2824void
792set_char (OpenCL::Kernel this, cl_uint idx, cl_char value) 2825function_name (OpenCL::Kernel self)
2826 PPCODE:
2827 size_t size;
2828 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size));
2829 char *value = tmpbuf (size);
2830 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0));
2831 EXTEND (SP, 1);
2832 const int i = 0;
2833 PUSHs (sv_2mortal (newSVpv (value, 0)));
2834
2835void
2836num_args (OpenCL::Kernel self)
2837 ALIAS:
2838 num_args = CL_KERNEL_NUM_ARGS
2839 reference_count = CL_KERNEL_REFERENCE_COUNT
2840 PPCODE:
2841 cl_uint value [1];
2842 NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0));
2843 EXTEND (SP, 1);
2844 const int i = 0;
2845 PUSHs (sv_2mortal (newSVuv (value [i])));
2846
2847void
2848context (OpenCL::Kernel self)
2849 PPCODE:
2850 cl_context value [1];
2851 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0));
2852 EXTEND (SP, 1);
2853 const int i = 0;
2854 {
2855 NEED_SUCCESS (RetainContext, (value [i]));
2856 PUSH_CLOBJ (stash_context, value [i]);
2857 }
2858
2859void
2860program (OpenCL::Kernel self)
2861 PPCODE:
2862 cl_program value [1];
2863 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0));
2864 EXTEND (SP, 1);
2865 const int i = 0;
2866 {
2867 NEED_SUCCESS (RetainProgram, (value [i]));
2868 PUSH_CLOBJ (stash_program, value [i]);
2869 }
2870
2871#END:kernel
2872
2873void
2874work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name)
793 CODE: 2875 PPCODE:
794 clSetKernelArg (this, idx, sizeof (value), &value); 2876 size_t size;
2877 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size));
2878 SV *sv = sv_2mortal (newSV (size));
2879 SvUPGRADE (sv, SVt_PV);
2880 SvPOK_only (sv);
2881 SvCUR_set (sv, size);
2882 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0));
2883 XPUSHs (sv);
795 2884
2885#BEGIN:kernel_work_group
2886
796void 2887void
797set_uchar (OpenCL::Kernel this, cl_uint idx, cl_uchar value) 2888work_group_size (OpenCL::Kernel self, OpenCL::Device device)
2889 ALIAS:
2890 work_group_size = CL_KERNEL_WORK_GROUP_SIZE
2891 preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE
2892 PPCODE:
2893 size_t value [1];
2894 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2895 EXTEND (SP, 1);
2896 const int i = 0;
2897 PUSHs (sv_2mortal (newSVuv (value [i])));
2898
2899void
2900compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device)
2901 PPCODE:
2902 size_t size;
2903 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size));
2904 size_t *value = tmpbuf (size);
2905 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0));
2906 int i, n = size / sizeof (*value);
2907 EXTEND (SP, n);
2908 for (i = 0; i < n; ++i)
2909 PUSHs (sv_2mortal (newSVuv (value [i])));
2910
2911void
2912local_mem_size (OpenCL::Kernel self, OpenCL::Device device)
2913 ALIAS:
2914 local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE
2915 private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE
2916 PPCODE:
2917 cl_ulong value [1];
2918 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2919 EXTEND (SP, 1);
2920 const int i = 0;
2921 PUSHs (sv_2mortal (newSVuv (value [i])));
2922
2923#END:kernel_work_group
2924
2925#if CL_VERSION_1_2
2926
2927void
2928arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name)
798 CODE: 2929 PPCODE:
799 clSetKernelArg (this, idx, sizeof (value), &value); 2930 size_t size;
2931 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size));
2932 SV *sv = sv_2mortal (newSV (size));
2933 SvUPGRADE (sv, SVt_PV);
2934 SvPOK_only (sv);
2935 SvCUR_set (sv, size);
2936 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0));
2937 XPUSHs (sv);
800 2938
801void 2939#BEGIN:kernel_arg
802set_short (OpenCL::Kernel this, cl_uint idx, cl_short value)
803 CODE:
804 clSetKernelArg (this, idx, sizeof (value), &value);
805 2940
806void 2941void
807set_ushort (OpenCL::Kernel this, cl_uint idx, cl_ushort value) 2942arg_address_qualifier (OpenCL::Kernel self, cl_uint idx)
808 CODE: 2943 PPCODE:
809 clSetKernelArg (this, idx, sizeof (value), &value); 2944 cl_kernel_arg_address_qualifier value [1];
2945 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0));
2946 EXTEND (SP, 1);
2947 const int i = 0;
2948 PUSHs (sv_2mortal (newSVuv (value [i])));
810 2949
811void 2950void
812set_int (OpenCL::Kernel this, cl_uint idx, cl_int value) 2951arg_access_qualifier (OpenCL::Kernel self, cl_uint idx)
813 CODE: 2952 PPCODE:
814 clSetKernelArg (this, idx, sizeof (value), &value); 2953 cl_kernel_arg_access_qualifier value [1];
2954 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0));
2955 EXTEND (SP, 1);
2956 const int i = 0;
2957 PUSHs (sv_2mortal (newSVuv (value [i])));
815 2958
816void 2959void
817set_uint (OpenCL::Kernel this, cl_uint idx, cl_uint value) 2960arg_type_name (OpenCL::Kernel self, cl_uint idx)
818 CODE: 2961 ALIAS:
819 clSetKernelArg (this, idx, sizeof (value), &value); 2962 arg_type_name = CL_KERNEL_ARG_TYPE_NAME
2963 arg_name = CL_KERNEL_ARG_NAME
2964 PPCODE:
2965 size_t size;
2966 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size));
2967 char *value = tmpbuf (size);
2968 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0));
2969 EXTEND (SP, 1);
2970 const int i = 0;
2971 PUSHs (sv_2mortal (newSVpv (value, 0)));
820 2972
821void 2973void
822set_long (OpenCL::Kernel this, cl_uint idx, cl_long value) 2974arg_type_qualifier (OpenCL::Kernel self, cl_uint idx)
823 CODE: 2975 PPCODE:
824 clSetKernelArg (this, idx, sizeof (value), &value); 2976 cl_kernel_arg_type_qualifier value [1];
2977 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0));
2978 EXTEND (SP, 1);
2979 const int i = 0;
2980 PUSHs (sv_2mortal (newSVuv (value [i])));
825 2981
826void 2982#END:kernel_arg
827set_ulong (OpenCL::Kernel this, cl_uint idx, cl_ulong value)
828 CODE:
829 clSetKernelArg (this, idx, sizeof (value), &value);
830 2983
831void 2984#endif
832set_half (OpenCL::Kernel this, cl_uint idx, cl_half value)
833 CODE:
834 clSetKernelArg (this, idx, sizeof (value), &value);
835
836void
837set_float (OpenCL::Kernel this, cl_uint idx, cl_float value)
838 CODE:
839 clSetKernelArg (this, idx, sizeof (value), &value);
840
841void
842set_double (OpenCL::Kernel this, cl_uint idx, cl_double value)
843 CODE:
844 clSetKernelArg (this, idx, sizeof (value), &value);
845
846void
847set_memory (OpenCL::Kernel this, cl_uint idx, OpenCL::Memory_ornull value)
848 CODE:
849 clSetKernelArg (this, idx, sizeof (value), &value);
850
851void
852set_buffer (OpenCL::Kernel this, cl_uint idx, OpenCL::Buffer_ornull value)
853 CODE:
854 clSetKernelArg (this, idx, sizeof (value), &value);
855
856void
857set_image2d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image2D_ornull value)
858 CODE:
859 clSetKernelArg (this, idx, sizeof (value), &value);
860
861void
862set_image3d (OpenCL::Kernel this, cl_uint idx, OpenCL::Image3D_ornull value)
863 CODE:
864 clSetKernelArg (this, idx, sizeof (value), &value);
865
866void
867set_sampler (OpenCL::Kernel this, cl_uint idx, OpenCL::Sampler value)
868 CODE:
869 clSetKernelArg (this, idx, sizeof (value), &value);
870
871void
872set_event (OpenCL::Kernel this, cl_uint idx, OpenCL::Event value)
873 CODE:
874 clSetKernelArg (this, idx, sizeof (value), &value);
875 2985
876MODULE = OpenCL PACKAGE = OpenCL::Event 2986MODULE = OpenCL PACKAGE = OpenCL::Event
877 2987
878void 2988void
879DESTROY (OpenCL::Event this) 2989DESTROY (OpenCL::Event self)
880 CODE: 2990 CODE:
881 clReleaseEvent (this); 2991 clReleaseEvent (self);
882 2992
883void 2993void
2994wait (OpenCL::Event self)
2995 CODE:
2996 clWaitForEvents (1, &self);
2997
2998void
2999cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb)
3000 CODE:
3001 clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb)));
3002
3003void
884info (OpenCL::Event this, cl_event_info name) 3004info (OpenCL::Event self, cl_event_info name)
885 PPCODE: 3005 PPCODE:
886 INFO (Event) 3006 INFO (Event)
887 3007
3008#BEGIN:event
3009
888void 3010void
889wait (OpenCL::Event this) 3011command_queue (OpenCL::Event self)
3012 PPCODE:
3013 cl_command_queue value [1];
3014 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0));
3015 EXTEND (SP, 1);
3016 const int i = 0;
3017 {
3018 NEED_SUCCESS (RetainCommandQueue, (value [i]));
3019 PUSH_CLOBJ (stash_queue, value [i]);
3020 }
3021
3022void
3023command_type (OpenCL::Event self)
3024 PPCODE:
3025 cl_command_type value [1];
3026 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_TYPE, sizeof (value), value, 0));
3027 EXTEND (SP, 1);
3028 const int i = 0;
3029 PUSHs (sv_2mortal (newSVuv (value [i])));
3030
3031void
3032reference_count (OpenCL::Event self)
3033 ALIAS:
3034 reference_count = CL_EVENT_REFERENCE_COUNT
3035 command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS
3036 PPCODE:
3037 cl_uint value [1];
3038 NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0));
3039 EXTEND (SP, 1);
3040 const int i = 0;
3041 PUSHs (sv_2mortal (newSVuv (value [i])));
3042
3043void
3044context (OpenCL::Event self)
3045 PPCODE:
3046 cl_context value [1];
3047 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0));
3048 EXTEND (SP, 1);
3049 const int i = 0;
3050 {
3051 NEED_SUCCESS (RetainContext, (value [i]));
3052 PUSH_CLOBJ (stash_context, value [i]);
3053 }
3054
3055#END:event
3056
3057void
3058profiling_info (OpenCL::Event self, cl_profiling_info name)
890 CODE: 3059 PPCODE:
891 clWaitForEvents (1, &this); 3060 INFO (EventProfiling)
3061
3062#BEGIN:profiling
3063
3064void
3065profiling_command_queued (OpenCL::Event self)
3066 ALIAS:
3067 profiling_command_queued = CL_PROFILING_COMMAND_QUEUED
3068 profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT
3069 profiling_command_start = CL_PROFILING_COMMAND_START
3070 profiling_command_end = CL_PROFILING_COMMAND_END
3071 PPCODE:
3072 cl_ulong value [1];
3073 NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0));
3074 EXTEND (SP, 1);
3075 const int i = 0;
3076 PUSHs (sv_2mortal (newSVuv (value [i])));
3077
3078#END:profiling
892 3079
893MODULE = OpenCL PACKAGE = OpenCL::UserEvent 3080MODULE = OpenCL PACKAGE = OpenCL::UserEvent
894 3081
895void 3082void
896set_status (OpenCL::UserEvent this, cl_int execution_status) 3083set_status (OpenCL::UserEvent self, cl_int execution_status)
897 CODE: 3084 CODE:
898 clSetUserEventStatus (this, execution_status); 3085 clSetUserEventStatus (self, execution_status);
899 3086
3087MODULE = OpenCL PACKAGE = OpenCL::Mapped
3088
3089void
3090DESTROY (SV *self)
3091 CODE:
3092 OpenCL__Mapped mapped = SvMAPPED (self);
3093
3094 clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0);
3095 mapped_detach (self, mapped);
3096
3097 clReleaseCommandQueue (mapped->queue);
3098 clReleaseEvent (mapped->event);
3099 Safefree (mapped);
3100
3101void
3102unmap (OpenCL::Mapped self, ...)
3103 CODE:
3104 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
3105
3106bool
3107mapped (OpenCL::Mapped self)
3108 CODE:
3109 RETVAL = !!self->ptr;
3110 OUTPUT:
3111 RETVAL
3112
3113void
3114wait (OpenCL::Mapped self)
3115 PPCODE:
3116 if (self->event)
3117 NEED_SUCCESS (WaitForEvents, (1, &self->event));
3118
3119void
3120event (OpenCL::Mapped self)
3121 PPCODE:
3122 if (!self->event)
3123 XSRETURN_UNDEF;
3124
3125 clRetainEvent (self->event);
3126 XPUSH_CLOBJ (stash_event, self->event);
3127
3128size_t
3129size (OpenCL::Mapped self)
3130 CODE:
3131 RETVAL = self->cb;
3132 OUTPUT:
3133 RETVAL
3134
3135IV
3136ptr (OpenCL::Mapped self)
3137 CODE:
3138 RETVAL = PTR2IV (self->ptr);
3139 OUTPUT:
3140 RETVAL
3141
3142void
3143set (OpenCL::Mapped self, size_t offset, SV *data)
3144 CODE:
3145 STRLEN len;
3146 const char *ptr = SvPVbyte (data, len);
3147
3148 if (offset + len > self->cb)
3149 croak ("OpenCL::Mapped::set out of bound condition detected");
3150
3151 memcpy (offset + (char *)self->ptr, ptr, len);
3152
3153MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer
3154
3155MODULE = OpenCL PACKAGE = OpenCL::MappedImage
3156
3157IV
3158row_pitch (OpenCL::Mapped self)
3159 ALIAS:
3160 slice_pitch = 1
3161 CODE:
3162 RETVAL = ix ? self->slice_pitch : self->row_pitch;
3163 OUTPUT:
3164 RETVAL
3165
3166
3167

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines