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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.5 by root, Wed Nov 16 00:35:30 2011 UTC vs.
Revision 1.61 by root, Tue May 1 16:37:23 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#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
6#include "xthread.h"
7#include "schmorp.h"
8
9#ifdef I_DLFCN
10 #include <dlfcn.h>
11#endif
12
13// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
14// but then fail to define the api functions unless you ALSO define
15// this. This breaks 100% of the opencl 1.1 apps, for what reason?
16// after all, the functions are deprecated, not removed.
17// in addition, you cannot test for this in any future-proof way.
18// each time a new opencl version comes out, you need to make a new
19// release.
20#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
21
22#ifndef PREFER_1_1
23 #define PREFER_1_1 1
24#endif
25
26#if PREFER_1_1
27 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
28#endif
29
30#ifdef __APPLE__
31 #include <OpenCL/opencl.h>
32#else
5#include <CL/opencl.h> 33 #include <CL/opencl.h>
34#endif
6 35
36#ifndef CL_VERSION_1_2
37 #undef PREFER_1_1
38 #define PREFER_1_1 1
39#endif
40
7typedef cl_platform_id OpenCL__Platform; 41typedef cl_platform_id OpenCL__Platform;
8typedef cl_device_id OpenCL__Device; 42typedef cl_device_id OpenCL__Device;
9typedef cl_context OpenCL__Context; 43typedef cl_context OpenCL__Context;
10typedef cl_command_queue OpenCL__Queue; 44typedef cl_command_queue OpenCL__Queue;
11typedef cl_mem OpenCL__Memory; 45typedef cl_mem OpenCL__Memory;
12typedef cl_mem OpenCL__Buffer; 46typedef cl_mem OpenCL__Buffer;
47typedef cl_mem OpenCL__BufferObj;
13typedef cl_mem OpenCL__Image; 48typedef cl_mem OpenCL__Image;
14typedef cl_mem OpenCL__Image2D; 49typedef cl_mem OpenCL__Image2D;
15typedef cl_mem OpenCL__Image3D; 50typedef cl_mem OpenCL__Image3D;
16typedef cl_mem OpenCL__Memory_ornull; 51typedef cl_mem OpenCL__Memory_ornull;
17typedef cl_mem OpenCL__Buffer_ornull; 52typedef cl_mem OpenCL__Buffer_ornull;
18typedef cl_mem OpenCL__Image_ornull; 53typedef cl_mem OpenCL__Image_ornull;
19typedef cl_mem OpenCL__Image2D_ornull; 54typedef cl_mem OpenCL__Image2D_ornull;
20typedef cl_mem OpenCL__Image3D_ornull; 55typedef cl_mem OpenCL__Image3D_ornull;
21typedef cl_sampler OpenCL__Sampler; 56typedef cl_sampler OpenCL__Sampler;
22typedef cl_program OpenCL__Program; 57typedef cl_program OpenCL__Program;
23typedef cl_kernel OpenCL__Kernel; 58typedef cl_kernel OpenCL__Kernel;
24typedef cl_event OpenCL__Event; 59typedef cl_event OpenCL__Event;
25typedef cl_event OpenCL__UserEvent; 60typedef cl_event OpenCL__UserEvent;
61
62typedef struct mapped * OpenCL__Mapped;
63
64typedef SV *FUTURE;
65
66static HV
67 *stash_platform,
68 *stash_device,
69 *stash_context,
70 *stash_queue,
71 *stash_program,
72 *stash_kernel,
73 *stash_sampler,
74 *stash_event,
75 *stash_userevent,
76 *stash_memory,
77 *stash_buffer,
78 *stash_bufferobj,
79 *stash_image,
80 *stash_image1d,
81 *stash_image1darray,
82 *stash_image1dbuffer,
83 *stash_image2d,
84 *stash_image2darray,
85 *stash_image3d,
86 *stash_mapped,
87 *stash_mappedbuffer,
88 *stash_mappedimage;
89
90/*****************************************************************************/
91
92// name must include a leading underscore
93// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
94// for perl. doh.
95static void *
96glsym (const char *name)
97{
98 void *fun = 0;
99
100 #if defined I_DLFCN && defined RTLD_DEFAULT
101 fun = dlsym (RTLD_DEFAULT, name + 1);
102 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
103
104 if (!fun)
105 {
106 static void *libgl;
107 static const char *glso[] = {
108 "libGL.so.1",
109 "libGL.so.3",
110 "libGL.so.4.0",
111 "libGL.so",
112 "/usr/lib/libGL.so",
113 "/usr/X11R6/lib/libGL.1.dylib"
114 };
115 int i;
116
117 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
118 {
119 libgl = dlopen (glso [i], RTLD_LAZY);
120 if (libgl)
121 break;
122 }
123
124 if (libgl)
125 {
126 fun = dlsym (libgl, name + 1);
127 if (!fun) fun = dlsym (libgl, name);
128 }
129 }
130 #endif
131
132 return fun;
133}
26 134
27/*****************************************************************************/ 135/*****************************************************************************/
28 136
29/* up to two temporary buffers */ 137/* up to two temporary buffers */
30static void * 138static void *
31tmpbuf (size_t size) 139tmpbuf (size_t size)
32{ 140{
141 enum { buffers = 3 };
33 static int idx; 142 static int idx;
34 static void *buf [2]; 143 static void *buf [buffers];
35 static size_t len [2]; 144 static size_t len [buffers];
36 145
37 idx ^= 1; 146 idx = (idx + 1) % buffers;
38 147
39 if (len [idx] < size) 148 if (len [idx] < size)
40 { 149 {
41 free (buf [idx]); 150 free (buf [idx]);
42 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 151 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
90 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); 199 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)");
91} 200}
92 201
93/*****************************************************************************/ 202/*****************************************************************************/
94 203
95static cl_int last_error; 204static cl_int res;
96 205
97#define FAIL(name,err) \ 206#define FAIL(name) \
98 croak ("cl" # name ": %s", err2str (last_error = err)); 207 croak ("cl" # name ": %s", err2str (res));
99 208
100#define NEED_SUCCESS(name,args) \ 209#define NEED_SUCCESS(name,args) \
101 do { \ 210 do { \
102 cl_int res = cl ## name args; \ 211 res = cl ## name args; \
103 \ 212 \
104 if (res) \ 213 if (res) \
105 FAIL (name, res); \ 214 FAIL (name); \
106 } while (0) 215 } while (0)
107 216
217#define NEED_SUCCESS_ARG(retdecl, name, args) \
218 retdecl = cl ## name args; \
219 if (res) \
220 FAIL (name);
221
108/*****************************************************************************/ 222/*****************************************************************************/
109 223
110#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) 224static cl_context_properties *
111#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) 225SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
226{
227 if (!sv || !SvOK (sv))
228 if (extra)
229 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
230 else
231 return 0;
112 232
233 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
234 {
235 AV *av = (AV *)SvRV (sv);
236 int i, len = av_len (av) + 1;
237 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
238 cl_context_properties *l = p;
239
240 if (len & 1)
241 croak ("%s: %s is not a property list (must be even number of elements)", func, svname);
242
243 while (extracount--)
244 *l++ = *extra++;
245
246 for (i = 0; i < len; i += 2)
247 {
248 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
249 SV *p_sv = *av_fetch (av, i + 1, 0);
250 cl_context_properties v = SvIV (p_sv); // code below can override
251
252 switch (t)
253 {
254 case CL_GLX_DISPLAY_KHR:
255 if (!SvOK (p_sv))
256 {
257 void *func = glsym ("_glXGetCurrentDisplay");
258 if (func)
259 v = (cl_context_properties)((void *(*)(void))func)();
260 }
261 break;
262
263 case CL_GL_CONTEXT_KHR:
264 if (!SvOK (p_sv))
265 {
266 void *func = glsym ("_glXGetCurrentContext");
267 if (func)
268 v = (cl_context_properties)((void *(*)(void))func)();
269 }
270 break;
271
272 default:
273 /* unknown property, treat as int */
274 break;
275 }
276
277 *l++ = t;
278 *l++ = v;
279 }
280
281 *l = 0;
282
283 return p;
284 }
285
286 croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname);
287}
288
289/*****************************************************************************/
290
291static SV *
292new_clobj (HV *stash, IV id)
293{
294 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash));
295}
296
297#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id)))
298#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
299
300/* cl objects are either \$iv, or [$iv, ...] */
301/* they can be upgraded at runtime to the array form */
113static void * 302static void *
114SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) 303SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg)
115{ 304{
116 if (SvROK (sv) && sv_derived_from (sv, pkg)) 305 if (SvROK (sv) && sv_derived_from (sv, pkg))
117 return (void *)SvIV (SvRV (sv)); 306 return (void *)SvIV (SvRV (sv));
118 307
119 croak ("%s: %s is not of type %s", func, svname, pkg); 308 croak ("%s: %s is not of type %s", func, svname, pkg);
120} 309}
121 310
122/*****************************************************************************/ 311/*****************************************************************************/
312/* callback stuff */
313
314/* default context callback, log to stderr */
315static void CL_CALLBACK
316context_default_notify (const char *msg, const void *info, size_t cb, void *data)
317{
318 fprintf (stderr, "OpenCL Context Notify: %s\n", msg);
319}
320
321typedef struct
322{
323 int free_cb;
324 void (*push)(void *data1, void *data2, void *data3);
325} eq_vtbl;
326
327typedef struct eq_item
328{
329 struct eq_item *next;
330 eq_vtbl *vtbl;
331 SV *cb;
332 void *data1, *data2, *data3;
333} eq_item;
334
335static void (*eq_signal_func)(void *signal_arg, int value);
336static void *eq_signal_arg;
337static xmutex_t eq_lock = X_MUTEX_INIT;
338static eq_item *eq_head, *eq_tail;
339
340static void
341eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3)
342{
343 eq_item *item = malloc (sizeof (eq_item));
344
345 item->next = 0;
346 item->vtbl = vtbl;
347 item->cb = cb;
348 item->data1 = data1;
349 item->data2 = data2;
350 item->data3 = data3;
351
352 X_LOCK (eq_lock);
353
354 *(eq_head ? &eq_tail->next : &eq_head) = item;
355 eq_tail = item;
356
357 X_UNLOCK (eq_lock);
358
359 eq_signal_func (eq_signal_arg, 0);
360}
361
362static eq_item *
363eq_dec (void)
364{
365 eq_item *res;
366
367 X_LOCK (eq_lock);
368
369 res = eq_head;
370 if (res)
371 eq_head = res->next;
372
373 X_UNLOCK (eq_lock);
374
375 return res;
376}
377
378static void
379eq_poll (void)
380{
381 eq_item *item;
382
383 while ((item = eq_dec ()))
384 {
385 ENTER;
386 SAVETMPS;
387
388 dSP;
389 PUSHMARK (SP);
390 EXTEND (SP, 2);
391
392 if (item->vtbl->free_cb)
393 sv_2mortal (item->cb);
394
395 PUTBACK;
396 item->vtbl->push (item->data1, item->data2, item->data3);
397
398 SV *cb = item->cb;
399 free (item);
400
401 call_sv (cb, G_DISCARD | G_VOID);
402
403 FREETMPS;
404 LEAVE;
405 }
406}
407
408static void
409eq_poll_interrupt (pTHX_ void *c_arg, int value)
410{
411 eq_poll ();
412}
413
414/*****************************************************************************/
415/* context notify */
416
417static void
418eq_context_push (void *data1, void *data2, void *data3)
419{
420 dSP;
421 PUSHs (sv_2mortal (newSVpv (data1, 0)));
422 PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3)));
423 PUTBACK;
424
425 free (data1);
426 free (data2);
427}
428
429static eq_vtbl eq_context_vtbl = { 0, eq_context_push };
430
431static void CL_CALLBACK
432eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data)
433{
434 void *pvt_copy = malloc (cb);
435 memcpy (pvt_copy, pvt, cb);
436 eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb);
437}
438
439#define CONTEXT_NOTIFY_CALLBACK \
440 void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \
441 void *user_data = 0; \
442 \
443 if (SvOK (notify)) \
444 { \
445 pfn_notify = eq_context_notify; \
446 user_data = s_get_cv (notify); \
447 }
448
449static SV *
450new_clobj_context (cl_context ctx, void *user_data)
451{
452 SV *sv = new_clobj (stash_context, (IV)ctx);
453
454 if (user_data)
455 sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0);
456
457 return sv;
458}
459
460#define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data));
461
462/*****************************************************************************/
463/* build/compile/link notify */
464
465static void
466eq_program_push (void *data1, void *data2, void *data3)
467{
468 dSP;
469 PUSH_CLOBJ (stash_program, data1);
470 PUTBACK;
471}
472
473static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
474
475static void CL_CALLBACK
476eq_program_notify (cl_program program, void *user_data)
477{
478 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
479}
480
481struct build_args
482{
483 cl_program program;
484 char *options;
485 void *user_data;
486 cl_uint num_devices;
487};
488
489X_THREAD_PROC (build_program_thread)
490{
491 struct build_args *arg = thr_arg;
492
493 clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0);
494
495 if (arg->user_data)
496 eq_program_notify (arg->program, arg->user_data);
497 else
498 clReleaseProgram (arg->program);
499
500 free (arg->options);
501 free (arg);
502}
503
504static void
505build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data)
506{
507 struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices);
508
509 arg->program = program;
510 arg->options = strdup (options);
511 arg->user_data = user_data;
512 arg->num_devices = num_devices;
513 memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices);
514
515 xthread_t id;
516 thread_create (&id, build_program_thread, arg);
517}
518
519/*****************************************************************************/
520/* event objects */
521
522static void
523eq_event_push (void *data1, void *data2, void *data3)
524{
525 dSP;
526 PUSH_CLOBJ (stash_event, data1);
527 PUSHs (sv_2mortal (newSViv ((IV)data2)));
528 PUTBACK;
529}
530
531static eq_vtbl eq_event_vtbl = { 1, eq_event_push };
532
533static void CL_CALLBACK
534eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data)
535{
536 clRetainEvent (event);
537 eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0);
538}
539
540/*****************************************************************************/
541/* mapped_xxx */
542
543static OpenCL__Mapped
544SvMAPPED (SV *sv)
545{
546 // no typechecking atm., keep your fingers crossed
547 return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr;
548}
549
550struct mapped
551{
552 cl_command_queue queue;
553 cl_mem memobj;
554 void *ptr;
555 size_t cb;
556 cl_event event;
557 size_t row_pitch;
558 size_t slice_pitch;
559};
560
561static SV *
562mapped_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)
563{
564 SV *data = newSV (0);
565 SvUPGRADE (data, SVt_PVMG);
566
567 OpenCL__Mapped mapped;
568 New (0, mapped, 1, struct mapped);
569
570 clRetainCommandQueue (queue);
571
572 mapped->queue = queue;
573 mapped->memobj = memobj;
574 mapped->ptr = ptr;
575 mapped->cb = cb;
576 mapped->event = ev;
577 mapped->row_pitch = row_pitch;
578 mapped->slice_pitch = slice_pitch;
579
580 sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0);
581
582 if (!(flags & CL_MAP_WRITE))
583 SvREADONLY_on (data);
584
585 if (SvLEN (data))
586 Safefree (data);
587
588 SvPVX (data) = (char *)ptr;
589 SvCUR_set (data, cb);
590 SvLEN_set (data, 0);
591 SvPOK_only (data);
592
593 return sv_2mortal (sv_bless (newRV_noinc (data), stash));
594}
595
596static void
597mapped_detach (SV *sv, OpenCL__Mapped mapped)
598{
599 SV *data = SvRV (sv);
600
601 if (SvPVX (data) != (char *)mapped->ptr)
602 warn ("FATAL: OpenCL memory mapped scalar changed location, detected");
603 else
604 {
605 SvREADONLY_off (data);
606 SvCUR_set (data, 0);
607 SvPVX (data) = 0;
608 SvOK_off (data);
609 }
610
611 mapped->ptr = 0;
612}
613
614/*****************************************************************************/
615
616static size_t
617img_row_pitch (cl_mem img)
618{
619 size_t res;
620 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
621 return res;
622}
123 623
124static cl_event * 624static cl_event *
125event_list (SV **items, int count) 625event_list (SV **items, cl_uint *rcount, cl_event extra)
126{ 626{
627 cl_uint count = *rcount;
628
629 if (!count && !extra)
630 return 0;
631
127 cl_event *list = tmpbuf (sizeof (cl_event) * count); 632 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
633 int i = 0;
128 634
129 while (count--) 635 while (count)
636 {
637 --count;
638 if (SvOK (items [count]))
130 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 639 list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event");
640 }
131 641
642 if (extra)
643 list [i++] = extra;
644
645 *rcount = i;
646
132 return list; 647 return i ? list : 0;
133} 648}
134 649
135#define EVENT_LIST(items,count) \ 650#define EVENT_LIST(skip) \
136 cl_uint event_list_count = (count); \ 651 cl_uint event_list_count = items - skip; \
137 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 652 cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0)
138 653
139#define INFO(class) \ 654#define INFO(class) \
140{ \ 655{ \
141 size_t size; \ 656 size_t size; \
142 SV *sv; \
143 \
144 NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ 657 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
145 sv = sv_2mortal (newSV (size)); \ 658 SV *sv = sv_2mortal (newSV (size)); \
146 SvUPGRADE (sv, SVt_PV); \ 659 SvUPGRADE (sv, SVt_PV); \
147 SvPOK_only (sv); \ 660 SvPOK_only (sv); \
148 SvCUR_set (sv, size); \ 661 SvCUR_set (sv, size); \
149 NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ 662 NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \
150 XPUSHs (sv); \ 663 XPUSHs (sv); \
151} 664}
152 665
153MODULE = OpenCL PACKAGE = OpenCL 666MODULE = OpenCL PACKAGE = OpenCL
154 667
155PROTOTYPES: ENABLE 668PROTOTYPES: ENABLE
156 669
670void
671poll ()
672 CODE:
673 eq_poll ();
674
675void
676_eq_initialise (IV func, IV arg)
677 CODE:
678 eq_signal_func = (void (*)(void *, int))func;
679 eq_signal_arg = (void*)arg;
680
157BOOT: 681BOOT:
158{ 682{
159 HV *stash = gv_stashpv ("OpenCL", 1); 683 HV *stash = gv_stashpv ("OpenCL", 1);
160 static const ivstr *civ, const_iv[] = { 684 static const ivstr *civ, const_iv[] = {
161 { sizeof (cl_char ), "SIZEOF_CHAR" }, 685 { sizeof (cl_char ), "SIZEOF_CHAR" },
162 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 686 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
163 { sizeof (cl_short ), "SIZEOF_SHORT" }, 687 { sizeof (cl_short ), "SIZEOF_SHORT" },
164 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 688 { sizeof (cl_ushort), "SIZEOF_USHORT" },
165 { sizeof (cl_int ), "SIZEOF_INT" }, 689 { sizeof (cl_int ), "SIZEOF_INT" },
166 { sizeof (cl_uint ), "SIZEOF_UINT" }, 690 { sizeof (cl_uint ), "SIZEOF_UINT" },
167 { sizeof (cl_long ), "SIZEOF_LONG" }, 691 { sizeof (cl_long ), "SIZEOF_LONG" },
168 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 692 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
169 { sizeof (cl_half ), "SIZEOF_HALF" }, 693 { sizeof (cl_half ), "SIZEOF_HALF" },
170 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 694 { sizeof (cl_float ), "SIZEOF_FLOAT" },
171 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 695 { sizeof (cl_double), "SIZEOF_DOUBLE" },
172#include "constiv.h" 696#include "constiv.h"
173 }; 697 };
698
174 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 699 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)); 700 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
701
702 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
703 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD);
704 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
705 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD);
706 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD);
707 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD);
708 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD);
709 stash_event = gv_stashpv ("OpenCL::Event", GV_ADD);
710 stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD);
711 stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD);
712 stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD);
713 stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD);
714 stash_image = gv_stashpv ("OpenCL::Image", GV_ADD);
715 stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD);
716 stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD);
717 stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD);
718 stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD);
719 stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD);
720 stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD);
721 stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD);
722 stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD);
723 stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD);
724
725 sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt);
176} 726}
177 727
178cl_int 728cl_int
179errno () 729errno ()
180 CODE: 730 CODE:
181 errno = last_error; 731 RETVAL = res;
732 OUTPUT:
733 RETVAL
182 734
183const char * 735const char *
184err2str (cl_int err) 736err2str (cl_int err = res)
185 737
186const char * 738const char *
187enum2str (cl_uint value) 739enum2str (cl_uint value)
188 740
189void 741void
190platforms () 742platforms ()
191 PPCODE: 743 PPCODE:
192{
193 cl_platform_id *list; 744 cl_platform_id *list;
194 cl_uint count; 745 cl_uint count;
195 int i; 746 int i;
196 747
197 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); 748 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count));
198 list = tmpbuf (sizeof (*list) * count); 749 list = tmpbuf (sizeof (*list) * count);
199 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 750 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
200 751
201 EXTEND (SP, count); 752 EXTEND (SP, count);
202 for (i = 0; i < count; ++i) 753 for (i = 0; i < count; ++i)
203 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 754 PUSH_CLOBJ (stash_platform, (IV)list [i]);
204}
205 755
206void 756void
207context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 757context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
208 PPCODE: 758 PPCODE:
209{ 759 CONTEXT_NOTIFY_CALLBACK;
210 cl_int res; 760 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res));
211 cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); 761 XPUSH_CLOBJ_CONTEXT;
212 762
213 if (res) 763void
214 FAIL (CreateContextFromType, res); 764context (FUTURE properties, FUTURE devices, FUTURE notify)
215 765 PPCODE:
216 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 766 /* der Gipfel der Kunst */
217}
218 767
219void 768void
220wait_for_events (...) 769wait_for_events (...)
221 CODE: 770 CODE:
222{
223 EVENT_LIST (0, items); 771 EVENT_LIST (0);
224 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); 772 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr));
225}
226 773
227PROTOTYPES: DISABLE 774PROTOTYPES: DISABLE
228 775
229MODULE = OpenCL PACKAGE = OpenCL::Platform 776MODULE = OpenCL PACKAGE = OpenCL::Platform
230 777
231void 778void
232info (OpenCL::Platform this, cl_platform_info name) 779info (OpenCL::Platform self, cl_platform_info name)
233 PPCODE: 780 PPCODE:
234 INFO (Platform) 781 INFO (Platform)
235 782
236void 783void
784unload_compiler (OpenCL::Platform self)
785 CODE:
786#if CL_VERSION_1_2
787 clUnloadPlatformCompiler (self);
788#endif
789
790#BEGIN:platform
791
792void
793profile (OpenCL::Platform self)
794 ALIAS:
795 profile = CL_PLATFORM_PROFILE
796 version = CL_PLATFORM_VERSION
797 name = CL_PLATFORM_NAME
798 vendor = CL_PLATFORM_VENDOR
799 extensions = CL_PLATFORM_EXTENSIONS
800 PPCODE:
801 size_t size;
802 NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size));
803 char *value = tmpbuf (size);
804 NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0));
805 EXTEND (SP, 1);
806 const int i = 0;
807 PUSHs (sv_2mortal (newSVpv (value, 0)));
808
809#END:platform
810
811void
237devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) 812devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL)
238 PPCODE: 813 PPCODE:
239{
240 cl_device_id *list; 814 cl_device_id *list;
241 cl_uint count; 815 cl_uint count;
242 int i; 816 int i;
243 817
244 NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); 818 NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count));
245 list = tmpbuf (sizeof (*list) * count); 819 list = tmpbuf (sizeof (*list) * count);
246 NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); 820 NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0));
247 821
248 EXTEND (SP, count); 822 EXTEND (SP, count);
249 for (i = 0; i < count; ++i) 823 for (i = 0; i < count; ++i)
250 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 824 PUSH_CLOBJ (stash_device, list [i]);
251}
252 825
253void 826void
827context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
828 PPCODE:
829 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
830 croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call");
831
832 AV *av = (AV *)SvRV (devices);
833 cl_uint num_devices = av_len (av) + 1;
834 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices);
835
836 int i;
837 for (i = num_devices; i--; )
838 device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
839
840 CONTEXT_NOTIFY_CALLBACK;
841 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res));
842 XPUSH_CLOBJ_CONTEXT;
843
844void
254context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 845context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
255 PPCODE: 846 PPCODE:
256{
257 cl_int res;
258 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; 847 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
848 cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2);
849
850 CONTEXT_NOTIFY_CALLBACK;
259 cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); 851 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res));
260 852 XPUSH_CLOBJ_CONTEXT;
261 if (res)
262 FAIL (CreateContextFromType, res);
263
264 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
265}
266 853
267MODULE = OpenCL PACKAGE = OpenCL::Device 854MODULE = OpenCL PACKAGE = OpenCL::Device
268 855
269void 856void
270info (OpenCL::Device this, cl_device_info name) 857info (OpenCL::Device self, cl_device_info name)
271 PPCODE: 858 PPCODE:
272 INFO (Device) 859 INFO (Device)
273 860
861#BEGIN:device
862
274void 863void
275context_simple (OpenCL::Device this) 864type (OpenCL::Device self)
276 PPCODE: 865 PPCODE:
277{ 866 cl_device_type value [1];
278 cl_int res; 867 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0));
279 cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); 868 EXTEND (SP, 1);
869 const int i = 0;
870 PUSHs (sv_2mortal (newSVuv (value [i])));
280 871
281 if (res) 872void
282 FAIL (CreateContext, res); 873vendor_id (OpenCL::Device self)
874 ALIAS:
875 vendor_id = CL_DEVICE_VENDOR_ID
876 max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS
877 max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS
878 preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR
879 preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT
880 preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT
881 preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG
882 preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT
883 preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE
884 max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY
885 max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS
886 max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS
887 image_support = CL_DEVICE_IMAGE_SUPPORT
888 max_samplers = CL_DEVICE_MAX_SAMPLERS
889 mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN
890 min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE
891 global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE
892 max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS
893 preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF
894 native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR
895 native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT
896 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
897 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
898 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
899 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
900 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
901 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
902 PPCODE:
903 cl_uint value [1];
904 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
905 EXTEND (SP, 1);
906 const int i = 0;
907 PUSHs (sv_2mortal (newSVuv (value [i])));
283 908
284 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 909void
285} 910max_work_group_size (OpenCL::Device self)
911 ALIAS:
912 max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE
913 image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH
914 image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT
915 image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH
916 image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT
917 image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH
918 max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE
919 profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION
920 PPCODE:
921 size_t value [1];
922 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
923 EXTEND (SP, 1);
924 const int i = 0;
925 PUSHs (sv_2mortal (newSVuv (value [i])));
926
927void
928max_work_item_sizes (OpenCL::Device self)
929 PPCODE:
930 size_t size;
931 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size));
932 size_t *value = tmpbuf (size);
933 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0));
934 int i, n = size / sizeof (*value);
935 EXTEND (SP, n);
936 for (i = 0; i < n; ++i)
937 PUSHs (sv_2mortal (newSVuv (value [i])));
938
939void
940address_bits (OpenCL::Device self)
941 PPCODE:
942 cl_bitfield value [1];
943 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_ADDRESS_BITS, sizeof (value), value, 0));
944 EXTEND (SP, 1);
945 const int i = 0;
946 PUSHs (sv_2mortal (newSVuv (value [i])));
947
948void
949max_mem_alloc_size (OpenCL::Device self)
950 ALIAS:
951 max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE
952 global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE
953 global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE
954 max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE
955 local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE
956 PPCODE:
957 cl_ulong value [1];
958 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
959 EXTEND (SP, 1);
960 const int i = 0;
961 PUSHs (sv_2mortal (newSVuv (value [i])));
962
963void
964single_fp_config (OpenCL::Device self)
965 ALIAS:
966 single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG
967 double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG
968 half_fp_config = CL_DEVICE_HALF_FP_CONFIG
969 PPCODE:
970 cl_device_fp_config value [1];
971 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
972 EXTEND (SP, 1);
973 const int i = 0;
974 PUSHs (sv_2mortal (newSVuv (value [i])));
975
976void
977global_mem_cache_type (OpenCL::Device self)
978 PPCODE:
979 cl_device_mem_cache_type value [1];
980 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_GLOBAL_MEM_CACHE_TYPE, sizeof (value), value, 0));
981 EXTEND (SP, 1);
982 const int i = 0;
983 PUSHs (sv_2mortal (newSVuv (value [i])));
984
985void
986local_mem_type (OpenCL::Device self)
987 PPCODE:
988 cl_device_local_mem_type value [1];
989 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_LOCAL_MEM_TYPE, sizeof (value), value, 0));
990 EXTEND (SP, 1);
991 const int i = 0;
992 PUSHs (sv_2mortal (newSVuv (value [i])));
993
994void
995error_correction_support (OpenCL::Device self)
996 ALIAS:
997 error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT
998 endian_little = CL_DEVICE_ENDIAN_LITTLE
999 available = CL_DEVICE_AVAILABLE
1000 compiler_available = CL_DEVICE_COMPILER_AVAILABLE
1001 host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY
1002 PPCODE:
1003 cl_bool value [1];
1004 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1005 EXTEND (SP, 1);
1006 const int i = 0;
1007 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
1008
1009void
1010execution_capabilities (OpenCL::Device self)
1011 PPCODE:
1012 cl_device_exec_capabilities value [1];
1013 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_EXECUTION_CAPABILITIES, sizeof (value), value, 0));
1014 EXTEND (SP, 1);
1015 const int i = 0;
1016 PUSHs (sv_2mortal (newSVuv (value [i])));
1017
1018void
1019properties (OpenCL::Device self)
1020 PPCODE:
1021 cl_command_queue_properties value [1];
1022 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0));
1023 EXTEND (SP, 1);
1024 const int i = 0;
1025 PUSHs (sv_2mortal (newSVuv (value [i])));
1026
1027void
1028platform (OpenCL::Device self)
1029 PPCODE:
1030 cl_platform_id value [1];
1031 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0));
1032 EXTEND (SP, 1);
1033 const int i = 0;
1034 {
1035 PUSH_CLOBJ (stash_platform, value [i]);
1036 }
1037
1038void
1039name (OpenCL::Device self)
1040 ALIAS:
1041 name = CL_DEVICE_NAME
1042 vendor = CL_DEVICE_VENDOR
1043 driver_version = CL_DRIVER_VERSION
1044 profile = CL_DEVICE_PROFILE
1045 version = CL_DEVICE_VERSION
1046 extensions = CL_DEVICE_EXTENSIONS
1047 PPCODE:
1048 size_t size;
1049 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1050 char *value = tmpbuf (size);
1051 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1052 EXTEND (SP, 1);
1053 const int i = 0;
1054 PUSHs (sv_2mortal (newSVpv (value, 0)));
1055
1056void
1057parent_device_ext (OpenCL::Device self)
1058 PPCODE:
1059 cl_device_id value [1];
1060 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0));
1061 EXTEND (SP, 1);
1062 const int i = 0;
1063 {
1064 PUSH_CLOBJ (stash_device, value [i]);
1065 }
1066
1067void
1068partition_types_ext (OpenCL::Device self)
1069 ALIAS:
1070 partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT
1071 affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT
1072 partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT
1073 PPCODE:
1074 size_t size;
1075 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1076 cl_device_partition_property_ext *value = tmpbuf (size);
1077 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1078 int i, n = size / sizeof (*value);
1079 EXTEND (SP, n);
1080 for (i = 0; i < n; ++i)
1081 PUSHs (sv_2mortal (newSVuv (value [i])));
1082
1083#END:device
286 1084
287MODULE = OpenCL PACKAGE = OpenCL::Context 1085MODULE = OpenCL PACKAGE = OpenCL::Context
288 1086
289void 1087void
290DESTROY (OpenCL::Context context) 1088DESTROY (OpenCL::Context context)
291 CODE: 1089 CODE:
292 clReleaseContext (context); 1090 clReleaseContext (context);
293 1091
294void 1092void
295info (OpenCL::Context this, cl_context_info name) 1093info (OpenCL::Context self, cl_context_info name)
296 PPCODE: 1094 PPCODE:
297 INFO (Context) 1095 INFO (Context)
298 1096
299void 1097void
300command_queue_simple (OpenCL::Context this, OpenCL::Device device) 1098queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0)
301 PPCODE: 1099 PPCODE:
302{ 1100 NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res));
303 cl_int res; 1101 XPUSH_CLOBJ (stash_queue, queue);
304 cl_command_queue queue = clCreateCommandQueue (this, device, 0, &res);
305 1102
306 if (res)
307 FAIL (CreateCommandQueue, res);
308
309 XPUSH_NEW_OBJ ("OpenCL::Queue", queue);
310}
311
312void 1103void
313user_event (OpenCL::Context this) 1104user_event (OpenCL::Context self)
314 PPCODE: 1105 PPCODE:
315{ 1106 NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res));
316 cl_int res; 1107 XPUSH_CLOBJ (stash_userevent, ev);
317 cl_event ev = clCreateUserEvent (this, &res);
318 1108
319 if (res)
320 FAIL (CreateUserevent, res);
321
322 XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev);
323}
324
325void 1109void
326buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) 1110buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
327 PPCODE: 1111 PPCODE:
328{
329 cl_int res;
330 cl_mem mem;
331
332 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 1112 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?"); 1113 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
334 1114
335 mem = clCreateBuffer (this, flags, len, 0, &res); 1115 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
1116 XPUSH_CLOBJ (stash_bufferobj, mem);
336 1117
337 if (res)
338 FAIL (CreateBuffer, res);
339
340 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
341}
342
343void 1118void
344buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) 1119buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
345 PPCODE: 1120 PPCODE:
346{
347 STRLEN len; 1121 STRLEN len;
348 char *ptr = SvPVbyte (data, len); 1122 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))) 1123 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?"); 1124 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
354 1125 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
355 mem = clCreateBuffer (this, flags, len, ptr, &res); 1126 XPUSH_CLOBJ (stash_bufferobj, mem);
356 1127
357 if (res) 1128#if CL_VERSION_1_2
358 FAIL (CreateBuffer, res);
359 1129
360 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
361}
362
363void 1130void
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) 1131image (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: 1132 PPCODE:
366{
367 STRLEN len; 1133 STRLEN len;
368 char *ptr = SvPVbyte (data, len); 1134 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
369 const cl_image_format format = { channel_order, channel_type }; 1135 const cl_image_format format = { channel_order, channel_type };
370 cl_int res; 1136 const cl_image_desc desc = {
371 cl_mem mem = clCreateImage2D (this, flags, &format, width, height, len / height, ptr, &res); 1137 type,
1138 width, height, depth,
1139 array_size, row_pitch, slice_pitch,
1140 num_mip_level, num_samples,
1141 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0
1142 };
1143 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1144 HV *stash = stash_image;
1145 switch (type)
1146 {
1147 case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break;
1148 case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break;
1149 case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break;
1150 case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break;
1151 case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break;
1152 case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break;
1153 }
1154 XPUSH_CLOBJ (stash, mem);
372 1155
373 if (res) 1156#endif
374 FAIL (CreateImage2D, res);
375 1157
376 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
377}
378
379void 1158void
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) 1159image2d (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: 1160 PPCODE:
382{
383 STRLEN len; 1161 STRLEN len;
384 char *ptr = SvPVbyte (data, len); 1162 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
385 const cl_image_format format = { channel_order, channel_type }; 1163 const cl_image_format format = { channel_order, channel_type };
386 cl_int res; 1164#if PREFER_1_1
387 cl_mem mem = clCreateImage3D (this, flags, &format, width, height, 1165 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); 1166#else
1167 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
1168 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1169#endif
1170 XPUSH_CLOBJ (stash_image2d, mem);
389 1171
390 if (res)
391 FAIL (CreateImage3D, res);
392
393 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
394}
395
396void 1172void
1173image3d (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)
1174 PPCODE:
1175 STRLEN len;
1176 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1177 const cl_image_format format = { channel_order, channel_type };
1178#if PREFER_1_1
1179 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
1180#else
1181 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
1182 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1183#endif
1184 XPUSH_CLOBJ (stash_image3d, mem);
1185
1186#if cl_apple_gl_sharing || cl_khr_gl_sharing
1187
1188void
1189gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
1190 PPCODE:
1191 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
1192 XPUSH_CLOBJ (stash_bufferobj, mem);
1193
1194void
1195gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
1196 PPCODE:
1197 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
1198 XPUSH_CLOBJ (stash_image2d, mem);
1199
1200#if CL_VERSION_1_2
1201
1202void
1203gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1204 ALIAS:
1205 PPCODE:
1206 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
1207 cl_gl_object_type type;
1208 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
1209 HV *stash = stash_memory;
1210 switch (type)
1211 {
1212 case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break;
1213 case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break;
1214 case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break;
1215 case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break;
1216 case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break;
1217 case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break;
1218 }
1219 XPUSH_CLOBJ (stash, mem);
1220
1221#endif
1222
1223void
1224gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1225 PPCODE:
1226#if PREFER_1_1
1227 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
1228#else
1229 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1230#endif
1231 XPUSH_CLOBJ (stash_image2d, mem);
1232
1233void
1234gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1235 PPCODE:
1236#if PREFER_1_1
1237 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
1238#else
1239 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1240#endif
1241 XPUSH_CLOBJ (stash_image3d, mem);
1242
1243#endif
1244
1245void
397supported_image_formats (OpenCL::Context this, cl_mem_flags flags, cl_mem_object_type image_type) 1246supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
398 PPCODE: 1247 PPCODE:
399{ 1248{
400 cl_uint count; 1249 cl_uint count;
401 cl_image_format *list; 1250 cl_image_format *list;
402 int i; 1251 int i;
403 1252
404 NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, 0, 0, &count)); 1253 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count));
405 Newx (list, count, cl_image_format); 1254 Newx (list, count, cl_image_format);
406 NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, count, list, 0)); 1255 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0));
407 1256
408 EXTEND (SP, count); 1257 EXTEND (SP, count);
409 for (i = 0; i < count; ++i) 1258 for (i = 0; i < count; ++i)
410 { 1259 {
411 AV *av = newAV (); 1260 AV *av = newAV ();
414 PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); 1263 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
415 } 1264 }
416} 1265}
417 1266
418void 1267void
419sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) 1268sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode)
420 PPCODE: 1269 PPCODE:
421{
422 cl_int res;
423 cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); 1270 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); 1271 XPUSH_CLOBJ (stash_sampler, sampler);
429}
430 1272
431void 1273void
432program_with_source (OpenCL::Context this, SV *program) 1274program_with_source (OpenCL::Context self, SV *program)
433 PPCODE: 1275 PPCODE:
434{
435 STRLEN len; 1276 STRLEN len;
436 size_t len2; 1277 size_t len2;
437 const char *ptr = SvPVbyte (program, len); 1278 const char *ptr = SvPVbyte (program, len);
438 cl_int res;
439 cl_program prog;
440 1279
441 len2 = len; 1280 len2 = len;
442 prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); 1281 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); 1282 XPUSH_CLOBJ (stash_program, prog);
448} 1283
1284#BEGIN:context
1285
1286void
1287reference_count (OpenCL::Context self)
1288 ALIAS:
1289 reference_count = CL_CONTEXT_REFERENCE_COUNT
1290 num_devices = CL_CONTEXT_NUM_DEVICES
1291 PPCODE:
1292 cl_uint value [1];
1293 NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0));
1294 EXTEND (SP, 1);
1295 const int i = 0;
1296 PUSHs (sv_2mortal (newSVuv (value [i])));
1297
1298void
1299devices (OpenCL::Context self)
1300 PPCODE:
1301 size_t size;
1302 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size));
1303 cl_device_id *value = tmpbuf (size);
1304 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0));
1305 int i, n = size / sizeof (*value);
1306 EXTEND (SP, n);
1307 for (i = 0; i < n; ++i)
1308 {
1309 PUSH_CLOBJ (stash_device, value [i]);
1310 }
1311
1312void
1313properties (OpenCL::Context self)
1314 PPCODE:
1315 size_t size;
1316 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size));
1317 cl_context_properties *value = tmpbuf (size);
1318 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0));
1319 int i, n = size / sizeof (*value);
1320 EXTEND (SP, n);
1321 for (i = 0; i < n; ++i)
1322 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
1323
1324#END:context
449 1325
450MODULE = OpenCL PACKAGE = OpenCL::Queue 1326MODULE = OpenCL PACKAGE = OpenCL::Queue
451 1327
452void 1328void
453DESTROY (OpenCL::Queue this) 1329DESTROY (OpenCL::Queue self)
454 CODE: 1330 CODE:
455 clReleaseCommandQueue (this); 1331 clReleaseCommandQueue (self);
456 1332
457void 1333void
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, ...) 1334read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1335 ALIAS:
1336 enqueue_read_buffer = 0
464 PPCODE: 1337 PPCODE:
465{
466 cl_event ev = 0; 1338 cl_event ev = 0;
467 EVENT_LIST (6, items - 6); 1339 EVENT_LIST (6);
468 1340
469 SvUPGRADE (data, SVt_PV); 1341 SvUPGRADE (data, SVt_PV);
470 SvGROW (data, len); 1342 SvGROW (data, len);
471 SvPOK_only (data); 1343 SvPOK_only (data);
472 SvCUR_set (data, len); 1344 SvCUR_set (data, len);
473 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1345 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
474 1346
475 if (ev) 1347 if (ev)
476 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1348 XPUSH_CLOBJ (stash_event, ev);
477}
478 1349
479void 1350void
480enqueue_write_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) 1351write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1352 ALIAS:
1353 enqueue_write_buffer = 0
481 PPCODE: 1354 PPCODE:
482{
483 cl_event ev = 0; 1355 cl_event ev = 0;
484 STRLEN len; 1356 STRLEN len;
485 char *ptr = SvPVbyte (data, len); 1357 char *ptr = SvPVbyte (data, len);
486 EVENT_LIST (5, items - 5); 1358 EVENT_LIST (5);
487 1359
488 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1360 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
489 1361
490 if (ev) 1362 if (ev)
491 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1363 XPUSH_CLOBJ (stash_event, ev);
492}
493 1364
1365#if CL_VERSION_1_2
1366
494void 1367void
495enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1368fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1369 ALIAS:
1370 enqueue_fill_buffer = 0
496 PPCODE: 1371 PPCODE:
497{
498 cl_event ev = 0; 1372 cl_event ev = 0;
1373 STRLEN len;
1374 char *ptr = SvPVbyte (data, len);
499 EVENT_LIST (6, items - 6); 1375 EVENT_LIST (5);
500 1376
501 NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1377 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
502 1378
503 if (ev) 1379 if (ev)
504 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1380 XPUSH_CLOBJ (stash_event, ev);
505}
506 1381
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 1382void
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, ...) 1383fill_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, ...)
1384 ALIAS:
1385 enqueue_fill_image = 0
512 PPCODE: 1386 PPCODE:
513{
514 cl_event ev = 0; 1387 cl_event ev = 0;
1388 STRLEN len;
1389 const size_t origin [3] = { x, y, z };
1390 const size_t region [3] = { width, height, depth };
1391 EVENT_LIST (12);
1392
1393 const cl_float c_f [4] = { r, g, b, a };
1394 const cl_uint c_u [4] = { r, g, b, a };
1395 const cl_int c_s [4] = { r, g, b, a };
1396 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1397 static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 };
1398 cl_image_format format;
1399 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1400 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1401 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1402 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1403
1404 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1405 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1406
1407 if (ev)
1408 XPUSH_CLOBJ (stash_event, ev);
1409
1410#endif
1411
1412void
1413copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1414 ALIAS:
1415 enqueue_copy_buffer = 0
1416 PPCODE:
1417 cl_event ev = 0;
1418 EVENT_LIST (6);
1419
1420 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1421
1422 if (ev)
1423 XPUSH_CLOBJ (stash_event, ev);
1424
1425void
1426read_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, ...)
1427 ALIAS:
1428 enqueue_read_buffer_rect = 0
1429 PPCODE:
1430 cl_event ev = 0;
1431 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 }; 1432 const size_t host_origin[3] = { host_x, host_y, host_z };
516 const size_t region[3] = { width, height, depth }; 1433 const size_t region[3] = { width, height, depth };
1434 EVENT_LIST (17);
1435
1436 if (!buf_row_pitch)
1437 buf_row_pitch = region [0];
1438
1439 if (!buf_slice_pitch)
1440 buf_slice_pitch = region [1] * buf_row_pitch;
1441
1442 if (!host_row_pitch)
1443 host_row_pitch = region [0];
1444
1445 if (!host_slice_pitch)
1446 host_slice_pitch = region [1] * host_row_pitch;
1447
517 size_t len = row_pitch * slice_pitch * depth; 1448 size_t len = host_row_pitch * host_slice_pitch * region [2];
518 EVENT_LIST (11, items - 11);
519 1449
520 SvUPGRADE (data, SVt_PV); 1450 SvUPGRADE (data, SVt_PV);
521 SvGROW (data, len); 1451 SvGROW (data, len);
522 SvPOK_only (data); 1452 SvPOK_only (data);
523 SvCUR_set (data, len); 1453 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)); 1454 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 1455
526 if (ev) 1456 if (ev)
527 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1457 XPUSH_CLOBJ (stash_event, ev);
528}
529 1458
530void 1459void
1460write_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, ...)
1461 ALIAS:
1462 enqueue_write_buffer_rect = 0
1463 PPCODE:
1464 cl_event ev = 0;
1465 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1466 const size_t host_origin[3] = { host_x, host_y, host_z };
1467 const size_t region[3] = { width, height, depth };
1468 STRLEN len;
1469 char *ptr = SvPVbyte (data, len);
1470 EVENT_LIST (17);
1471
1472 if (!buf_row_pitch)
1473 buf_row_pitch = region [0];
1474
1475 if (!buf_slice_pitch)
1476 buf_slice_pitch = region [1] * buf_row_pitch;
1477
1478 if (!host_row_pitch)
1479 host_row_pitch = region [0];
1480
1481 if (!host_slice_pitch)
1482 host_slice_pitch = region [1] * host_row_pitch;
1483
1484 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
1485
1486 if (len < min_len)
1487 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1488
1489 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));
1490
1491 if (ev)
1492 XPUSH_CLOBJ (stash_event, ev);
1493
1494void
1495copy_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, ...)
1496 ALIAS:
1497 enqueue_copy_buffer_rect = 0
1498 PPCODE:
1499 cl_event ev = 0;
1500 const size_t src_origin[3] = { src_x, src_y, src_z };
1501 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1502 const size_t region[3] = { width, height, depth };
1503 EVENT_LIST (16);
1504
1505 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));
1506
1507 if (ev)
1508 XPUSH_CLOBJ (stash_event, ev);
1509
1510void
1511read_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, ...)
1512 ALIAS:
1513 enqueue_read_image = 0
1514 PPCODE:
1515 cl_event ev = 0;
1516 const size_t src_origin[3] = { src_x, src_y, src_z };
1517 const size_t region[3] = { width, height, depth };
1518 EVENT_LIST (12);
1519
1520 if (!row_pitch)
1521 row_pitch = img_row_pitch (src);
1522
1523 if (depth > 1 && !slice_pitch)
1524 slice_pitch = row_pitch * height;
1525
1526 size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1527
1528 SvUPGRADE (data, SVt_PV);
1529 SvGROW (data, len);
1530 SvPOK_only (data);
1531 SvCUR_set (data, len);
1532 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));
1533
1534 if (ev)
1535 XPUSH_CLOBJ (stash_event, ev);
1536
1537void
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, ...) 1538write_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, ...)
1539 ALIAS:
1540 enqueue_write_image = 0
532 PPCODE: 1541 PPCODE:
533{
534 cl_event ev = 0; 1542 cl_event ev = 0;
535 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1543 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
536 const size_t region[3] = { width, height, depth }; 1544 const size_t region[3] = { width, height, depth };
537 STRLEN len; 1545 STRLEN len;
538 char *ptr = SvPVbyte (data, len); 1546 char *ptr = SvPVbyte (data, len);
1547 EVENT_LIST (12);
1548
1549 if (!row_pitch)
1550 row_pitch = img_row_pitch (dst);
1551
1552 if (depth > 1 && !slice_pitch)
539 size_t slice_pitch = len / (row_pitch * height); 1553 slice_pitch = row_pitch * height;
540 EVENT_LIST (11, items - 11);
541 1554
1555 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1556
1557 if (len < min_len)
1558 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1559
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)); 1560 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 1561
544 if (ev) 1562 if (ev)
545 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1563 XPUSH_CLOBJ (stash_event, ev);
546}
547 1564
548void 1565void
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, ...) 1566copy_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, ...)
1567 ALIAS:
1568 enqueue_copy_image = 0
550 PPCODE: 1569 PPCODE:
551{
552 cl_event ev = 0; 1570 cl_event ev = 0;
553 const size_t src_origin[3] = { src_x, src_y, src_z }; 1571 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 }; 1572 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
555 const size_t region[3] = { width, height, depth }; 1573 const size_t region[3] = { width, height, depth };
556 EVENT_LIST (16, items - 16); 1574 EVENT_LIST (12);
557 1575
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)); 1576 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
559 1577
560 if (ev) 1578 if (ev)
561 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1579 XPUSH_CLOBJ (stash_event, ev);
562}
563 1580
564void 1581void
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, ...) 1582copy_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, ...)
1583 ALIAS:
1584 enqueue_copy_image_to_buffer = 0
566 PPCODE: 1585 PPCODE:
567{
568 cl_event ev = 0; 1586 cl_event ev = 0;
1587 const size_t src_origin[3] = { src_x, src_y, src_z };
1588 const size_t region [3] = { width, height, depth };
1589 EVENT_LIST (10);
1590
1591 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1592
1593 if (ev)
1594 XPUSH_CLOBJ (stash_event, ev);
1595
1596void
1597copy_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, ...)
1598 ALIAS:
1599 enqueue_copy_buffer_to_image = 0
1600 PPCODE:
1601 cl_event ev = 0;
569 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1602 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1603 const size_t region [3] = { width, height, depth };
1604 EVENT_LIST (10);
1605
1606 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1607
1608 if (ev)
1609 XPUSH_CLOBJ (stash_event, ev);
1610
1611void
1612map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, size_t cb = 0, ...)
1613 ALIAS:
1614 enqueue_map_buffer = 0
1615 PPCODE:
1616 cl_event ev;
1617 EVENT_LIST (7);
1618
1619 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
1620 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0));
1621
1622void
1623map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, size_t width = 0, size_t height = 0, size_t depth = 0, ...)
1624 ALIAS:
1625 enqueue_map_image = 0
1626 PPCODE:
1627 cl_event ev;
1628 const size_t origin[3] = { x, y, z };
570 const size_t region[3] = { width, height, depth }; 1629 const size_t region[3] = { width, height, depth };
1630 size_t row_pitch, slice_pitch;
571 EVENT_LIST (10, items - 10); 1631 EVENT_LIST (11);
572 1632
573 NEED_SUCCESS (EnqueueCopyBufferToImage, (this, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1633 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));
1634
1635 size_t cb = slice_pitch ? slice_pitch * region [2]
1636 : row_pitch ? row_pitch * region [1]
1637 : region [0];
1638
1639 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch));
1640
1641void
1642unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1643 PPCODE:
1644 cl_uint event_list_count = items - 2;
1645 cl_event *event_list_ptr = event_list (&ST (2), &event_list_count, mapped->event);
1646
1647 NEED_SUCCESS (EnqueueUnmapMemObject, (self, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &mapped->event));
1648 mapped_detach (ST (1), mapped);
1649
1650 if (GIMME_V != G_VOID)
1651 XPUSH_CLOBJ (stash_event, mapped->event);
1652
1653void
1654task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1655 ALIAS:
1656 enqueue_task = 0
1657 PPCODE:
1658 cl_event ev = 0;
1659 EVENT_LIST (2);
1660
1661 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
574 1662
575 if (ev) 1663 if (ev)
576 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1664 XPUSH_CLOBJ (stash_event, ev);
577}
578 1665
579void 1666void
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, ...) 1667nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1668 ALIAS:
1669 enqueue_nd_range_kernel = 0
625 PPCODE: 1670 PPCODE:
626{
627 cl_event ev = 0; 1671 cl_event ev = 0;
628 size_t *gwo = 0, *gws, *lws = 0; 1672 size_t *gwo = 0, *gws, *lws = 0;
629 int gws_len; 1673 int gws_len;
630 size_t *lists; 1674 size_t *lists;
631 int i; 1675 int i;
632 EVENT_LIST (5, items - 5); 1676 EVENT_LIST (5);
633 1677
634 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) 1678 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
635 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1679 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
636 1680
637 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1681 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
638 1682
639 lists = tmpbuf (sizeof (size_t) * 3 * gws_len); 1683 lists = tmpbuf (sizeof (size_t) * 3 * gws_len);
640 1684
641 gws = lists + gws_len * 0; 1685 gws = lists + gws_len * 0;
642 for (i = 0; i < gws_len; ++i) 1686 for (i = 0; i < gws_len; ++i)
1687 {
643 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); 1688 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]);
1689 // at least nvidia crashes for 0-sized work group sizes, work around
1690 if (!gws [i])
1691 croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i);
1692 }
644 1693
645 if (SvOK (global_work_offset)) 1694 if (SvOK (global_work_offset))
646 { 1695 {
647 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) 1696 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"); 1697 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference");
655 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1704 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
656 } 1705 }
657 1706
658 if (SvOK (local_work_size)) 1707 if (SvOK (local_work_size))
659 { 1708 {
660 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1709 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"); 1710 croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference");
662 1711
663 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1712 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"); 1713 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
665 1714
666 lws = lists + gws_len * 2; 1715 lws = lists + gws_len * 2;
667 for (i = 0; i < gws_len; ++i) 1716 for (i = 0; i < gws_len; ++i)
1717 {
668 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); 1718 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]);
1719 // at least nvidia crashes for 0-sized work group sizes, work around
1720 if (!lws [i])
1721 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
1722 }
669 } 1723 }
670 1724
671 NEED_SUCCESS (EnqueueNDRangeKernel, (this, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1725 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
672 1726
673 if (ev) 1727 if (ev)
674 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1728 XPUSH_CLOBJ (stash_event, ev);
675}
676 1729
1730#if cl_apple_gl_sharing || cl_khr_gl_sharing
1731
677void 1732void
678enqueue_marker (OpenCL::Queue this) 1733acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
1734 ALIAS:
1735 release_gl_objects = 1
1736 enqueue_acquire_gl_objects = 0
1737 enqueue_release_gl_objects = 1
679 PPCODE: 1738 PPCODE:
680{ 1739 if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV)
1740 croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call");
1741
681 cl_event ev; 1742 cl_event ev = 0;
682 NEED_SUCCESS (EnqueueMarker, (this, &ev)); 1743 EVENT_LIST (2);
683 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1744 AV *av = (AV *)SvRV (objects);
684} 1745 cl_uint num_objects = av_len (av) + 1;
1746 cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects);
1747 int i;
685 1748
1749 for (i = num_objects; i--; )
1750 object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory");
1751
1752 if (ix)
1753 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1754 else
1755 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1756
1757 if (ev)
1758 XPUSH_CLOBJ (stash_event, ev);
1759
1760#endif
1761
686void 1762void
687enqueue_wait_for_events (OpenCL::Queue this, ...) 1763wait_for_events (OpenCL::Queue self, ...)
1764 ALIAS:
1765 enqueue_wait_for_events = 0
688 CODE: 1766 CODE:
689{
690 EVENT_LIST (1, items - 1); 1767 EVENT_LIST (1);
1768#if PREFER_1_1
691 NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); 1769 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
692} 1770#else
1771 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0));
1772#endif
693 1773
694void 1774void
695enqueue_barrier (OpenCL::Queue this) 1775marker (OpenCL::Queue self, ...)
1776 ALIAS:
1777 enqueue_marker = 0
696 CODE: 1778 PPCODE:
1779 cl_event ev = 0;
1780 EVENT_LIST (1);
1781#if PREFER_1_1
1782 if (!event_list_count)
1783 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1784 else
1785#if CL_VERSION_1_2
1786 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1787#else
1788 {
1789 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier
1790 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
1791 }
1792#endif
1793#else
1794 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1795#endif
1796 if (ev)
1797 XPUSH_CLOBJ (stash_event, ev);
1798
1799void
1800barrier (OpenCL::Queue self, ...)
1801 ALIAS:
1802 enqueue_barrier = 0
1803 PPCODE:
1804 cl_event ev = 0;
1805 EVENT_LIST (1);
1806#if PREFER_1_1
1807 if (!event_list_count && GIMME_V == G_VOID)
697 NEED_SUCCESS (EnqueueBarrier, (this)); 1808 NEED_SUCCESS (EnqueueBarrier, (self));
1809 else
1810#if CL_VERSION_1_2
1811 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1812#else
1813 {
1814 if (event_list_count)
1815 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
698 1816
1817 if (GIMME_V != G_VOID)
1818 NEED_SUCCESS (EnqueueMarker, (self, &ev));
1819 }
1820#endif
1821#else
1822 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1823#endif
1824 if (ev)
1825 XPUSH_CLOBJ (stash_event, ev);
1826
699void 1827void
700flush (OpenCL::Queue this) 1828flush (OpenCL::Queue self)
701 CODE: 1829 CODE:
702 NEED_SUCCESS (Flush, (this)); 1830 NEED_SUCCESS (Flush, (self));
703 1831
704void 1832void
705finish (OpenCL::Queue this) 1833finish (OpenCL::Queue self)
706 CODE: 1834 CODE:
707 NEED_SUCCESS (Finish, (this)); 1835 NEED_SUCCESS (Finish, (self));
1836
1837void
1838info (OpenCL::Queue self, cl_command_queue_info name)
1839 PPCODE:
1840 INFO (CommandQueue)
1841
1842#BEGIN:command_queue
1843
1844void
1845context (OpenCL::Queue self)
1846 PPCODE:
1847 cl_context value [1];
1848 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0));
1849 EXTEND (SP, 1);
1850 const int i = 0;
1851 {
1852 NEED_SUCCESS (RetainContext, (value [i]));
1853 PUSH_CLOBJ (stash_context, value [i]);
1854 }
1855
1856void
1857device (OpenCL::Queue self)
1858 PPCODE:
1859 cl_device_id value [1];
1860 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0));
1861 EXTEND (SP, 1);
1862 const int i = 0;
1863 {
1864 PUSH_CLOBJ (stash_device, value [i]);
1865 }
1866
1867void
1868reference_count (OpenCL::Queue self)
1869 PPCODE:
1870 cl_uint value [1];
1871 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0));
1872 EXTEND (SP, 1);
1873 const int i = 0;
1874 PUSHs (sv_2mortal (newSVuv (value [i])));
1875
1876void
1877properties (OpenCL::Queue self)
1878 PPCODE:
1879 cl_command_queue_properties value [1];
1880 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0));
1881 EXTEND (SP, 1);
1882 const int i = 0;
1883 PUSHs (sv_2mortal (newSVuv (value [i])));
1884
1885#END:command_queue
708 1886
709MODULE = OpenCL PACKAGE = OpenCL::Memory 1887MODULE = OpenCL PACKAGE = OpenCL::Memory
710 1888
711void 1889void
712DESTROY (OpenCL::Memory this) 1890DESTROY (OpenCL::Memory self)
713 CODE: 1891 CODE:
714 clReleaseMemObject (this); 1892 clReleaseMemObject (self);
715 1893
716void 1894void
717info (OpenCL::Memory this, cl_mem_info name) 1895info (OpenCL::Memory self, cl_mem_info name)
718 PPCODE: 1896 PPCODE:
719 INFO (MemObject) 1897 INFO (MemObject)
720 1898
1899#BEGIN:mem
1900
1901void
1902type (OpenCL::Memory self)
1903 PPCODE:
1904 cl_mem_object_type value [1];
1905 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0));
1906 EXTEND (SP, 1);
1907 const int i = 0;
1908 PUSHs (sv_2mortal (newSVuv (value [i])));
1909
1910void
1911flags (OpenCL::Memory self)
1912 PPCODE:
1913 cl_mem_flags value [1];
1914 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0));
1915 EXTEND (SP, 1);
1916 const int i = 0;
1917 PUSHs (sv_2mortal (newSVuv (value [i])));
1918
1919void
1920size (OpenCL::Memory self)
1921 ALIAS:
1922 size = CL_MEM_SIZE
1923 offset = CL_MEM_OFFSET
1924 PPCODE:
1925 size_t value [1];
1926 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
1927 EXTEND (SP, 1);
1928 const int i = 0;
1929 PUSHs (sv_2mortal (newSVuv (value [i])));
1930
1931void
1932host_ptr (OpenCL::Memory self)
1933 PPCODE:
1934 void * value [1];
1935 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0));
1936 EXTEND (SP, 1);
1937 const int i = 0;
1938 PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i])));
1939
1940void
1941map_count (OpenCL::Memory self)
1942 ALIAS:
1943 map_count = CL_MEM_MAP_COUNT
1944 reference_count = CL_MEM_REFERENCE_COUNT
1945 PPCODE:
1946 cl_uint value [1];
1947 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
1948 EXTEND (SP, 1);
1949 const int i = 0;
1950 PUSHs (sv_2mortal (newSVuv (value [i])));
1951
1952void
1953context (OpenCL::Memory self)
1954 PPCODE:
1955 cl_context value [1];
1956 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0));
1957 EXTEND (SP, 1);
1958 const int i = 0;
1959 {
1960 NEED_SUCCESS (RetainContext, (value [i]));
1961 PUSH_CLOBJ (stash_context, value [i]);
1962 }
1963
1964void
1965associated_memobject (OpenCL::Memory self)
1966 PPCODE:
1967 cl_mem value [1];
1968 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0));
1969 EXTEND (SP, 1);
1970 const int i = 0;
1971 {
1972 NEED_SUCCESS (RetainMemObject, (value [i]));
1973 PUSH_CLOBJ (stash_memory, value [i]);
1974 }
1975
1976#END:mem
1977
1978#if cl_apple_gl_sharing || cl_khr_gl_sharing
1979
1980void
1981gl_object_info (OpenCL::Memory self)
1982 PPCODE:
1983 cl_gl_object_type type;
1984 cl_GLuint name;
1985 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
1986 EXTEND (SP, 2);
1987 PUSHs (sv_2mortal (newSVuv (type)));
1988 PUSHs (sv_2mortal (newSVuv (name)));
1989
1990#endif
1991
1992MODULE = OpenCL PACKAGE = OpenCL::BufferObj
1993
1994void
1995sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
1996 PPCODE:
1997 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR))
1998 croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!");
1999
2000 cl_buffer_region crdata = { origin, size };
2001
2002 NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res));
2003 XPUSH_CLOBJ (stash_buffer, mem);
2004
2005MODULE = OpenCL PACKAGE = OpenCL::Image
2006
2007void
2008image_info (OpenCL::Image self, cl_image_info name)
2009 PPCODE:
2010 INFO (Image)
2011
2012void
2013format (OpenCL::Image self)
2014 PPCODE:
2015 cl_image_format format;
2016 NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
2017 EXTEND (SP, 2);
2018 PUSHs (sv_2mortal (newSVuv (format.image_channel_order)));
2019 PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type)));
2020
2021#BEGIN:image
2022
2023void
2024element_size (OpenCL::Image self)
2025 ALIAS:
2026 element_size = CL_IMAGE_ELEMENT_SIZE
2027 row_pitch = CL_IMAGE_ROW_PITCH
2028 slice_pitch = CL_IMAGE_SLICE_PITCH
2029 width = CL_IMAGE_WIDTH
2030 height = CL_IMAGE_HEIGHT
2031 depth = CL_IMAGE_DEPTH
2032 PPCODE:
2033 size_t value [1];
2034 NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0));
2035 EXTEND (SP, 1);
2036 const int i = 0;
2037 PUSHs (sv_2mortal (newSVuv (value [i])));
2038
2039#END:image
2040
2041#if cl_apple_gl_sharing || cl_khr_gl_sharing
2042
2043#BEGIN:gl_texture
2044
2045void
2046target (OpenCL::Image self)
2047 PPCODE:
2048 cl_GLenum value [1];
2049 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
2050 EXTEND (SP, 1);
2051 const int i = 0;
2052 PUSHs (sv_2mortal (newSVuv (value [i])));
2053
2054void
2055gl_mipmap_level (OpenCL::Image self)
2056 PPCODE:
2057 cl_GLint value [1];
2058 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
2059 EXTEND (SP, 1);
2060 const int i = 0;
2061 PUSHs (sv_2mortal (newSViv (value [i])));
2062
2063#END:gl_texture
2064
2065#endif
2066
721MODULE = OpenCL PACKAGE = OpenCL::Sampler 2067MODULE = OpenCL PACKAGE = OpenCL::Sampler
722 2068
723void 2069void
724DESTROY (OpenCL::Sampler this) 2070DESTROY (OpenCL::Sampler self)
725 CODE: 2071 CODE:
726 clReleaseSampler (this); 2072 clReleaseSampler (self);
727 2073
728void 2074void
729info (OpenCL::Sampler this, cl_sampler_info name) 2075info (OpenCL::Sampler self, cl_sampler_info name)
730 PPCODE: 2076 PPCODE:
731 INFO (Sampler) 2077 INFO (Sampler)
732 2078
2079#BEGIN:sampler
2080
2081void
2082reference_count (OpenCL::Sampler self)
2083 PPCODE:
2084 cl_uint value [1];
2085 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_REFERENCE_COUNT, sizeof (value), value, 0));
2086 EXTEND (SP, 1);
2087 const int i = 0;
2088 PUSHs (sv_2mortal (newSVuv (value [i])));
2089
2090void
2091context (OpenCL::Sampler self)
2092 PPCODE:
2093 cl_context value [1];
2094 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0));
2095 EXTEND (SP, 1);
2096 const int i = 0;
2097 {
2098 NEED_SUCCESS (RetainContext, (value [i]));
2099 PUSH_CLOBJ (stash_context, value [i]);
2100 }
2101
2102void
2103normalized_coords (OpenCL::Sampler self)
2104 PPCODE:
2105 cl_addressing_mode value [1];
2106 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0));
2107 EXTEND (SP, 1);
2108 const int i = 0;
2109 PUSHs (sv_2mortal (newSVuv (value [i])));
2110
2111void
2112addressing_mode (OpenCL::Sampler self)
2113 PPCODE:
2114 cl_filter_mode value [1];
2115 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0));
2116 EXTEND (SP, 1);
2117 const int i = 0;
2118 PUSHs (sv_2mortal (newSVuv (value [i])));
2119
2120void
2121filter_mode (OpenCL::Sampler self)
2122 PPCODE:
2123 cl_bool value [1];
2124 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0));
2125 EXTEND (SP, 1);
2126 const int i = 0;
2127 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
2128
2129#END:sampler
2130
733MODULE = OpenCL PACKAGE = OpenCL::Program 2131MODULE = OpenCL PACKAGE = OpenCL::Program
734 2132
735void 2133void
736DESTROY (OpenCL::Program this) 2134DESTROY (OpenCL::Program self)
737 CODE: 2135 CODE:
738 clReleaseProgram (this); 2136 clReleaseProgram (self);
739 2137
740void 2138void
741info (OpenCL::Program this, cl_program_info name) 2139build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2140 ALIAS:
2141 build_async = 1
742 PPCODE: 2142 CODE:
743 INFO (Program) 2143 void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0;
2144 void *user_data = 0;
2145 cl_uint num_devices = 0;
2146 cl_device_id *device_list = 0;
744 2147
745void 2148 if (SvOK (devices))
746build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) 2149 {
747 CODE: 2150 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV)
2151 croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects.");
2152
2153 AV *av = (AV *)SvRV (devices);
2154 num_devices = av_len (av) + 1;
2155
2156 if (num_devices)
2157 {
2158 device_list = tmpbuf (sizeof (*device_list) * num_devices);
2159 int count;
2160 for (count = 0; count < num_devices; ++count)
2161 device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device");
2162 }
2163 }
2164
2165 if (SvOK (notify))
2166 {
2167 NEED_SUCCESS (RetainProgram, (self));
2168 pfn_notify = eq_program_notify;
2169 user_data = SvREFCNT_inc (s_get_cv (notify));
2170 }
2171
2172 if (ix)
2173 build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data);
2174 else
748 NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); 2175 NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
749 2176
750void 2177void
751build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) 2178build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
752 PPCODE: 2179 PPCODE:
753{
754 size_t size; 2180 size_t size;
755 SV *sv;
756
757 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); 2181 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size));
758 sv = sv_2mortal (newSV (size)); 2182 SV *sv = sv_2mortal (newSV (size));
759 SvUPGRADE (sv, SVt_PV); 2183 SvUPGRADE (sv, SVt_PV);
760 SvPOK_only (sv); 2184 SvPOK_only (sv);
761 SvCUR_set (sv, size); 2185 SvCUR_set (sv, size);
762 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); 2186 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0));
763 XPUSHs (sv); 2187 XPUSHs (sv);
764} 2188
2189#BEGIN:program_build
2190
2191void
2192build_status (OpenCL::Program self, OpenCL::Device device)
2193 PPCODE:
2194 cl_build_status value [1];
2195 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0));
2196 EXTEND (SP, 1);
2197 const int i = 0;
2198 PUSHs (sv_2mortal (newSViv (value [i])));
2199
2200void
2201build_options (OpenCL::Program self, OpenCL::Device device)
2202 ALIAS:
2203 build_options = CL_PROGRAM_BUILD_OPTIONS
2204 build_log = CL_PROGRAM_BUILD_LOG
2205 PPCODE:
2206 size_t size;
2207 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size));
2208 char *value = tmpbuf (size);
2209 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2210 EXTEND (SP, 1);
2211 const int i = 0;
2212 PUSHs (sv_2mortal (newSVpv (value, 0)));
2213
2214#END:program_build
765 2215
766void 2216void
767kernel (OpenCL::Program program, SV *function) 2217kernel (OpenCL::Program program, SV *function)
768 PPCODE: 2218 PPCODE:
769{
770 cl_int res;
771 cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); 2219 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
2220 XPUSH_CLOBJ (stash_kernel, kernel);
772 2221
773 if (res) 2222void
774 FAIL (CreateKernel, res); 2223kernels_in_program (OpenCL::Program program)
2224 PPCODE:
2225 cl_uint num_kernels;
2226 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
2227 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
2228 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
775 2229
776 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 2230 int i;
777} 2231 EXTEND (SP, num_kernels);
2232 for (i = 0; i < num_kernels; ++i)
2233 PUSH_CLOBJ (stash_kernel, kernels [i]);
2234
2235void
2236info (OpenCL::Program self, cl_program_info name)
2237 PPCODE:
2238 INFO (Program)
2239
2240void
2241binaries (OpenCL::Program self)
2242 PPCODE:
2243 cl_uint n, i;
2244 size_t size;
2245
2246 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0));
2247 if (!n) XSRETURN_EMPTY;
2248
2249 size_t *sizes = tmpbuf (sizeof (*sizes) * n);
2250 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size));
2251 if (size != sizeof (*sizes) * n) XSRETURN_EMPTY;
2252 unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n);
2253
2254 EXTEND (SP, n);
2255 for (i = 0; i < n; ++i)
2256 {
2257 SV *sv = sv_2mortal (newSV (sizes [i]));
2258 SvUPGRADE (sv, SVt_PV);
2259 SvPOK_only (sv);
2260 SvCUR_set (sv, sizes [i]);
2261 ptrs [i] = (void *)SvPVX (sv);
2262 PUSHs (sv);
2263 }
2264
2265 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
2266 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
2267
2268#BEGIN:program
2269
2270void
2271reference_count (OpenCL::Program self)
2272 ALIAS:
2273 reference_count = CL_PROGRAM_REFERENCE_COUNT
2274 num_devices = CL_PROGRAM_NUM_DEVICES
2275 PPCODE:
2276 cl_uint value [1];
2277 NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0));
2278 EXTEND (SP, 1);
2279 const int i = 0;
2280 PUSHs (sv_2mortal (newSVuv (value [i])));
2281
2282void
2283context (OpenCL::Program self)
2284 PPCODE:
2285 cl_context value [1];
2286 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0));
2287 EXTEND (SP, 1);
2288 const int i = 0;
2289 {
2290 NEED_SUCCESS (RetainContext, (value [i]));
2291 PUSH_CLOBJ (stash_context, value [i]);
2292 }
2293
2294void
2295devices (OpenCL::Program self)
2296 PPCODE:
2297 size_t size;
2298 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size));
2299 cl_device_id *value = tmpbuf (size);
2300 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0));
2301 int i, n = size / sizeof (*value);
2302 EXTEND (SP, n);
2303 for (i = 0; i < n; ++i)
2304 {
2305 PUSH_CLOBJ (stash_device, value [i]);
2306 }
2307
2308void
2309source (OpenCL::Program self)
2310 PPCODE:
2311 size_t size;
2312 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size));
2313 char *value = tmpbuf (size);
2314 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0));
2315 EXTEND (SP, 1);
2316 const int i = 0;
2317 PUSHs (sv_2mortal (newSVpv (value, 0)));
2318
2319void
2320binary_sizes (OpenCL::Program self)
2321 PPCODE:
2322 size_t size;
2323 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size));
2324 size_t *value = tmpbuf (size);
2325 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0));
2326 int i, n = size / sizeof (*value);
2327 EXTEND (SP, n);
2328 for (i = 0; i < n; ++i)
2329 PUSHs (sv_2mortal (newSVuv (value [i])));
2330
2331#END:program
778 2332
779MODULE = OpenCL PACKAGE = OpenCL::Kernel 2333MODULE = OpenCL PACKAGE = OpenCL::Kernel
780 2334
781void 2335void
782DESTROY (OpenCL::Kernel this) 2336DESTROY (OpenCL::Kernel self)
783 CODE: 2337 CODE:
784 clReleaseKernel (this); 2338 clReleaseKernel (self);
785 2339
786void 2340void
2341setf (OpenCL::Kernel self, const char *format, ...)
2342 CODE:
2343 int i;
2344 for (i = 2; ; ++i)
2345 {
2346 while (*format == ' ')
2347 ++format;
2348
2349 char type = *format++;
2350
2351 if (!type)
2352 break;
2353
2354 if (i >= items)
2355 croak ("OpenCL::Kernel::setf format string too long (not enough arguments)");
2356
2357 SV *sv = ST (i);
2358
2359 union
2360 {
2361 cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS;
2362 cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL;
2363 cl_half ch; cl_float cf; cl_double cd;
2364 cl_mem cm;
2365 cl_sampler ca;
2366 size_t cz;
2367 cl_event ce;
2368 } arg;
2369 size_t size;
2370 int nullarg = 0;
2371
2372 switch (type)
2373 {
2374 case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break;
2375 case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break;
2376 case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break;
2377 case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break;
2378 case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break;
2379 case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break;
2380 case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break;
2381 case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break;
2382
2383 case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break;
2384 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2385 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2386
2387 case 'z': nullarg = 1; size = SvIV (sv); break;
2388
2389 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2390 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2391 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2392
2393 default:
2394 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2395 }
2396
2397 res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg);
2398 if (res)
2399 croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res));
2400 }
2401
2402 if (i != items)
2403 croak ("OpenCL::Kernel::setf format string too short (too many arguments)");
2404
2405void
2406set_char (OpenCL::Kernel self, cl_uint idx, cl_char value)
2407 CODE:
2408 clSetKernelArg (self, idx, sizeof (value), &value);
2409
2410void
2411set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value)
2412 CODE:
2413 clSetKernelArg (self, idx, sizeof (value), &value);
2414
2415void
2416set_short (OpenCL::Kernel self, cl_uint idx, cl_short value)
2417 CODE:
2418 clSetKernelArg (self, idx, sizeof (value), &value);
2419
2420void
2421set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value)
2422 CODE:
2423 clSetKernelArg (self, idx, sizeof (value), &value);
2424
2425void
2426set_int (OpenCL::Kernel self, cl_uint idx, cl_int value)
2427 CODE:
2428 clSetKernelArg (self, idx, sizeof (value), &value);
2429
2430void
2431set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value)
2432 CODE:
2433 clSetKernelArg (self, idx, sizeof (value), &value);
2434
2435void
2436set_long (OpenCL::Kernel self, cl_uint idx, cl_long value)
2437 CODE:
2438 clSetKernelArg (self, idx, sizeof (value), &value);
2439
2440void
2441set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value)
2442 CODE:
2443 clSetKernelArg (self, idx, sizeof (value), &value);
2444
2445void
2446set_half (OpenCL::Kernel self, cl_uint idx, cl_half value)
2447 CODE:
2448 clSetKernelArg (self, idx, sizeof (value), &value);
2449
2450void
2451set_float (OpenCL::Kernel self, cl_uint idx, cl_float value)
2452 CODE:
2453 clSetKernelArg (self, idx, sizeof (value), &value);
2454
2455void
2456set_double (OpenCL::Kernel self, cl_uint idx, cl_double value)
2457 CODE:
2458 clSetKernelArg (self, idx, sizeof (value), &value);
2459
2460void
2461set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value)
2462 CODE:
2463 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2464
2465void
2466set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value)
2467 CODE:
2468 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2469
2470void
2471set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value)
2472 ALIAS:
2473 set_image2d = 0
2474 set_image3d = 0
2475 CODE:
2476 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2477
2478void
2479set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
2480 CODE:
2481 clSetKernelArg (self, idx, sizeof (value), &value);
2482
2483void
2484set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
2485 CODE:
2486 clSetKernelArg (self, idx, size, 0);
2487
2488void
2489set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
2490 CODE:
2491 clSetKernelArg (self, idx, sizeof (value), &value);
2492
2493void
787info (OpenCL::Kernel this, cl_kernel_info name) 2494info (OpenCL::Kernel self, cl_kernel_info name)
788 PPCODE: 2495 PPCODE:
789 INFO (Kernel) 2496 INFO (Kernel)
790 2497
2498#BEGIN:kernel
2499
791void 2500void
792set_char (OpenCL::Kernel this, cl_uint idx, cl_char value) 2501function_name (OpenCL::Kernel self)
2502 PPCODE:
2503 size_t size;
2504 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size));
2505 char *value = tmpbuf (size);
2506 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0));
2507 EXTEND (SP, 1);
2508 const int i = 0;
2509 PUSHs (sv_2mortal (newSVpv (value, 0)));
2510
2511void
2512num_args (OpenCL::Kernel self)
2513 ALIAS:
2514 num_args = CL_KERNEL_NUM_ARGS
2515 reference_count = CL_KERNEL_REFERENCE_COUNT
2516 PPCODE:
2517 cl_uint value [1];
2518 NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0));
2519 EXTEND (SP, 1);
2520 const int i = 0;
2521 PUSHs (sv_2mortal (newSVuv (value [i])));
2522
2523void
2524context (OpenCL::Kernel self)
2525 PPCODE:
2526 cl_context value [1];
2527 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0));
2528 EXTEND (SP, 1);
2529 const int i = 0;
2530 {
2531 NEED_SUCCESS (RetainContext, (value [i]));
2532 PUSH_CLOBJ (stash_context, value [i]);
2533 }
2534
2535void
2536program (OpenCL::Kernel self)
2537 PPCODE:
2538 cl_program value [1];
2539 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0));
2540 EXTEND (SP, 1);
2541 const int i = 0;
2542 {
2543 NEED_SUCCESS (RetainProgram, (value [i]));
2544 PUSH_CLOBJ (stash_program, value [i]);
2545 }
2546
2547#END:kernel
2548
2549void
2550work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name)
793 CODE: 2551 PPCODE:
794 clSetKernelArg (this, idx, sizeof (value), &value); 2552 size_t size;
2553 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size));
2554 SV *sv = sv_2mortal (newSV (size));
2555 SvUPGRADE (sv, SVt_PV);
2556 SvPOK_only (sv);
2557 SvCUR_set (sv, size);
2558 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0));
2559 XPUSHs (sv);
795 2560
796void 2561#BEGIN:kernel_work_group
797set_uchar (OpenCL::Kernel this, cl_uint idx, cl_uchar value)
798 CODE:
799 clSetKernelArg (this, idx, sizeof (value), &value);
800 2562
801void 2563void
802set_short (OpenCL::Kernel this, cl_uint idx, cl_short value) 2564work_group_size (OpenCL::Kernel self, OpenCL::Device device)
803 CODE: 2565 ALIAS:
804 clSetKernelArg (this, idx, sizeof (value), &value); 2566 work_group_size = CL_KERNEL_WORK_GROUP_SIZE
2567 preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE
2568 PPCODE:
2569 size_t value [1];
2570 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2571 EXTEND (SP, 1);
2572 const int i = 0;
2573 PUSHs (sv_2mortal (newSVuv (value [i])));
805 2574
806void 2575void
807set_ushort (OpenCL::Kernel this, cl_uint idx, cl_ushort value) 2576compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device)
808 CODE: 2577 PPCODE:
809 clSetKernelArg (this, idx, sizeof (value), &value); 2578 size_t size;
2579 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size));
2580 size_t *value = tmpbuf (size);
2581 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0));
2582 int i, n = size / sizeof (*value);
2583 EXTEND (SP, n);
2584 for (i = 0; i < n; ++i)
2585 PUSHs (sv_2mortal (newSVuv (value [i])));
810 2586
811void 2587void
812set_int (OpenCL::Kernel this, cl_uint idx, cl_int value) 2588local_mem_size (OpenCL::Kernel self, OpenCL::Device device)
813 CODE: 2589 ALIAS:
814 clSetKernelArg (this, idx, sizeof (value), &value); 2590 local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE
2591 private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE
2592 PPCODE:
2593 cl_ulong value [1];
2594 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2595 EXTEND (SP, 1);
2596 const int i = 0;
2597 PUSHs (sv_2mortal (newSVuv (value [i])));
815 2598
816void 2599#END:kernel_work_group
817set_uint (OpenCL::Kernel this, cl_uint idx, cl_uint value)
818 CODE:
819 clSetKernelArg (this, idx, sizeof (value), &value);
820
821void
822set_long (OpenCL::Kernel this, cl_uint idx, cl_long value)
823 CODE:
824 clSetKernelArg (this, idx, sizeof (value), &value);
825
826void
827set_ulong (OpenCL::Kernel this, cl_uint idx, cl_ulong value)
828 CODE:
829 clSetKernelArg (this, idx, sizeof (value), &value);
830
831void
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 2600
876MODULE = OpenCL PACKAGE = OpenCL::Event 2601MODULE = OpenCL PACKAGE = OpenCL::Event
877 2602
878void 2603void
879DESTROY (OpenCL::Event this) 2604DESTROY (OpenCL::Event self)
880 CODE: 2605 CODE:
881 clReleaseEvent (this); 2606 clReleaseEvent (self);
882 2607
883void 2608void
2609wait (OpenCL::Event self)
2610 CODE:
2611 clWaitForEvents (1, &self);
2612
2613void
2614cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb)
2615 CODE:
2616 clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb)));
2617
2618void
884info (OpenCL::Event this, cl_event_info name) 2619info (OpenCL::Event self, cl_event_info name)
885 PPCODE: 2620 PPCODE:
886 INFO (Event) 2621 INFO (Event)
887 2622
2623#BEGIN:event
2624
888void 2625void
889wait (OpenCL::Event this) 2626command_queue (OpenCL::Event self)
2627 PPCODE:
2628 cl_command_queue value [1];
2629 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0));
2630 EXTEND (SP, 1);
2631 const int i = 0;
2632 {
2633 NEED_SUCCESS (RetainCommandQueue, (value [i]));
2634 PUSH_CLOBJ (stash_queue, value [i]);
2635 }
2636
2637void
2638command_type (OpenCL::Event self)
2639 PPCODE:
2640 cl_command_type value [1];
2641 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_TYPE, sizeof (value), value, 0));
2642 EXTEND (SP, 1);
2643 const int i = 0;
2644 PUSHs (sv_2mortal (newSVuv (value [i])));
2645
2646void
2647reference_count (OpenCL::Event self)
2648 ALIAS:
2649 reference_count = CL_EVENT_REFERENCE_COUNT
2650 command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS
2651 PPCODE:
2652 cl_uint value [1];
2653 NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0));
2654 EXTEND (SP, 1);
2655 const int i = 0;
2656 PUSHs (sv_2mortal (newSVuv (value [i])));
2657
2658void
2659context (OpenCL::Event self)
2660 PPCODE:
2661 cl_context value [1];
2662 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0));
2663 EXTEND (SP, 1);
2664 const int i = 0;
2665 {
2666 NEED_SUCCESS (RetainContext, (value [i]));
2667 PUSH_CLOBJ (stash_context, value [i]);
2668 }
2669
2670#END:event
2671
2672void
2673profiling_info (OpenCL::Event self, cl_profiling_info name)
890 CODE: 2674 PPCODE:
891 clWaitForEvents (1, &this); 2675 INFO (EventProfiling)
2676
2677#BEGIN:profiling
2678
2679void
2680profiling_command_queued (OpenCL::Event self)
2681 ALIAS:
2682 profiling_command_queued = CL_PROFILING_COMMAND_QUEUED
2683 profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT
2684 profiling_command_start = CL_PROFILING_COMMAND_START
2685 profiling_command_end = CL_PROFILING_COMMAND_END
2686 PPCODE:
2687 cl_ulong value [1];
2688 NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0));
2689 EXTEND (SP, 1);
2690 const int i = 0;
2691 PUSHs (sv_2mortal (newSVuv (value [i])));
2692
2693#END:profiling
892 2694
893MODULE = OpenCL PACKAGE = OpenCL::UserEvent 2695MODULE = OpenCL PACKAGE = OpenCL::UserEvent
894 2696
895void 2697void
896set_status (OpenCL::UserEvent this, cl_int execution_status) 2698set_status (OpenCL::UserEvent self, cl_int execution_status)
897 CODE: 2699 CODE:
898 clSetUserEventStatus (this, execution_status); 2700 clSetUserEventStatus (self, execution_status);
899 2701
2702MODULE = OpenCL PACKAGE = OpenCL::Mapped
2703
2704void
2705DESTROY (SV *self)
2706 CODE:
2707 OpenCL__Mapped mapped = SvMAPPED (self);
2708
2709 clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0);
2710 mapped_detach (self, mapped);
2711
2712 clReleaseCommandQueue (mapped->queue);
2713
2714 if (mapped->event)
2715 clReleaseEvent (mapped->event);
2716
2717 Safefree (mapped);
2718
2719bool
2720mapped (OpenCL::Mapped self)
2721 CODE:
2722 RETVAL = !!self->ptr;
2723 OUTPUT:
2724 RETVAL
2725
2726void
2727wait (OpenCL::Mapped self)
2728 PPCODE:
2729 if (self->event)
2730 NEED_SUCCESS (WaitForEvents, (1, &self->event));
2731
2732void
2733event (OpenCL::Mapped self)
2734 PPCODE:
2735 if (!self->event)
2736 XSRETURN_UNDEF;
2737
2738 clRetainEvent (self->event);
2739 XPUSH_CLOBJ (stash_event, self->event);
2740
2741size_t
2742size (OpenCL::Mapped self)
2743 CODE:
2744 RETVAL = self->cb;
2745 OUTPUT:
2746 RETVAL
2747
2748IV
2749ptr (OpenCL::Mapped self)
2750 CODE:
2751 RETVAL = PTR2IV (self->ptr);
2752 OUTPUT:
2753 RETVAL
2754
2755MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer
2756
2757MODULE = OpenCL PACKAGE = OpenCL::MappedImage
2758
2759IV
2760row_pitch (OpenCL::Mapped self)
2761 ALIAS:
2762 slice_pitch = 1
2763 CODE:
2764 RETVAL = ix ? self->slice_pitch : self->row_pitch;
2765 OUTPUT:
2766 RETVAL
2767
2768
2769

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines