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 1.80 by root, Sat May 5 16:44:02 2012 UTC

1#include <stddef.h>
2
1#include "EXTERN.h" 3#include "EXTERN.h"
2#include "perl.h" 4#include "perl.h"
3#include "XSUB.h" 5#include "XSUB.h"
4 6
7#include "ecb.h"//D
8
9#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
10#include "xthread.h"
11#include "schmorp.h"
12
13#ifdef I_DLFCN
14 #include <dlfcn.h>
15#endif
16
17// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
18// but then fail to define the api functions unless you ALSO define
19// this. This breaks 100% of the opencl 1.1 apps, for what reason?
20// after all, the functions are deprecated, not removed.
21// in addition, you cannot test for this in any future-proof way.
22// each time a new opencl version comes out, you need to make a new
23// release.
24#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
25
26#ifndef PREFER_1_1
27 #define PREFER_1_1 1
28#endif
29
30#if PREFER_1_1
31 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
32#endif
33
34#ifdef __APPLE__
35 #define CLHDR(name) <OpenCL/name>
36#else
37 #define CLHDR(name) <CL/name>
38#endif
39
5#include <CL/opencl.h> 40#include CLHDR(opencl.h)
6 41
42#ifndef CL_VERSION_1_2
43 #include CLHDR(cl_d3d9.h)
44#endif
45
46#if _WIN32
47 #include CLHDR(cl_d3d10.h)
48 #if CL_VERSION_1_2
49 #include CLHDR<cl_d3d11.h>
50 #endif
51 #include CLHDR<cl_dx9_media_sharing.h.h>
52#endif
53
54#ifndef CL_VERSION_1_2
55 #undef PREFER_1_1
56 #define PREFER_1_1 1
57#endif
58
59// make sure all constants we might use are actually defined
60#include "default.h"
61
7typedef cl_platform_id OpenCL__Platform; 62typedef cl_platform_id OpenCL__Platform;
8typedef cl_device_id OpenCL__Device; 63typedef cl_device_id OpenCL__Device;
64typedef cl_device_id OpenCL__SubDevice;
9typedef cl_context OpenCL__Context; 65typedef cl_context OpenCL__Context;
10typedef cl_command_queue OpenCL__Queue; 66typedef cl_command_queue OpenCL__Queue;
11typedef cl_mem OpenCL__Memory; 67typedef cl_mem OpenCL__Memory;
12typedef cl_mem OpenCL__Buffer; 68typedef cl_mem OpenCL__Buffer;
69typedef cl_mem OpenCL__BufferObj;
13typedef cl_mem OpenCL__Image; 70typedef cl_mem OpenCL__Image;
14typedef cl_mem OpenCL__Image2D;
15typedef cl_mem OpenCL__Image3D;
16typedef cl_mem OpenCL__Memory_ornull; 71typedef cl_mem OpenCL__Memory_ornull;
17typedef cl_mem OpenCL__Buffer_ornull; 72typedef cl_mem OpenCL__Buffer_ornull;
18typedef cl_mem OpenCL__Image_ornull; 73typedef cl_mem OpenCL__Image_ornull;
19typedef cl_mem OpenCL__Image2D_ornull;
20typedef cl_mem OpenCL__Image3D_ornull;
21typedef cl_sampler OpenCL__Sampler; 74typedef cl_sampler OpenCL__Sampler;
22typedef cl_program OpenCL__Program; 75typedef cl_program OpenCL__Program;
23typedef cl_kernel OpenCL__Kernel; 76typedef cl_kernel OpenCL__Kernel;
24typedef cl_event OpenCL__Event; 77typedef cl_event OpenCL__Event;
25typedef cl_event OpenCL__UserEvent; 78typedef cl_event OpenCL__UserEvent;
79
80typedef struct mapped * OpenCL__Mapped;
81
82static HV
83 *stash_platform,
84 *stash_device,
85 *stash_subdevice,
86 *stash_context,
87 *stash_queue,
88 *stash_program,
89 *stash_kernel,
90 *stash_sampler,
91 *stash_event,
92 *stash_userevent,
93 *stash_memory,
94 *stash_buffer,
95 *stash_bufferobj,
96 *stash_image,
97 *stash_image1d,
98 *stash_image1darray,
99 *stash_image1dbuffer,
100 *stash_image2d,
101 *stash_image2darray,
102 *stash_image3d,
103 *stash_mapped,
104 *stash_mappedbuffer,
105 *stash_mappedimage;
26 106
27/*****************************************************************************/ 107/*****************************************************************************/
28 108
109// name must include a leading underscore
110// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
111// for perl. doh.
112static void *
113glsym (const char *name)
114{
115 void *fun = 0;
116
117 #if defined I_DLFCN && defined RTLD_DEFAULT
118 fun = dlsym (RTLD_DEFAULT, name + 1);
119 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
120
121 if (!fun)
122 {
123 static void *libgl;
124 static const char *glso[] = {
125 "libGL.so.1",
126 "libGL.so.3",
127 "libGL.so.4.0",
128 "libGL.so",
129 "/usr/lib/libGL.so",
130 "/usr/X11R6/lib/libGL.1.dylib"
131 };
132 int i;
133
134 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
135 {
136 libgl = dlopen (glso [i], RTLD_LAZY);
137 if (libgl)
138 break;
139 }
140
141 if (libgl)
142 {
143 fun = dlsym (libgl, name + 1);
144 if (!fun) fun = dlsym (libgl, name);
145 }
146 }
147 #endif
148
149 return fun;
150}
151
152/*****************************************************************************/
153
29/* up to two temporary buffers */ 154/* up to two temporary buffers */
30static void * 155static void * ecb_noinline
31tmpbuf (size_t size) 156tmpbuf (size_t size)
32{ 157{
158 enum { buffers = 4 };
33 static int idx; 159 static int idx;
34 static void *buf [2]; 160 static void *buf [buffers];
35 static size_t len [2]; 161 static size_t len [buffers];
36 162
37 idx ^= 1; 163 idx = (idx + 1) % buffers;
38 164
39 if (len [idx] < size) 165 if (len [idx] < size)
40 { 166 {
41 free (buf [idx]); 167 free (buf [idx]);
42 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 168 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
44 } 170 }
45 171
46 return buf [idx]; 172 return buf [idx];
47} 173}
48 174
175static const char * ecb_noinline
176cv_get_name (CV *cv)
177{
178 static char fullname [256];
179
180 GV *gv = CvGV (cv); // gv better be nonzero
181
182 HV *stash = GvSTASH (gv);
183 const char *hvname = HvNAME_get (stash); // stash also better be nonzero
184 const char *gvname = GvNAME (gv);
185
186 snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname);
187 return fullname;
188}
189
49/*****************************************************************************/ 190/*****************************************************************************/
50 191
51typedef struct 192typedef struct
52{ 193{
53 IV iv; 194 IV iv;
54 const char *name; 195 const char *name;
55 #define const_iv(name) { (IV)CL_ ## name, # name }, 196 #define const_iv(name) { (IV)CL_ ## name, # name },
56} ivstr; 197} ivstr;
57 198
199typedef struct
200{
201 NV nv;
202 const char *name;
203 #define const_nv(name) { (NV)CL_ ## name, # name },
204} nvstr;
205
58static const char * 206static const char *
59iv2str (IV value, const ivstr *base, int count, const char *fallback) 207iv2str (IV value, const ivstr *base, int count, const char *fallback)
60{ 208{
61 int i; 209 int i;
62 static char strbuf [32]; 210 static char strbuf [32];
90 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)"); 238 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)");
91} 239}
92 240
93/*****************************************************************************/ 241/*****************************************************************************/
94 242
95static cl_int last_error; 243static cl_int res;
96 244
97#define FAIL(name,err) \ 245#define FAIL(name) \
98 croak ("cl" # name ": %s", err2str (last_error = err)); 246 croak ("cl" # name ": %s", err2str (res));
99 247
100#define NEED_SUCCESS(name,args) \ 248#define NEED_SUCCESS(name,args) \
101 do { \ 249 do { \
102 cl_int res = cl ## name args; \ 250 res = cl ## name args; \
103 \ 251 \
104 if (res) \ 252 if (res) \
105 FAIL (name, res); \ 253 FAIL (name); \
106 } while (0) 254 } while (0)
107 255
256#define NEED_SUCCESS_ARG(retdecl, name, args) \
257 retdecl = cl ## name args; \
258 if (res) \
259 FAIL (name);
260
108/*****************************************************************************/ 261/*****************************************************************************/
109 262
110#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr)
111#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr))
112
113static void * 263static SV *
264new_clobj (HV *stash, IV id)
265{
266 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash));
267}
268
269#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id)))
270#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
271
272/* cl objects are either \$iv, or [$iv, ...] */
273/* they can be upgraded at runtime to the array form */
274static void * ecb_noinline
114SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) 275SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg)
115{ 276{
277 // sv_derived_from is quite slow :(
116 if (SvROK (sv) && sv_derived_from (sv, pkg)) 278 if (SvROK (sv) && sv_derived_from (sv, pkg))
117 return (void *)SvIV (SvRV (sv)); 279 return (void *)SvIV (SvRV (sv));
118 280
119 croak ("%s: %s is not of type %s", func, svname, pkg); 281 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg);
282}
283
284// the "no-inherit" version of the above
285static void * ecb_noinline
286SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash)
287{
288 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash)
289 return (void *)SvIV (SvRV (sv));
290
291 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash));
120} 292}
121 293
122/*****************************************************************************/ 294/*****************************************************************************/
123 295
124static cl_event * 296static cl_context_properties * ecb_noinline
125event_list (SV **items, int count) 297SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
126{ 298{
299 if (!sv || !SvOK (sv))
300 if (extra)
301 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
302 else
303 return 0;
304
305 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
306 {
307 AV *av = (AV *)SvRV (sv);
308 int i, len = av_len (av) + 1;
309 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
310 cl_context_properties *l = p;
311
312 if (len & 1)
313 croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname);
314
315 while (extracount--)
316 *l++ = *extra++;
317
318 for (i = 0; i < len; i += 2)
319 {
320 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
321 SV *p_sv = *av_fetch (av, i + 1, 0);
322 cl_context_properties v = SvIV (p_sv); // code below can override
323
324 switch (t)
325 {
326 case CL_CONTEXT_PLATFORM:
327 if (SvROK (p_sv))
328 v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform");
329 break;
330
331 case CL_GLX_DISPLAY_KHR:
332 if (!SvOK (p_sv))
333 {
334 void *func = glsym ("_glXGetCurrentDisplay");
335 if (func)
336 v = (cl_context_properties)((void *(*)(void))func)();
337 }
338 break;
339
340 case CL_GL_CONTEXT_KHR:
341 if (!SvOK (p_sv))
342 {
343 void *func = glsym ("_glXGetCurrentContext");
344 if (func)
345 v = (cl_context_properties)((void *(*)(void))func)();
346 }
347 break;
348
349 default:
350 /* unknown property, treat as int */
351 break;
352 }
353
354 *l++ = t;
355 *l++ = v;
356 }
357
358 *l = 0;
359
360 return p;
361 }
362
363 croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname);
364}
365
366// parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation
367// is a pointer (and only on well-behaved systems).
368static void * ecb_noinline
369object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount)
370{
371 if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV)
372 croak ("%s: '%s' parameter must be %sa reference to an array of %s objects",
373 cv_get_name (cv), argname, or_undef ? "undef or " : "", klass);
374
375 AV *av = (AV *)SvRV (arg);
376 void **list = 0;
377 cl_uint count = av_len (av) + 1;
378
379 if (count)
380 {
381 list = tmpbuf (sizeof (*list) * count);
382 int i;
383 for (i = 0; i < count; ++i)
384 list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass);
385 }
386
387 if (!count && !or_undef)
388 croak ("%s: '%s' must contain at least one %s object",
389 cv_get_name (cv), argname, klass);
390
391 *rcount = count;
392 return (void *)list;
393}
394
395/*****************************************************************************/
396/* callback stuff */
397
398/* default context callback, log to stderr */
399static void CL_CALLBACK
400context_default_notify (const char *msg, const void *info, size_t cb, void *data)
401{
402 fprintf (stderr, "OpenCL Context Notify: %s\n", msg);
403}
404
405typedef struct
406{
407 int free_cb;
408 void (*push)(void *data1, void *data2, void *data3);
409} eq_vtbl;
410
411typedef struct eq_item
412{
413 struct eq_item *next;
414 eq_vtbl *vtbl;
415 SV *cb;
416 void *data1, *data2, *data3;
417} eq_item;
418
419static void (*eq_signal_func)(void *signal_arg, int value);
420static void *eq_signal_arg;
421static xmutex_t eq_lock = X_MUTEX_INIT;
422static eq_item *eq_head, *eq_tail;
423
424static void ecb_noinline
425eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3)
426{
427 eq_item *item = malloc (sizeof (eq_item));
428
429 item->next = 0;
430 item->vtbl = vtbl;
431 item->cb = cb;
432 item->data1 = data1;
433 item->data2 = data2;
434 item->data3 = data3;
435
436 X_LOCK (eq_lock);
437
438 *(eq_head ? &eq_tail->next : &eq_head) = item;
439 eq_tail = item;
440
441 X_UNLOCK (eq_lock);
442
443 eq_signal_func (eq_signal_arg, 0);
444}
445
446static eq_item *
447eq_dec (void)
448{
449 eq_item *res;
450
451 X_LOCK (eq_lock);
452
453 res = eq_head;
454 if (res)
455 eq_head = res->next;
456
457 X_UNLOCK (eq_lock);
458
459 return res;
460}
461
462static void
463eq_poll (void)
464{
465 eq_item *item;
466
467 while ((item = eq_dec ()))
468 {
469 ENTER;
470 SAVETMPS;
471
472 dSP;
473 PUSHMARK (SP);
474 EXTEND (SP, 2);
475
476 if (item->vtbl->free_cb)
477 sv_2mortal (item->cb);
478
479 PUTBACK;
480 item->vtbl->push (item->data1, item->data2, item->data3);
481
482 SV *cb = item->cb;
483 free (item);
484
485 call_sv (cb, G_DISCARD | G_VOID);
486
487 FREETMPS;
488 LEAVE;
489 }
490}
491
492static void
493eq_poll_interrupt (pTHX_ void *c_arg, int value)
494{
495 eq_poll ();
496}
497
498/*****************************************************************************/
499/* context notify */
500
501static void ecb_noinline
502eq_context_push (void *data1, void *data2, void *data3)
503{
504 dSP;
505 PUSHs (sv_2mortal (newSVpv (data1, 0)));
506 PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3)));
507 PUTBACK;
508
509 free (data1);
510 free (data2);
511}
512
513static eq_vtbl eq_context_vtbl = { 0, eq_context_push };
514
515static void CL_CALLBACK
516eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data)
517{
518 void *pvt_copy = malloc (cb);
519 memcpy (pvt_copy, pvt, cb);
520 eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb);
521}
522
523#define CONTEXT_NOTIFY_CALLBACK \
524 void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \
525 void *user_data = 0; \
526 \
527 if (SvOK (notify)) \
528 { \
529 pfn_notify = eq_context_notify; \
530 user_data = s_get_cv (notify); \
531 }
532
533static SV * ecb_noinline
534new_clobj_context (cl_context ctx, void *user_data)
535{
536 SV *sv = new_clobj (stash_context, (IV)ctx);
537
538 if (user_data)
539 sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0);
540
541 return sv;
542}
543
544#define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data));
545
546/*****************************************************************************/
547/* build/compile/link notify */
548
549static void
550eq_program_push (void *data1, void *data2, void *data3)
551{
552 dSP;
553 PUSH_CLOBJ (stash_program, data1);
554 PUTBACK;
555}
556
557static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
558
559static void CL_CALLBACK
560eq_program_notify (cl_program program, void *user_data)
561{
562 clRetainProgram (program);
563
564 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
565}
566
567typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data);
568
569static program_callback ecb_noinline
570make_program_callback (SV *notify, void **ruser_data)
571{
572 if (SvOK (notify))
573 {
574 *ruser_data = SvREFCNT_inc (s_get_cv (notify));
575 return eq_program_notify;
576 }
577 else
578 {
579 *ruser_data = 0;
580 return 0;
581 }
582}
583
584struct build_args
585{
586 cl_program program;
587 char *options;
588 void *user_data;
589 cl_uint num_devices;
590};
591
592X_THREAD_PROC (build_program_thread)
593{
594 struct build_args *arg = thr_arg;
595
596 clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0);
597
598 if (arg->user_data)
599 eq_program_notify (arg->program, arg->user_data);
600 else
601 clReleaseProgram (arg->program);
602
603 free (arg->options);
604 free (arg);
605
606 return 0;
607}
608
609static void
610build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data)
611{
612 struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices);
613
614 arg->program = program;
615 arg->options = strdup (options);
616 arg->user_data = user_data;
617 arg->num_devices = num_devices;
618 memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices);
619
620 xthread_t id;
621 thread_create (&id, build_program_thread, arg);
622}
623
624/*****************************************************************************/
625/* mem object destructor notify */
626
627static void ecb_noinline
628eq_destructor_push (void *data1, void *data2, void *data3)
629{
630}
631
632static eq_vtbl eq_destructor_vtbl = { 0, eq_destructor_push };
633
634static void CL_CALLBACK
635eq_destructor_notify (cl_mem memobj, void *user_data)
636{
637 eq_enq (&eq_destructor_vtbl, (SV *)user_data, (void *)memobj, 0, 0);
638}
639
640/*****************************************************************************/
641/* event objects */
642
643static void
644eq_event_push (void *data1, void *data2, void *data3)
645{
646 dSP;
647 PUSH_CLOBJ (stash_event, data1);
648 PUSHs (sv_2mortal (newSViv ((IV)data2)));
649 PUTBACK;
650}
651
652static eq_vtbl eq_event_vtbl = { 1, eq_event_push };
653
654static void CL_CALLBACK
655eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data)
656{
657 clRetainEvent (event);
658 eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0);
659}
660
661/*****************************************************************************/
662/* utilities for XS code */
663
664static size_t
665img_row_pitch (cl_mem img)
666{
667 size_t res;
668 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
669 return res;
670}
671
672static cl_event * ecb_noinline
673event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
674{
675 cl_uint count = *rcount;
676
677 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed
678 *rcount = count = 0;
679
680 if (!count && !extra)
681 return 0;
682
127 cl_event *list = tmpbuf (sizeof (cl_event) * count); 683 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
684 int i = 0;
128 685
129 while (count--) 686 while (count--)
687 if (SvOK (items [count]))
130 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 688 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
131 689
690 if (extra)
691 list [i++] = extra;
692
693 *rcount = i;
694
132 return list; 695 return i ? list : 0;
133} 696}
134 697
135#define EVENT_LIST(items,count) \ 698#define EVENT_LIST(skip) \
136 cl_uint event_list_count = (count); \ 699 cl_uint event_list_count = items - (skip); \
137 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 700 cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0)
138 701
139#define INFO(class) \ 702#define INFO(class) \
140{ \ 703{ \
141 size_t size; \ 704 size_t size; \
142 SV *sv; \
143 \
144 NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ 705 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
145 sv = sv_2mortal (newSV (size)); \ 706 SV *sv = sv_2mortal (newSV (size)); \
146 SvUPGRADE (sv, SVt_PV); \ 707 SvUPGRADE (sv, SVt_PV); \
147 SvPOK_only (sv); \ 708 SvPOK_only (sv); \
148 SvCUR_set (sv, size); \ 709 SvCUR_set (sv, size); \
149 NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ 710 NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \
150 XPUSHs (sv); \ 711 XPUSHs (sv); \
151} 712}
152 713
714/*****************************************************************************/
715/* mapped_xxx */
716
717static OpenCL__Mapped
718SvMAPPED (SV *sv)
719{
720 // no typechecking atm., keep your fingers crossed
721 return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr;
722}
723
724struct mapped
725{
726 cl_command_queue queue;
727 cl_mem memobj;
728 void *ptr;
729 size_t cb;
730 cl_event event;
731 size_t row_pitch;
732 size_t slice_pitch;
733
734 size_t element_size;
735 size_t width, height, depth;
736};
737
738static SV *
739mapped_new (
740 HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags,
741 void *ptr, size_t cb, cl_event ev,
742 size_t row_pitch, size_t slice_pitch, size_t element_size,
743 size_t width, size_t height, size_t depth
744)
745{
746 SV *data = newSV (0);
747 SvUPGRADE (data, SVt_PVMG);
748
749 OpenCL__Mapped mapped;
750 New (0, mapped, 1, struct mapped);
751
752 clRetainCommandQueue (queue);
753
754 mapped->queue = queue;
755 mapped->memobj = memobj;
756 mapped->ptr = ptr;
757 mapped->cb = cb;
758 mapped->event = ev;
759 mapped->row_pitch = row_pitch;
760 mapped->slice_pitch = slice_pitch;
761
762 mapped->element_size = element_size;
763 mapped->width = width;
764 mapped->height = height;
765 mapped->depth = depth;
766
767 sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0);
768
769 if (SvLEN (data))
770 Safefree (data);
771
772 SvPVX (data) = (char *)ptr;
773 SvCUR_set (data, cb);
774 SvLEN_set (data, 0);
775 SvPOK_only (data);
776
777 SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash));
778
779 if (!(flags & CL_MAP_WRITE))
780 SvREADONLY_on (data);
781
782 return obj;
783}
784
785static void
786mapped_detach (SV *sv, OpenCL__Mapped mapped)
787{
788 SV *data = SvRV (sv);
789
790 // the next check checks both before AND after detach, where SvPVX should STILL be 0
791 if (SvPVX (data) != (char *)mapped->ptr)
792 warn ("FATAL: OpenCL memory mapped scalar changed location, detected");
793 else
794 {
795 SvREADONLY_off (data);
796 SvCUR_set (data, 0);
797 SvPVX (data) = 0;
798 SvOK_off (data);
799 }
800
801 mapped->ptr = 0;
802}
803
804static void
805mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count)
806{
807 cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event);
808 cl_event ev;
809
810 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev));
811
812 clReleaseEvent (mapped->event);
813 mapped->event = ev;
814
815 mapped_detach (self, mapped);
816}
817
818static size_t
819mapped_element_size (OpenCL__Mapped self)
820{
821 if (!self->element_size)
822 clGetImageInfo (self->memobj, CL_IMAGE_ELEMENT_SIZE, sizeof (self->element_size), &self->element_size, 0);
823
824 return self->element_size;
825}
826
827/*****************************************************************************/
828
153MODULE = OpenCL PACKAGE = OpenCL 829MODULE = OpenCL PACKAGE = OpenCL
154 830
155PROTOTYPES: ENABLE 831PROTOTYPES: ENABLE
156 832
833void
834poll ()
835 CODE:
836 eq_poll ();
837
838void
839_eq_initialise (IV func, IV arg)
840 CODE:
841 eq_signal_func = (void (*)(void *, int))func;
842 eq_signal_arg = (void*)arg;
843
157BOOT: 844BOOT:
158{ 845{
159 HV *stash = gv_stashpv ("OpenCL", 1); 846 HV *stash = gv_stashpv ("OpenCL", 1);
847
160 static const ivstr *civ, const_iv[] = { 848 static const ivstr *civ, const_iv[] = {
161 { sizeof (cl_char ), "SIZEOF_CHAR" }, 849 { sizeof (cl_char ), "SIZEOF_CHAR" },
162 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 850 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
163 { sizeof (cl_short ), "SIZEOF_SHORT" }, 851 { sizeof (cl_short ), "SIZEOF_SHORT" },
164 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 852 { sizeof (cl_ushort), "SIZEOF_USHORT" },
165 { sizeof (cl_int ), "SIZEOF_INT" }, 853 { sizeof (cl_int ), "SIZEOF_INT" },
166 { sizeof (cl_uint ), "SIZEOF_UINT" }, 854 { sizeof (cl_uint ), "SIZEOF_UINT" },
167 { sizeof (cl_long ), "SIZEOF_LONG" }, 855 { sizeof (cl_long ), "SIZEOF_LONG" },
168 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 856 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
169 { sizeof (cl_half ), "SIZEOF_HALF" }, 857 { sizeof (cl_half ), "SIZEOF_HALF" },
170 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 858 { sizeof (cl_float ), "SIZEOF_FLOAT" },
171 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 859 { sizeof (cl_double), "SIZEOF_DOUBLE" },
860 { PREFER_1_1 , "PREFER_1_1" },
172#include "constiv.h" 861#include "constiv.h"
173 }; 862 };
863
174 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 864 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)); 865 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
866
867 static const nvstr *cnv, const_nv[] = {
868#include "constnv.h"
869 };
870
871 for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--)
872 newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv));
873
874 newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call
875
876 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
877 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD);
878 stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD);
879 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
880 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD);
881 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD);
882 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD);
883 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD);
884 stash_event = gv_stashpv ("OpenCL::Event", GV_ADD);
885 stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD);
886 stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD);
887 stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD);
888 stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD);
889 stash_image = gv_stashpv ("OpenCL::Image", GV_ADD);
890 stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD);
891 stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD);
892 stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD);
893 stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD);
894 stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD);
895 stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD);
896 stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD);
897 stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD);
898 stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD);
899
900 sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt);
176} 901}
177 902
178cl_int 903cl_int
179errno () 904errno ()
180 CODE: 905 CODE:
181 errno = last_error; 906 RETVAL = res;
907 OUTPUT:
908 RETVAL
182 909
183const char * 910const char *
184err2str (cl_int err) 911err2str (cl_int err = res)
185 912
186const char * 913const char *
187enum2str (cl_uint value) 914enum2str (cl_uint value)
188 915
189void 916void
190platforms () 917platforms ()
191 PPCODE: 918 PPCODE:
192{
193 cl_platform_id *list; 919 cl_platform_id *list;
194 cl_uint count; 920 cl_uint count;
195 int i; 921 int i;
196 922
197 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); 923 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count));
198 list = tmpbuf (sizeof (*list) * count); 924 list = tmpbuf (sizeof (*list) * count);
199 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 925 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
200 926
201 EXTEND (SP, count); 927 EXTEND (SP, count);
202 for (i = 0; i < count; ++i) 928 for (i = 0; i < count; ++i)
203 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 929 PUSH_CLOBJ (stash_platform, list [i]);
204}
205 930
206void 931void
207context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 932context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
208 PPCODE: 933 PPCODE:
209{ 934 CONTEXT_NOTIFY_CALLBACK;
210 cl_int res; 935 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
211 cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); 936 XPUSH_CLOBJ_CONTEXT;
212 937
213 if (res) 938void
214 FAIL (CreateContextFromType, res); 939context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
940 PPCODE:
941 cl_uint device_count;
942 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
215 943
216 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 944 CONTEXT_NOTIFY_CALLBACK;
217} 945 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res));
946 XPUSH_CLOBJ_CONTEXT;
218 947
219void 948void
220wait_for_events (...) 949wait_for_events (...)
221 CODE: 950 CODE:
222{
223 EVENT_LIST (0, items); 951 EVENT_LIST (0);
224 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); 952 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr));
225}
226 953
227PROTOTYPES: DISABLE 954PROTOTYPES: DISABLE
228 955
229MODULE = OpenCL PACKAGE = OpenCL::Platform 956MODULE = OpenCL PACKAGE = OpenCL::Platform
230 957
231void 958void
232info (OpenCL::Platform this, cl_platform_info name) 959info (OpenCL::Platform self, cl_platform_info name)
233 PPCODE: 960 PPCODE:
234 INFO (Platform) 961 INFO (Platform)
235 962
236void 963void
964unload_compiler (OpenCL::Platform self)
965 CODE:
966#if CL_VERSION_1_2
967 clUnloadPlatformCompiler (self);
968#endif
969
970#BEGIN:platform
971
972void
973profile (OpenCL::Platform self)
974 ALIAS:
975 profile = CL_PLATFORM_PROFILE
976 version = CL_PLATFORM_VERSION
977 name = CL_PLATFORM_NAME
978 vendor = CL_PLATFORM_VENDOR
979 extensions = CL_PLATFORM_EXTENSIONS
980 PPCODE:
981 size_t size;
982 NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size));
983 char *value = tmpbuf (size);
984 NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0));
985 EXTEND (SP, 1);
986 const int i = 0;
987 PUSHs (sv_2mortal (newSVpv (value, 0)));
988
989#END:platform
990
991void
237devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) 992devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL)
238 PPCODE: 993 PPCODE:
239{
240 cl_device_id *list; 994 cl_device_id *list;
241 cl_uint count; 995 cl_uint count;
242 int i; 996 int i;
243 997
244 NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); 998 NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count));
245 list = tmpbuf (sizeof (*list) * count); 999 list = tmpbuf (sizeof (*list) * count);
246 NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); 1000 NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0));
247 1001
248 EXTEND (SP, count); 1002 EXTEND (SP, count);
249 for (i = 0; i < count; ++i) 1003 for (i = 0; i < count; ++i)
250 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 1004 PUSH_CLOBJ (stash_device, list [i]);
251}
252 1005
253void 1006void
254context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 1007context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
255 PPCODE: 1008 PPCODE:
256{
257 cl_int res;
258 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; 1009 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
259 cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); 1010 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
260 1011
261 if (res) 1012 cl_uint device_count;
262 FAIL (CreateContextFromType, res); 1013 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
263 1014
264 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 1015 CONTEXT_NOTIFY_CALLBACK;
265} 1016 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
1017 XPUSH_CLOBJ_CONTEXT;
1018
1019void
1020context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
1021 PPCODE:
1022 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
1023 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
1024
1025 CONTEXT_NOTIFY_CALLBACK;
1026 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res));
1027 XPUSH_CLOBJ_CONTEXT;
266 1028
267MODULE = OpenCL PACKAGE = OpenCL::Device 1029MODULE = OpenCL PACKAGE = OpenCL::Device
268 1030
269void 1031void
270info (OpenCL::Device this, cl_device_info name) 1032info (OpenCL::Device self, cl_device_info name)
271 PPCODE: 1033 PPCODE:
272 INFO (Device) 1034 INFO (Device)
273 1035
1036#if CL_VERSION_1_2
1037
274void 1038void
275context_simple (OpenCL::Device this) 1039sub_devices (OpenCL::Device self, SV *properties)
1040 PPCODE:
1041 if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV)
1042 croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs");
1043
1044 properties = SvRV (properties);
1045
1046 cl_uint count = av_len ((AV *)properties) + 1;
1047 cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1);
1048
1049 int i;
1050 for (i = 0; i < count; ++i)
1051 props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0));
1052
1053 props [count] = 0;
1054
1055 cl_uint num_devices;
1056 NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices));
1057 cl_device_id *list = tmpbuf (sizeof (*list) * num_devices);
1058 NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0));
1059
1060 EXTEND (SP, num_devices);
1061 for (i = 0; i < count; ++i)
1062 PUSH_CLOBJ (stash_subdevice, list [i]);
1063
1064#endif
1065
1066#BEGIN:device
1067
1068void
1069type (OpenCL::Device self)
1070 ALIAS:
1071 type = CL_DEVICE_TYPE
1072 address_bits = CL_DEVICE_ADDRESS_BITS
1073 max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE
1074 single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG
1075 global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE
1076 global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE
1077 max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE
1078 local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE
1079 execution_capabilities = CL_DEVICE_EXECUTION_CAPABILITIES
1080 properties = CL_DEVICE_QUEUE_PROPERTIES
1081 double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG
1082 half_fp_config = CL_DEVICE_HALF_FP_CONFIG
276 PPCODE: 1083 PPCODE:
277{ 1084 cl_ulong value [1];
278 cl_int res; 1085 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
279 cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); 1086 EXTEND (SP, 1);
1087 const int i = 0;
1088 PUSHs (sv_2mortal (newSVuv (value [i])));
280 1089
281 if (res) 1090void
282 FAIL (CreateContext, res); 1091vendor_id (OpenCL::Device self)
1092 ALIAS:
1093 vendor_id = CL_DEVICE_VENDOR_ID
1094 max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS
1095 max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS
1096 preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR
1097 preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT
1098 preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT
1099 preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG
1100 preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT
1101 preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE
1102 max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY
1103 max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS
1104 max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS
1105 image_support = CL_DEVICE_IMAGE_SUPPORT
1106 max_samplers = CL_DEVICE_MAX_SAMPLERS
1107 mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN
1108 min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE
1109 global_mem_cache_type = CL_DEVICE_GLOBAL_MEM_CACHE_TYPE
1110 global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE
1111 max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS
1112 local_mem_type = CL_DEVICE_LOCAL_MEM_TYPE
1113 preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF
1114 native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR
1115 native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT
1116 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
1117 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
1118 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
1119 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
1120 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
1121 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
1122 PPCODE:
1123 cl_uint value [1];
1124 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1125 EXTEND (SP, 1);
1126 const int i = 0;
1127 PUSHs (sv_2mortal (newSVuv (value [i])));
283 1128
284 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 1129void
285} 1130max_work_group_size (OpenCL::Device self)
1131 ALIAS:
1132 max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE
1133 image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH
1134 image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT
1135 image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH
1136 image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT
1137 image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH
1138 max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE
1139 profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION
1140 PPCODE:
1141 size_t value [1];
1142 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1143 EXTEND (SP, 1);
1144 const int i = 0;
1145 PUSHs (sv_2mortal (newSVuv (value [i])));
1146
1147void
1148max_work_item_sizes (OpenCL::Device self)
1149 PPCODE:
1150 size_t size;
1151 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size));
1152 size_t *value = tmpbuf (size);
1153 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0));
1154 int i, n = size / sizeof (*value);
1155 EXTEND (SP, n);
1156 for (i = 0; i < n; ++i)
1157 PUSHs (sv_2mortal (newSVuv (value [i])));
1158
1159void
1160error_correction_support (OpenCL::Device self)
1161 ALIAS:
1162 error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT
1163 endian_little = CL_DEVICE_ENDIAN_LITTLE
1164 available = CL_DEVICE_AVAILABLE
1165 compiler_available = CL_DEVICE_COMPILER_AVAILABLE
1166 host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY
1167 PPCODE:
1168 cl_uint value [1];
1169 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1170 EXTEND (SP, 1);
1171 const int i = 0;
1172 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
1173
1174void
1175platform (OpenCL::Device self)
1176 PPCODE:
1177 cl_platform_id value [1];
1178 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0));
1179 EXTEND (SP, 1);
1180 const int i = 0;
1181 PUSH_CLOBJ (stash_platform, value [i]);
1182
1183void
1184name (OpenCL::Device self)
1185 ALIAS:
1186 name = CL_DEVICE_NAME
1187 vendor = CL_DEVICE_VENDOR
1188 driver_version = CL_DRIVER_VERSION
1189 profile = CL_DEVICE_PROFILE
1190 version = CL_DEVICE_VERSION
1191 extensions = CL_DEVICE_EXTENSIONS
1192 PPCODE:
1193 size_t size;
1194 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1195 char *value = tmpbuf (size);
1196 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1197 EXTEND (SP, 1);
1198 const int i = 0;
1199 PUSHs (sv_2mortal (newSVpv (value, 0)));
1200
1201void
1202parent_device_ext (OpenCL::Device self)
1203 PPCODE:
1204 cl_device_id value [1];
1205 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0));
1206 EXTEND (SP, 1);
1207 const int i = 0;
1208 PUSH_CLOBJ (stash_device, value [i]);
1209
1210void
1211partition_types_ext (OpenCL::Device self)
1212 ALIAS:
1213 partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT
1214 affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT
1215 partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT
1216 PPCODE:
1217 size_t size;
1218 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1219 cl_device_partition_property_ext *value = tmpbuf (size);
1220 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1221 int i, n = size / sizeof (*value);
1222 EXTEND (SP, n);
1223 for (i = 0; i < n; ++i)
1224 PUSHs (sv_2mortal (newSVuv (value [i])));
1225
1226#END:device
1227
1228MODULE = OpenCL PACKAGE = OpenCL::SubDevice
1229
1230#if CL_VERSION_1_2
1231
1232void
1233DESTROY (OpenCL::SubDevice self)
1234 CODE:
1235 clReleaseDevice (self);
1236
1237#endif
286 1238
287MODULE = OpenCL PACKAGE = OpenCL::Context 1239MODULE = OpenCL PACKAGE = OpenCL::Context
288 1240
289void 1241void
290DESTROY (OpenCL::Context context) 1242DESTROY (OpenCL::Context self)
291 CODE: 1243 CODE:
292 clReleaseContext (context); 1244 clReleaseContext (self);
293 1245
294void 1246void
295info (OpenCL::Context this, cl_context_info name) 1247info (OpenCL::Context self, cl_context_info name)
296 PPCODE: 1248 PPCODE:
297 INFO (Context) 1249 INFO (Context)
298 1250
299void 1251void
300command_queue_simple (OpenCL::Context this, OpenCL::Device device) 1252queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0)
301 PPCODE: 1253 PPCODE:
302{ 1254 NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res));
303 cl_int res; 1255 XPUSH_CLOBJ (stash_queue, queue);
304 cl_command_queue queue = clCreateCommandQueue (this, device, 0, &res);
305 1256
306 if (res)
307 FAIL (CreateCommandQueue, res);
308
309 XPUSH_NEW_OBJ ("OpenCL::Queue", queue);
310}
311
312void 1257void
313user_event (OpenCL::Context this) 1258user_event (OpenCL::Context self)
314 PPCODE: 1259 PPCODE:
315{ 1260 NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res));
316 cl_int res; 1261 XPUSH_CLOBJ (stash_userevent, ev);
317 cl_event ev = clCreateUserEvent (this, &res);
318 1262
319 if (res)
320 FAIL (CreateUserevent, res);
321
322 XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev);
323}
324
325void 1263void
326buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) 1264buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
327 PPCODE: 1265 PPCODE:
328{
329 cl_int res;
330 cl_mem mem;
331
332 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 1266 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?"); 1267 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
334 1268
335 mem = clCreateBuffer (this, flags, len, 0, &res); 1269 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
1270 XPUSH_CLOBJ (stash_bufferobj, mem);
336 1271
337 if (res)
338 FAIL (CreateBuffer, res);
339
340 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
341}
342
343void 1272void
344buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) 1273buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
345 PPCODE: 1274 PPCODE:
346{
347 STRLEN len; 1275 STRLEN len;
348 char *ptr = SvPVbyte (data, len); 1276 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))) 1277 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?"); 1278 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
354 1279 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
355 mem = clCreateBuffer (this, flags, len, ptr, &res); 1280 XPUSH_CLOBJ (stash_bufferobj, mem);
356 1281
357 if (res) 1282#if CL_VERSION_1_2
358 FAIL (CreateBuffer, res);
359 1283
360 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem);
361}
362
363void 1284void
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) 1285image (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: 1286 PPCODE:
366{
367 STRLEN len; 1287 STRLEN len;
368 char *ptr = SvPVbyte (data, len); 1288 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
369 const cl_image_format format = { channel_order, channel_type }; 1289 const cl_image_format format = { channel_order, channel_type };
370 cl_int res; 1290 const cl_image_desc desc = {
371 cl_mem mem = clCreateImage2D (this, flags, &format, width, height, len / height, ptr, &res); 1291 type,
1292 width, height, depth,
1293 array_size, row_pitch, slice_pitch,
1294 num_mip_level, num_samples,
1295 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0
1296 };
1297 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1298 HV *stash = stash_image;
1299 switch (type)
1300 {
1301 case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break;
1302 case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break;
1303 case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break;
1304 case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break;
1305 case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break;
1306 case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break;
1307 }
1308 XPUSH_CLOBJ (stash, mem);
372 1309
373 if (res) 1310#endif
374 FAIL (CreateImage2D, res);
375 1311
376 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem);
377}
378
379void 1312void
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) 1313image2d (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: 1314 PPCODE:
382{
383 STRLEN len; 1315 STRLEN len;
384 char *ptr = SvPVbyte (data, len); 1316 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
385 const cl_image_format format = { channel_order, channel_type }; 1317 const cl_image_format format = { channel_order, channel_type };
386 cl_int res; 1318#if PREFER_1_1
387 cl_mem mem = clCreateImage3D (this, flags, &format, width, height, 1319 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); 1320#else
1321 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
1322 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1323#endif
1324 XPUSH_CLOBJ (stash_image2d, mem);
389 1325
390 if (res)
391 FAIL (CreateImage3D, res);
392
393 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem);
394}
395
396void 1326void
1327image3d (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)
1328 PPCODE:
1329 STRLEN len;
1330 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1331 const cl_image_format format = { channel_order, channel_type };
1332#if PREFER_1_1
1333 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
1334#else
1335 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
1336 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1337#endif
1338 XPUSH_CLOBJ (stash_image3d, mem);
1339
1340#if cl_apple_gl_sharing || cl_khr_gl_sharing
1341
1342void
1343gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
1344 PPCODE:
1345 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
1346 XPUSH_CLOBJ (stash_bufferobj, mem);
1347
1348void
1349gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
1350 PPCODE:
1351 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
1352 XPUSH_CLOBJ (stash_image2d, mem);
1353
1354#if CL_VERSION_1_2
1355
1356void
1357gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1358 ALIAS:
1359 PPCODE:
1360 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
1361 cl_gl_object_type type;
1362 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
1363 HV *stash = stash_memory;
1364 switch (type)
1365 {
1366 case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break;
1367 case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break;
1368 case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break;
1369 case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break;
1370 case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break;
1371 case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break;
1372 }
1373 XPUSH_CLOBJ (stash, mem);
1374
1375#endif
1376
1377void
1378gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1379 PPCODE:
1380#if PREFER_1_1
1381 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
1382#else
1383 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1384#endif
1385 XPUSH_CLOBJ (stash_image2d, mem);
1386
1387void
1388gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1389 PPCODE:
1390#if PREFER_1_1
1391 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
1392#else
1393 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1394#endif
1395 XPUSH_CLOBJ (stash_image3d, mem);
1396
1397#endif
1398
1399void
397supported_image_formats (OpenCL::Context this, cl_mem_flags flags, cl_mem_object_type image_type) 1400supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
398 PPCODE: 1401 PPCODE:
399{ 1402{
400 cl_uint count; 1403 cl_uint count;
401 cl_image_format *list; 1404 cl_image_format *list;
402 int i; 1405 int i;
403 1406
404 NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, 0, 0, &count)); 1407 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count));
405 Newx (list, count, cl_image_format); 1408 Newx (list, count, cl_image_format);
406 NEED_SUCCESS (GetSupportedImageFormats, (this, flags, image_type, count, list, 0)); 1409 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0));
407 1410
408 EXTEND (SP, count); 1411 EXTEND (SP, count);
409 for (i = 0; i < count; ++i) 1412 for (i = 0; i < count; ++i)
410 { 1413 {
411 AV *av = newAV (); 1414 AV *av = newAV ();
414 PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); 1417 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
415 } 1418 }
416} 1419}
417 1420
418void 1421void
419sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) 1422sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode)
420 PPCODE: 1423 PPCODE:
421{
422 cl_int res;
423 cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); 1424 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); 1425 XPUSH_CLOBJ (stash_sampler, sampler);
429}
430 1426
431void 1427void
432program_with_source (OpenCL::Context this, SV *program) 1428program_with_source (OpenCL::Context self, SV *program)
433 PPCODE: 1429 PPCODE:
434{
435 STRLEN len; 1430 STRLEN len;
436 size_t len2; 1431 size_t len2;
437 const char *ptr = SvPVbyte (program, len); 1432 const char *ptr = SvPVbyte (program, len);
438 cl_int res;
439 cl_program prog;
440 1433
441 len2 = len; 1434 len2 = len;
442 prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); 1435 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); 1436 XPUSH_CLOBJ (stash_program, prog);
448} 1437
1438void
1439program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1440 PPCODE:
1441 cl_uint device_count;
1442 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1443
1444 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1445 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1446
1447 binaries = SvRV (binaries);
1448
1449 if (device_count != av_len ((AV *)binaries) + 1)
1450 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1451
1452 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1453 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1454 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1455
1456 int i;
1457 for (i = 0; i < device_count; ++i)
1458 {
1459 STRLEN len;
1460 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1461 length_list [i] = len;
1462 }
1463
1464 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list,
1465 length_list, binary_list,
1466 GIMME_V == G_ARRAY ? status_list : 0, &res));
1467
1468 EXTEND (SP, 2);
1469 PUSH_CLOBJ (stash_program, prog);
1470
1471 if (GIMME_V == G_ARRAY)
1472 {
1473 AV *av = newAV ();
1474 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1475
1476 for (i = device_count; i--; )
1477 av_store (av, i, newSViv (status_list [i]));
1478 }
1479
1480#if CL_VERSION_1_2
1481
1482void
1483program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names)
1484 PPCODE:
1485 cl_uint device_count;
1486 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1487
1488 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res));
1489
1490 XPUSH_CLOBJ (stash_program, prog);
1491
1492void
1493link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef)
1494 CODE:
1495 cl_uint device_count = 0;
1496 cl_device_id *device_list = 0;
1497
1498 if (SvOK (devices))
1499 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
1500
1501 cl_uint program_count;
1502 cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count);
1503
1504 void *user_data;
1505 program_callback pfn_notify = make_program_callback (notify, &user_data);
1506
1507 NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
1508 program_count, program_list, pfn_notify, user_data, &res));
1509
1510 XPUSH_CLOBJ (stash_program, prog);
1511
1512#endif
1513
1514#BEGIN:context
1515
1516void
1517reference_count (OpenCL::Context self)
1518 ALIAS:
1519 reference_count = CL_CONTEXT_REFERENCE_COUNT
1520 num_devices = CL_CONTEXT_NUM_DEVICES
1521 PPCODE:
1522 cl_uint value [1];
1523 NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0));
1524 EXTEND (SP, 1);
1525 const int i = 0;
1526 PUSHs (sv_2mortal (newSVuv (value [i])));
1527
1528void
1529devices (OpenCL::Context self)
1530 PPCODE:
1531 size_t size;
1532 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size));
1533 cl_device_id *value = tmpbuf (size);
1534 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0));
1535 int i, n = size / sizeof (*value);
1536 EXTEND (SP, n);
1537 for (i = 0; i < n; ++i)
1538 PUSH_CLOBJ (stash_device, value [i]);
1539
1540void
1541properties (OpenCL::Context self)
1542 PPCODE:
1543 size_t size;
1544 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size));
1545 cl_context_properties *value = tmpbuf (size);
1546 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0));
1547 int i, n = size / sizeof (*value);
1548 EXTEND (SP, n);
1549 for (i = 0; i < n; ++i)
1550 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
1551
1552#END:context
449 1553
450MODULE = OpenCL PACKAGE = OpenCL::Queue 1554MODULE = OpenCL PACKAGE = OpenCL::Queue
451 1555
452void 1556void
453DESTROY (OpenCL::Queue this) 1557DESTROY (OpenCL::Queue self)
454 CODE: 1558 CODE:
455 clReleaseCommandQueue (this); 1559 clReleaseCommandQueue (self);
456 1560
457void 1561void
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, ...) 1562read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1563 ALIAS:
1564 enqueue_read_buffer = 0
464 PPCODE: 1565 PPCODE:
465{
466 cl_event ev = 0;
467 EVENT_LIST (6, items - 6); 1566 EVENT_LIST (6);
468 1567
469 SvUPGRADE (data, SVt_PV); 1568 SvUPGRADE (data, SVt_PV);
470 SvGROW (data, len); 1569 SvGROW (data, len);
471 SvPOK_only (data); 1570 SvPOK_only (data);
472 SvCUR_set (data, len); 1571 SvCUR_set (data, len);
1572
1573 cl_event ev = 0;
473 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1574 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
474 1575
475 if (ev) 1576 if (ev)
476 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1577 XPUSH_CLOBJ (stash_event, ev);
477}
478 1578
479void 1579void
480enqueue_write_buffer (OpenCL::Queue this, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) 1580write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1581 ALIAS:
1582 enqueue_write_buffer = 0
481 PPCODE: 1583 PPCODE:
482{ 1584 EVENT_LIST (5);
483 cl_event ev = 0; 1585
484 STRLEN len; 1586 STRLEN len;
485 char *ptr = SvPVbyte (data, len); 1587 char *ptr = SvPVbyte (data, len);
486 EVENT_LIST (5, items - 5);
487 1588
1589 cl_event ev = 0;
488 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1590 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
489 1591
490 if (ev) 1592 if (ev)
491 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1593 XPUSH_CLOBJ (stash_event, ev);
492}
493 1594
1595#if CL_VERSION_1_2
1596
494void 1597void
495enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1598fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1599 ALIAS:
1600 enqueue_fill_buffer = 0
496 PPCODE: 1601 PPCODE:
497{ 1602 EVENT_LIST (5);
1603
1604 STRLEN len;
1605 char *ptr = SvPVbyte (data, len);
1606
498 cl_event ev = 0; 1607 cl_event ev = 0;
499 EVENT_LIST (6, items - 6);
500
501 NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1608 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
502 1609
503 if (ev) 1610 if (ev)
504 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1611 XPUSH_CLOBJ (stash_event, ev);
505}
506 1612
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 1613void
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, ...) 1614fill_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, ...)
1615 ALIAS:
1616 enqueue_fill_image = 0
512 PPCODE: 1617 PPCODE:
513{ 1618 EVENT_LIST (12);
1619
1620 const size_t origin [3] = { x, y, z };
1621 const size_t region [3] = { width, height, depth };
1622
1623 const cl_float c_f [4] = { r, g, b, a };
1624 const cl_uint c_u [4] = { r, g, b, a };
1625 const cl_int c_s [4] = { r, g, b, a };
1626 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1627 static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 };
1628 cl_image_format format;
1629 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1630 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1631 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1632 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1633
514 cl_event ev = 0; 1634 cl_event ev = 0;
1635 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1636 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1637
1638 if (ev)
1639 XPUSH_CLOBJ (stash_event, ev);
1640
1641#endif
1642
1643void
1644copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1645 ALIAS:
1646 enqueue_copy_buffer = 0
1647 PPCODE:
1648 EVENT_LIST (6);
1649
1650 cl_event ev = 0;
1651 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1652
1653 if (ev)
1654 XPUSH_CLOBJ (stash_event, ev);
1655
1656void
1657read_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, ...)
1658 ALIAS:
1659 enqueue_read_buffer_rect = 0
1660 PPCODE:
1661 EVENT_LIST (17);
1662
1663 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 }; 1664 const size_t host_origin[3] = { host_x, host_y, host_z };
516 const size_t region[3] = { width, height, depth }; 1665 const size_t region[3] = { width, height, depth };
1666
1667 if (!buf_row_pitch)
1668 buf_row_pitch = region [0];
1669
1670 if (!buf_slice_pitch)
1671 buf_slice_pitch = region [1] * buf_row_pitch;
1672
1673 if (!host_row_pitch)
1674 host_row_pitch = region [0];
1675
1676 if (!host_slice_pitch)
1677 host_slice_pitch = region [1] * host_row_pitch;
1678
517 size_t len = row_pitch * slice_pitch * depth; 1679 size_t len = host_row_pitch * host_slice_pitch * region [2];
518 EVENT_LIST (11, items - 11);
519 1680
520 SvUPGRADE (data, SVt_PV); 1681 SvUPGRADE (data, SVt_PV);
521 SvGROW (data, len); 1682 SvGROW (data, len);
522 SvPOK_only (data); 1683 SvPOK_only (data);
523 SvCUR_set (data, len); 1684 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)); 1685
1686 cl_event ev = 0;
1687 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 1688
526 if (ev) 1689 if (ev)
527 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1690 XPUSH_CLOBJ (stash_event, ev);
528}
529 1691
530void 1692void
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, ...) 1693write_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, ...)
1694 ALIAS:
1695 enqueue_write_buffer_rect = 0
532 PPCODE: 1696 PPCODE:
533{ 1697 EVENT_LIST (17);
1698
1699 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1700 const size_t host_origin[3] = { host_x, host_y, host_z };
1701 const size_t region[3] = { width, height, depth };
1702 STRLEN len;
1703 char *ptr = SvPVbyte (data, len);
1704
1705 if (!buf_row_pitch)
1706 buf_row_pitch = region [0];
1707
1708 if (!buf_slice_pitch)
1709 buf_slice_pitch = region [1] * buf_row_pitch;
1710
1711 if (!host_row_pitch)
1712 host_row_pitch = region [0];
1713
1714 if (!host_slice_pitch)
1715 host_slice_pitch = region [1] * host_row_pitch;
1716
1717 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
1718
1719 if (len < min_len)
1720 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1721
534 cl_event ev = 0; 1722 cl_event ev = 0;
1723 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));
1724
1725 if (ev)
1726 XPUSH_CLOBJ (stash_event, ev);
1727
1728void
1729copy_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, ...)
1730 ALIAS:
1731 enqueue_copy_buffer_rect = 0
1732 PPCODE:
1733 EVENT_LIST (16);
1734
1735 const size_t src_origin[3] = { src_x, src_y, src_z };
1736 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1737 const size_t region[3] = { width, height, depth };
1738
1739 cl_event ev = 0;
1740 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));
1741
1742 if (ev)
1743 XPUSH_CLOBJ (stash_event, ev);
1744
1745void
1746read_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, ...)
1747 ALIAS:
1748 enqueue_read_image = 0
1749 PPCODE:
1750 EVENT_LIST (12);
1751
1752 const size_t src_origin[3] = { src_x, src_y, src_z };
1753 const size_t region[3] = { width, height, depth };
1754
1755 if (!row_pitch)
1756 row_pitch = img_row_pitch (src);
1757
1758 if (depth > 1 && !slice_pitch)
1759 slice_pitch = row_pitch * height;
1760
1761 size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1762
1763 SvUPGRADE (data, SVt_PV);
1764 SvGROW (data, len);
1765 SvPOK_only (data);
1766 SvCUR_set (data, len);
1767
1768 cl_event ev = 0;
1769 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));
1770
1771 if (ev)
1772 XPUSH_CLOBJ (stash_event, ev);
1773
1774void
1775write_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, ...)
1776 ALIAS:
1777 enqueue_write_image = 0
1778 PPCODE:
1779 EVENT_LIST (12);
1780
535 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1781 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
536 const size_t region[3] = { width, height, depth }; 1782 const size_t region[3] = { width, height, depth };
537 STRLEN len; 1783 STRLEN len;
538 char *ptr = SvPVbyte (data, len); 1784 char *ptr = SvPVbyte (data, len);
1785
1786 if (!row_pitch)
1787 row_pitch = img_row_pitch (dst);
1788
1789 if (depth > 1 && !slice_pitch)
539 size_t slice_pitch = len / (row_pitch * height); 1790 slice_pitch = row_pitch * height;
540 EVENT_LIST (11, items - 11);
541 1791
1792 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1793
1794 if (len < min_len)
1795 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1796
1797 cl_event ev = 0;
542 NEED_SUCCESS (EnqueueWriteImage, (this, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1798 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 1799
544 if (ev) 1800 if (ev)
545 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1801 XPUSH_CLOBJ (stash_event, ev);
546}
547 1802
548void 1803void
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, ...) 1804copy_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, ...)
1805 ALIAS:
1806 enqueue_copy_image = 0
550 PPCODE: 1807 PPCODE:
551{ 1808 EVENT_LIST (12);
552 cl_event ev = 0; 1809
553 const size_t src_origin[3] = { src_x, src_y, src_z }; 1810 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 }; 1811 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
555 const size_t region[3] = { width, height, depth }; 1812 const size_t region[3] = { width, height, depth };
556 EVENT_LIST (16, items - 16);
557 1813
1814 cl_event ev = 0;
558 NEED_SUCCESS (EnqueueCopyBufferRect, (this, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1815 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
559 1816
560 if (ev) 1817 if (ev)
561 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1818 XPUSH_CLOBJ (stash_event, ev);
562}
563 1819
564void 1820void
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, ...) 1821copy_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, ...)
1822 ALIAS:
1823 enqueue_copy_image_to_buffer = 0
566 PPCODE: 1824 PPCODE:
567{ 1825 EVENT_LIST (10);
1826
1827 const size_t src_origin[3] = { src_x, src_y, src_z };
1828 const size_t region [3] = { width, height, depth };
1829
568 cl_event ev = 0; 1830 cl_event ev = 0;
1831 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1832
1833 if (ev)
1834 XPUSH_CLOBJ (stash_event, ev);
1835
1836void
1837copy_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, ...)
1838 ALIAS:
1839 enqueue_copy_buffer_to_image = 0
1840 PPCODE:
1841 EVENT_LIST (10);
1842
569 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1843 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1844 const size_t region [3] = { width, height, depth };
1845
1846 cl_event ev = 0;
1847 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1848
1849 if (ev)
1850 XPUSH_CLOBJ (stash_event, ev);
1851
1852void
1853map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, SV *cb_ = &PL_sv_undef, ...)
1854 ALIAS:
1855 enqueue_map_buffer = 0
1856 PPCODE:
1857 EVENT_LIST (6);
1858
1859 size_t cb = SvIV (cb_);
1860
1861 if (!SvOK (cb_))
1862 {
1863 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1864 cb -= offset;
1865 }
1866
1867 cl_event ev;
1868 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
1869 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0, 1, cb, 1, 1));
1870
1871void
1872map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...)
1873 ALIAS:
1874 enqueue_map_image = 0
1875 PPCODE:
1876 size_t width = SvIV (width_);
1877 if (!SvOK (width_))
1878 {
1879 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0));
1880 width -= x;
1881 }
1882
1883 size_t height = SvIV (width_);
1884 if (!SvOK (height_))
1885 {
1886 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0));
1887 height -= y;
1888
1889 // stupid opencl returns 0 for depth, but requires 1 for 2d images
1890 if (!height)
1891 height = 1;
1892 }
1893
1894 size_t depth = SvIV (width_);
1895 if (!SvOK (depth_))
1896 {
1897 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0));
1898 depth -= z;
1899
1900 // stupid opencl returns 0 for depth, but requires 1 for 2d images
1901 if (!depth)
1902 depth = 1;
1903 }
1904
1905 const size_t origin[3] = { x, y, z };
570 const size_t region[3] = { width, height, depth }; 1906 const size_t region[3] = { width, height, depth };
1907 size_t row_pitch, slice_pitch;
571 EVENT_LIST (10, items - 10); 1908 EVENT_LIST (10);
572 1909
573 NEED_SUCCESS (EnqueueCopyBufferToImage, (this, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1910 cl_event ev;
1911 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));
1912
1913 size_t cb = slice_pitch ? slice_pitch * region [2]
1914 : row_pitch ? row_pitch * region [1]
1915 : region [0];
1916
1917 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch, 0, width, height, depth));
1918
1919void
1920unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1921 PPCODE:
1922 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1923 if (GIMME_V != G_VOID)
1924 {
1925 clRetainEvent (mapped->event);
1926 XPUSH_CLOBJ (stash_event, mapped->event);
1927 }
1928
1929void
1930task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1931 ALIAS:
1932 enqueue_task = 0
1933 PPCODE:
1934 EVENT_LIST (2);
1935
1936 cl_event ev = 0;
1937 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
574 1938
575 if (ev) 1939 if (ev)
576 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1940 XPUSH_CLOBJ (stash_event, ev);
577}
578 1941
579void 1942void
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, ...) 1943nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1944 ALIAS:
1945 enqueue_nd_range_kernel = 0
625 PPCODE: 1946 PPCODE:
626{ 1947 EVENT_LIST (5);
627 cl_event ev = 0; 1948
628 size_t *gwo = 0, *gws, *lws = 0; 1949 size_t *gwo = 0, *gws, *lws = 0;
629 int gws_len; 1950 int gws_len;
630 size_t *lists; 1951 size_t *lists;
631 int i; 1952 int i;
632 EVENT_LIST (5, items - 5);
633 1953
634 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) 1954 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
635 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1955 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
636 1956
637 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1957 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
638 1958
639 lists = tmpbuf (sizeof (size_t) * 3 * gws_len); 1959 lists = tmpbuf (sizeof (size_t) * 3 * gws_len);
640 1960
641 gws = lists + gws_len * 0; 1961 gws = lists + gws_len * 0;
642 for (i = 0; i < gws_len; ++i) 1962 for (i = 0; i < gws_len; ++i)
1963 {
643 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); 1964 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]);
1965 // at least nvidia crashes for 0-sized work group sizes, work around
1966 if (!gws [i])
1967 croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i);
1968 }
644 1969
645 if (SvOK (global_work_offset)) 1970 if (SvOK (global_work_offset))
646 { 1971 {
647 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) 1972 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"); 1973 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference");
655 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 1980 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
656 } 1981 }
657 1982
658 if (SvOK (local_work_size)) 1983 if (SvOK (local_work_size))
659 { 1984 {
660 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 1985 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"); 1986 croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference");
662 1987
663 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 1988 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"); 1989 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
665 1990
666 lws = lists + gws_len * 2; 1991 lws = lists + gws_len * 2;
667 for (i = 0; i < gws_len; ++i) 1992 for (i = 0; i < gws_len; ++i)
1993 {
668 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); 1994 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]);
1995 // at least nvidia crashes for 0-sized work group sizes, work around
1996 if (!lws [i])
1997 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
1998 }
669 } 1999 }
670 2000
2001 cl_event ev = 0;
671 NEED_SUCCESS (EnqueueNDRangeKernel, (this, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2002 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
672 2003
673 if (ev) 2004 if (ev)
674 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 2005 XPUSH_CLOBJ (stash_event, ev);
675}
676 2006
2007#if CL_VERSION_1_2
2008
677void 2009void
678enqueue_marker (OpenCL::Queue this) 2010migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...)
2011 ALIAS:
2012 enqueue_migrate_mem_objects = 0
679 PPCODE: 2013 PPCODE:
680{ 2014 EVENT_LIST (3);
2015
2016 cl_uint obj_count;
2017 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2018
2019 cl_event ev = 0;
2020 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2021
2022 if (ev)
2023 XPUSH_CLOBJ (stash_event, ev);
2024
2025#endif
2026
2027#if cl_apple_gl_sharing || cl_khr_gl_sharing
2028
2029void
2030acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
2031 ALIAS:
2032 release_gl_objects = 1
2033 enqueue_acquire_gl_objects = 0
2034 enqueue_release_gl_objects = 1
2035 PPCODE:
2036 EVENT_LIST (2);
2037
2038 cl_uint obj_count;
2039 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2040
681 cl_event ev; 2041 cl_event ev = 0;
682 NEED_SUCCESS (EnqueueMarker, (this, &ev));
683 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
684}
685 2042
2043 if (ix)
2044 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2045 else
2046 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2047
2048 if (ev)
2049 XPUSH_CLOBJ (stash_event, ev);
2050
2051#endif
2052
686void 2053void
687enqueue_wait_for_events (OpenCL::Queue this, ...) 2054wait_for_events (OpenCL::Queue self, ...)
2055 ALIAS:
2056 enqueue_wait_for_events = 0
688 CODE: 2057 CODE:
689{
690 EVENT_LIST (1, items - 1); 2058 EVENT_LIST (1);
2059#if PREFER_1_1
691 NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); 2060 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
692} 2061#else
2062 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0));
2063#endif
693 2064
694void 2065void
695enqueue_barrier (OpenCL::Queue this) 2066marker (OpenCL::Queue self, ...)
2067 ALIAS:
2068 enqueue_marker = 0
696 CODE: 2069 PPCODE:
2070 EVENT_LIST (1);
2071 cl_event ev = 0;
2072#if PREFER_1_1
2073 if (!event_list_count)
2074 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2075 else
2076#if CL_VERSION_1_2
2077 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2078#else
2079 {
2080 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier
2081 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2082 }
2083#endif
2084#else
2085 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2086#endif
2087 if (ev)
2088 XPUSH_CLOBJ (stash_event, ev);
2089
2090void
2091barrier (OpenCL::Queue self, ...)
2092 ALIAS:
2093 enqueue_barrier = 0
2094 PPCODE:
2095 EVENT_LIST (1);
2096 cl_event ev = 0;
2097#if PREFER_1_1
2098 if (!event_list_count && GIMME_V == G_VOID)
697 NEED_SUCCESS (EnqueueBarrier, (this)); 2099 NEED_SUCCESS (EnqueueBarrier, (self));
2100 else
2101#if CL_VERSION_1_2
2102 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2103#else
2104 {
2105 if (event_list_count)
2106 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
698 2107
2108 if (GIMME_V != G_VOID)
2109 NEED_SUCCESS (EnqueueMarker, (self, &ev));
2110 }
2111#endif
2112#else
2113 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2114#endif
2115 if (ev)
2116 XPUSH_CLOBJ (stash_event, ev);
2117
699void 2118void
700flush (OpenCL::Queue this) 2119flush (OpenCL::Queue self)
701 CODE: 2120 CODE:
702 NEED_SUCCESS (Flush, (this)); 2121 NEED_SUCCESS (Flush, (self));
703 2122
704void 2123void
705finish (OpenCL::Queue this) 2124finish (OpenCL::Queue self)
706 CODE: 2125 CODE:
707 NEED_SUCCESS (Finish, (this)); 2126 NEED_SUCCESS (Finish, (self));
2127
2128void
2129info (OpenCL::Queue self, cl_command_queue_info name)
2130 PPCODE:
2131 INFO (CommandQueue)
2132
2133#BEGIN:command_queue
2134
2135void
2136context (OpenCL::Queue self)
2137 PPCODE:
2138 cl_context value [1];
2139 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0));
2140 EXTEND (SP, 1);
2141 const int i = 0;
2142 NEED_SUCCESS (RetainContext, (value [i]));
2143 PUSH_CLOBJ (stash_context, value [i]);
2144
2145void
2146device (OpenCL::Queue self)
2147 PPCODE:
2148 cl_device_id value [1];
2149 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0));
2150 EXTEND (SP, 1);
2151 const int i = 0;
2152 PUSH_CLOBJ (stash_device, value [i]);
2153
2154void
2155reference_count (OpenCL::Queue self)
2156 PPCODE:
2157 cl_uint value [1];
2158 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0));
2159 EXTEND (SP, 1);
2160 const int i = 0;
2161 PUSHs (sv_2mortal (newSVuv (value [i])));
2162
2163void
2164properties (OpenCL::Queue self)
2165 PPCODE:
2166 cl_ulong value [1];
2167 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0));
2168 EXTEND (SP, 1);
2169 const int i = 0;
2170 PUSHs (sv_2mortal (newSVuv (value [i])));
2171
2172#END:command_queue
708 2173
709MODULE = OpenCL PACKAGE = OpenCL::Memory 2174MODULE = OpenCL PACKAGE = OpenCL::Memory
710 2175
711void 2176void
712DESTROY (OpenCL::Memory this) 2177DESTROY (OpenCL::Memory self)
713 CODE: 2178 CODE:
714 clReleaseMemObject (this); 2179 clReleaseMemObject (self);
715 2180
716void 2181void
717info (OpenCL::Memory this, cl_mem_info name) 2182info (OpenCL::Memory self, cl_mem_info name)
718 PPCODE: 2183 PPCODE:
719 INFO (MemObject) 2184 INFO (MemObject)
720 2185
2186void
2187destructor_callback (OpenCL::Memory self, SV *notify)
2188 PPCODE:
2189 clSetMemObjectDestructorCallback (self, eq_destructor_notify, SvREFCNT_inc (s_get_cv (notify)));
2190
2191#BEGIN:mem
2192
2193void
2194type (OpenCL::Memory self)
2195 ALIAS:
2196 type = CL_MEM_TYPE
2197 map_count = CL_MEM_MAP_COUNT
2198 reference_count = CL_MEM_REFERENCE_COUNT
2199 PPCODE:
2200 cl_uint value [1];
2201 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
2202 EXTEND (SP, 1);
2203 const int i = 0;
2204 PUSHs (sv_2mortal (newSVuv (value [i])));
2205
2206void
2207flags (OpenCL::Memory self)
2208 PPCODE:
2209 cl_ulong value [1];
2210 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0));
2211 EXTEND (SP, 1);
2212 const int i = 0;
2213 PUSHs (sv_2mortal (newSVuv (value [i])));
2214
2215void
2216size (OpenCL::Memory self)
2217 ALIAS:
2218 size = CL_MEM_SIZE
2219 offset = CL_MEM_OFFSET
2220 PPCODE:
2221 size_t value [1];
2222 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
2223 EXTEND (SP, 1);
2224 const int i = 0;
2225 PUSHs (sv_2mortal (newSVuv (value [i])));
2226
2227void
2228host_ptr (OpenCL::Memory self)
2229 PPCODE:
2230 void * value [1];
2231 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0));
2232 EXTEND (SP, 1);
2233 const int i = 0;
2234 PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i])));
2235
2236void
2237context (OpenCL::Memory self)
2238 PPCODE:
2239 cl_context value [1];
2240 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0));
2241 EXTEND (SP, 1);
2242 const int i = 0;
2243 NEED_SUCCESS (RetainContext, (value [i]));
2244 PUSH_CLOBJ (stash_context, value [i]);
2245
2246void
2247associated_memobject (OpenCL::Memory self)
2248 PPCODE:
2249 cl_mem value [1];
2250 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0));
2251 EXTEND (SP, 1);
2252 const int i = 0;
2253 NEED_SUCCESS (RetainMemObject, (value [i]));
2254 PUSH_CLOBJ (stash_memory, value [i]);
2255
2256#END:mem
2257
2258#if cl_apple_gl_sharing || cl_khr_gl_sharing
2259
2260void
2261gl_object_info (OpenCL::Memory self)
2262 PPCODE:
2263 cl_gl_object_type type;
2264 cl_GLuint name;
2265 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
2266 EXTEND (SP, 2);
2267 PUSHs (sv_2mortal (newSVuv (type)));
2268 PUSHs (sv_2mortal (newSVuv (name)));
2269
2270#endif
2271
2272MODULE = OpenCL PACKAGE = OpenCL::BufferObj
2273
2274void
2275sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
2276 PPCODE:
2277 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR))
2278 croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!");
2279
2280 cl_buffer_region crdata = { origin, size };
2281
2282 NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res));
2283 XPUSH_CLOBJ (stash_buffer, mem);
2284
2285MODULE = OpenCL PACKAGE = OpenCL::Image
2286
2287void
2288image_info (OpenCL::Image self, cl_image_info name)
2289 PPCODE:
2290 INFO (Image)
2291
2292void
2293format (OpenCL::Image self)
2294 PPCODE:
2295 cl_image_format format;
2296 NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
2297 EXTEND (SP, 2);
2298 PUSHs (sv_2mortal (newSVuv (format.image_channel_order)));
2299 PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type)));
2300
2301#BEGIN:image
2302
2303void
2304element_size (OpenCL::Image self)
2305 ALIAS:
2306 element_size = CL_IMAGE_ELEMENT_SIZE
2307 row_pitch = CL_IMAGE_ROW_PITCH
2308 slice_pitch = CL_IMAGE_SLICE_PITCH
2309 width = CL_IMAGE_WIDTH
2310 height = CL_IMAGE_HEIGHT
2311 depth = CL_IMAGE_DEPTH
2312 PPCODE:
2313 size_t value [1];
2314 NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0));
2315 EXTEND (SP, 1);
2316 const int i = 0;
2317 PUSHs (sv_2mortal (newSVuv (value [i])));
2318
2319#END:image
2320
2321#if cl_apple_gl_sharing || cl_khr_gl_sharing
2322
2323#BEGIN:gl_texture
2324
2325void
2326target (OpenCL::Image self)
2327 PPCODE:
2328 cl_GLenum value [1];
2329 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
2330 EXTEND (SP, 1);
2331 const int i = 0;
2332 PUSHs (sv_2mortal (newSVuv (value [i])));
2333
2334void
2335gl_mipmap_level (OpenCL::Image self)
2336 PPCODE:
2337 cl_GLint value [1];
2338 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
2339 EXTEND (SP, 1);
2340 const int i = 0;
2341 PUSHs (sv_2mortal (newSViv (value [i])));
2342
2343#END:gl_texture
2344
2345#endif
2346
721MODULE = OpenCL PACKAGE = OpenCL::Sampler 2347MODULE = OpenCL PACKAGE = OpenCL::Sampler
722 2348
723void 2349void
724DESTROY (OpenCL::Sampler this) 2350DESTROY (OpenCL::Sampler self)
725 CODE: 2351 CODE:
726 clReleaseSampler (this); 2352 clReleaseSampler (self);
727 2353
728void 2354void
729info (OpenCL::Sampler this, cl_sampler_info name) 2355info (OpenCL::Sampler self, cl_sampler_info name)
730 PPCODE: 2356 PPCODE:
731 INFO (Sampler) 2357 INFO (Sampler)
732 2358
2359#BEGIN:sampler
2360
2361void
2362reference_count (OpenCL::Sampler self)
2363 ALIAS:
2364 reference_count = CL_SAMPLER_REFERENCE_COUNT
2365 normalized_coords = CL_SAMPLER_NORMALIZED_COORDS
2366 addressing_mode = CL_SAMPLER_ADDRESSING_MODE
2367 PPCODE:
2368 cl_uint value [1];
2369 NEED_SUCCESS (GetSamplerInfo, (self, ix, sizeof (value), value, 0));
2370 EXTEND (SP, 1);
2371 const int i = 0;
2372 PUSHs (sv_2mortal (newSVuv (value [i])));
2373
2374void
2375context (OpenCL::Sampler self)
2376 PPCODE:
2377 cl_context value [1];
2378 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0));
2379 EXTEND (SP, 1);
2380 const int i = 0;
2381 NEED_SUCCESS (RetainContext, (value [i]));
2382 PUSH_CLOBJ (stash_context, value [i]);
2383
2384void
2385filter_mode (OpenCL::Sampler self)
2386 PPCODE:
2387 cl_uint value [1];
2388 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0));
2389 EXTEND (SP, 1);
2390 const int i = 0;
2391 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
2392
2393#END:sampler
2394
733MODULE = OpenCL PACKAGE = OpenCL::Program 2395MODULE = OpenCL PACKAGE = OpenCL::Program
734 2396
735void 2397void
736DESTROY (OpenCL::Program this) 2398DESTROY (OpenCL::Program self)
737 CODE: 2399 CODE:
738 clReleaseProgram (this); 2400 clReleaseProgram (self);
739 2401
740void 2402void
741info (OpenCL::Program this, cl_program_info name) 2403build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2404 ALIAS:
2405 build_async = 1
742 PPCODE: 2406 CODE:
743 INFO (Program) 2407 cl_uint device_count = 0;
2408 cl_device_id *device_list = 0;
744 2409
2410 if (SvOK (devices))
2411 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2412
2413 void *user_data;
2414 program_callback pfn_notify = make_program_callback (notify, &user_data);
2415
2416 if (ix)
2417 build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data);
2418 else
2419 NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
2420
2421#if CL_VERSION_1_2
2422
745void 2423void
746build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) 2424compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef)
747 CODE: 2425 CODE:
2426 cl_uint device_count = 0;
2427 cl_device_id *device_list = 0;
2428
2429 if (SvOK (devices))
2430 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2431
2432 cl_uint header_count = 0;
2433 cl_program *header_list = 0;
2434 const char **header_name = 0;
2435
2436 if (SvOK (headers))
2437 {
2438 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2439 croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs");
2440
2441 HV *hv = (HV *)SvRV (devices);
2442
2443 header_count = hv_iterinit (hv);
2444 header_list = tmpbuf (sizeof (*header_list) * header_count);
2445 header_name = tmpbuf (sizeof (*header_name) * header_count);
2446
2447 HE *he;
2448 int i = 0;
2449 while (he = hv_iternext (hv))
2450 {
2451 header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he));
2452 header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program");
2453 ++i;
2454 }
2455 }
2456
2457 void *user_data;
2458 program_callback pfn_notify = make_program_callback (notify, &user_data);
2459
748 NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); 2460 NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
2461 header_count, header_list, header_name, pfn_notify, user_data));
749 2462
2463#endif
2464
750void 2465void
751build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) 2466build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
752 PPCODE: 2467 PPCODE:
753{
754 size_t size; 2468 size_t size;
755 SV *sv;
756
757 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); 2469 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size));
758 sv = sv_2mortal (newSV (size)); 2470 SV *sv = sv_2mortal (newSV (size));
759 SvUPGRADE (sv, SVt_PV); 2471 SvUPGRADE (sv, SVt_PV);
760 SvPOK_only (sv); 2472 SvPOK_only (sv);
761 SvCUR_set (sv, size); 2473 SvCUR_set (sv, size);
762 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); 2474 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0));
763 XPUSHs (sv); 2475 XPUSHs (sv);
764} 2476
2477#BEGIN:program_build
2478
2479void
2480build_status (OpenCL::Program self, OpenCL::Device device)
2481 PPCODE:
2482 cl_int value [1];
2483 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0));
2484 EXTEND (SP, 1);
2485 const int i = 0;
2486 PUSHs (sv_2mortal (newSViv (value [i])));
2487
2488void
2489build_options (OpenCL::Program self, OpenCL::Device device)
2490 ALIAS:
2491 build_options = CL_PROGRAM_BUILD_OPTIONS
2492 build_log = CL_PROGRAM_BUILD_LOG
2493 PPCODE:
2494 size_t size;
2495 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size));
2496 char *value = tmpbuf (size);
2497 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2498 EXTEND (SP, 1);
2499 const int i = 0;
2500 PUSHs (sv_2mortal (newSVpv (value, 0)));
2501
2502void
2503binary_type (OpenCL::Program self, OpenCL::Device device)
2504 PPCODE:
2505 cl_uint value [1];
2506 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0));
2507 EXTEND (SP, 1);
2508 const int i = 0;
2509 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
2510
2511#END:program_build
765 2512
766void 2513void
767kernel (OpenCL::Program program, SV *function) 2514kernel (OpenCL::Program program, SV *function)
768 PPCODE: 2515 PPCODE:
769{
770 cl_int res;
771 cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); 2516 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
2517 XPUSH_CLOBJ (stash_kernel, kernel);
772 2518
773 if (res) 2519void
774 FAIL (CreateKernel, res); 2520kernels_in_program (OpenCL::Program program)
2521 PPCODE:
2522 cl_uint num_kernels;
2523 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
2524 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
2525 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
775 2526
776 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 2527 int i;
777} 2528 EXTEND (SP, num_kernels);
2529 for (i = 0; i < num_kernels; ++i)
2530 PUSH_CLOBJ (stash_kernel, kernels [i]);
2531
2532void
2533info (OpenCL::Program self, cl_program_info name)
2534 PPCODE:
2535 INFO (Program)
2536
2537void
2538binaries (OpenCL::Program self)
2539 PPCODE:
2540 cl_uint n, i;
2541 size_t size;
2542
2543 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0));
2544 if (!n) XSRETURN_EMPTY;
2545
2546 size_t *sizes = tmpbuf (sizeof (*sizes) * n);
2547 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size));
2548 if (size != sizeof (*sizes) * n) XSRETURN_EMPTY;
2549 unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n);
2550
2551 EXTEND (SP, n);
2552 for (i = 0; i < n; ++i)
2553 {
2554 SV *sv = sv_2mortal (newSV (sizes [i]));
2555 SvUPGRADE (sv, SVt_PV);
2556 SvPOK_only (sv);
2557 SvCUR_set (sv, sizes [i]);
2558 ptrs [i] = (void *)SvPVX (sv);
2559 PUSHs (sv);
2560 }
2561
2562 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
2563 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
2564
2565#BEGIN:program
2566
2567void
2568reference_count (OpenCL::Program self)
2569 ALIAS:
2570 reference_count = CL_PROGRAM_REFERENCE_COUNT
2571 num_devices = CL_PROGRAM_NUM_DEVICES
2572 PPCODE:
2573 cl_uint value [1];
2574 NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0));
2575 EXTEND (SP, 1);
2576 const int i = 0;
2577 PUSHs (sv_2mortal (newSVuv (value [i])));
2578
2579void
2580context (OpenCL::Program self)
2581 PPCODE:
2582 cl_context value [1];
2583 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0));
2584 EXTEND (SP, 1);
2585 const int i = 0;
2586 NEED_SUCCESS (RetainContext, (value [i]));
2587 PUSH_CLOBJ (stash_context, value [i]);
2588
2589void
2590devices (OpenCL::Program self)
2591 PPCODE:
2592 size_t size;
2593 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size));
2594 cl_device_id *value = tmpbuf (size);
2595 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0));
2596 int i, n = size / sizeof (*value);
2597 EXTEND (SP, n);
2598 for (i = 0; i < n; ++i)
2599 PUSH_CLOBJ (stash_device, value [i]);
2600
2601void
2602source (OpenCL::Program self)
2603 PPCODE:
2604 size_t size;
2605 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size));
2606 char *value = tmpbuf (size);
2607 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0));
2608 EXTEND (SP, 1);
2609 const int i = 0;
2610 PUSHs (sv_2mortal (newSVpv (value, 0)));
2611
2612void
2613binary_sizes (OpenCL::Program self)
2614 PPCODE:
2615 size_t size;
2616 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size));
2617 size_t *value = tmpbuf (size);
2618 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0));
2619 int i, n = size / sizeof (*value);
2620 EXTEND (SP, n);
2621 for (i = 0; i < n; ++i)
2622 PUSHs (sv_2mortal (newSVuv (value [i])));
2623
2624#END:program
778 2625
779MODULE = OpenCL PACKAGE = OpenCL::Kernel 2626MODULE = OpenCL PACKAGE = OpenCL::Kernel
780 2627
781void 2628void
782DESTROY (OpenCL::Kernel this) 2629DESTROY (OpenCL::Kernel self)
783 CODE: 2630 CODE:
784 clReleaseKernel (this); 2631 clReleaseKernel (self);
785 2632
786void 2633void
2634setf (OpenCL::Kernel self, const char *format, ...)
2635 CODE:
2636 int i;
2637 for (i = 2; ; ++i)
2638 {
2639 while (*format == ' ')
2640 ++format;
2641
2642 char type = *format++;
2643
2644 if (!type)
2645 break;
2646
2647 if (i >= items)
2648 croak ("OpenCL::Kernel::setf format string too long (not enough arguments)");
2649
2650 SV *sv = ST (i);
2651
2652 union
2653 {
2654 cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS;
2655 cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL;
2656 cl_half ch; cl_float cf; cl_double cd;
2657 cl_mem cm;
2658 cl_sampler ca;
2659 size_t cz;
2660 cl_event ce;
2661 } arg;
2662 size_t size;
2663 int nullarg = 0;
2664
2665 switch (type)
2666 {
2667 case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break;
2668 case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break;
2669 case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break;
2670 case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break;
2671 case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break;
2672 case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break;
2673 case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break;
2674 case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break;
2675
2676 case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break;
2677 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2678 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2679
2680 case 'z': nullarg = 1; size = SvIV (sv); break;
2681
2682 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2683 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2684 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2685
2686 default:
2687 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2688 }
2689
2690 res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg);
2691 if (res)
2692 croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res));
2693 }
2694
2695 if (i != items)
2696 croak ("OpenCL::Kernel::setf format string too short (too many arguments)");
2697
2698void
2699set_char (OpenCL::Kernel self, cl_uint idx, cl_char value)
2700 CODE:
2701 clSetKernelArg (self, idx, sizeof (value), &value);
2702
2703void
2704set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value)
2705 CODE:
2706 clSetKernelArg (self, idx, sizeof (value), &value);
2707
2708void
2709set_short (OpenCL::Kernel self, cl_uint idx, cl_short value)
2710 CODE:
2711 clSetKernelArg (self, idx, sizeof (value), &value);
2712
2713void
2714set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value)
2715 CODE:
2716 clSetKernelArg (self, idx, sizeof (value), &value);
2717
2718void
2719set_int (OpenCL::Kernel self, cl_uint idx, cl_int value)
2720 CODE:
2721 clSetKernelArg (self, idx, sizeof (value), &value);
2722
2723void
2724set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value)
2725 CODE:
2726 clSetKernelArg (self, idx, sizeof (value), &value);
2727
2728void
2729set_long (OpenCL::Kernel self, cl_uint idx, cl_long value)
2730 CODE:
2731 clSetKernelArg (self, idx, sizeof (value), &value);
2732
2733void
2734set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value)
2735 CODE:
2736 clSetKernelArg (self, idx, sizeof (value), &value);
2737
2738void
2739set_half (OpenCL::Kernel self, cl_uint idx, cl_half value)
2740 CODE:
2741 clSetKernelArg (self, idx, sizeof (value), &value);
2742
2743void
2744set_float (OpenCL::Kernel self, cl_uint idx, cl_float value)
2745 CODE:
2746 clSetKernelArg (self, idx, sizeof (value), &value);
2747
2748void
2749set_double (OpenCL::Kernel self, cl_uint idx, cl_double value)
2750 CODE:
2751 clSetKernelArg (self, idx, sizeof (value), &value);
2752
2753void
2754set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value)
2755 CODE:
2756 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2757
2758void
2759set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value)
2760 CODE:
2761 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2762
2763void
2764set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value)
2765 CODE:
2766 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2767
2768void
2769set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
2770 CODE:
2771 clSetKernelArg (self, idx, sizeof (value), &value);
2772
2773void
2774set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
2775 CODE:
2776 clSetKernelArg (self, idx, size, 0);
2777
2778void
2779set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
2780 CODE:
2781 clSetKernelArg (self, idx, sizeof (value), &value);
2782
2783void
787info (OpenCL::Kernel this, cl_kernel_info name) 2784info (OpenCL::Kernel self, cl_kernel_info name)
788 PPCODE: 2785 PPCODE:
789 INFO (Kernel) 2786 INFO (Kernel)
790 2787
2788#BEGIN:kernel
2789
791void 2790void
792set_char (OpenCL::Kernel this, cl_uint idx, cl_char value) 2791function_name (OpenCL::Kernel self)
2792 PPCODE:
2793 size_t size;
2794 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size));
2795 char *value = tmpbuf (size);
2796 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0));
2797 EXTEND (SP, 1);
2798 const int i = 0;
2799 PUSHs (sv_2mortal (newSVpv (value, 0)));
2800
2801void
2802num_args (OpenCL::Kernel self)
2803 ALIAS:
2804 num_args = CL_KERNEL_NUM_ARGS
2805 reference_count = CL_KERNEL_REFERENCE_COUNT
2806 PPCODE:
2807 cl_uint value [1];
2808 NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0));
2809 EXTEND (SP, 1);
2810 const int i = 0;
2811 PUSHs (sv_2mortal (newSVuv (value [i])));
2812
2813void
2814context (OpenCL::Kernel self)
2815 PPCODE:
2816 cl_context value [1];
2817 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0));
2818 EXTEND (SP, 1);
2819 const int i = 0;
2820 NEED_SUCCESS (RetainContext, (value [i]));
2821 PUSH_CLOBJ (stash_context, value [i]);
2822
2823void
2824program (OpenCL::Kernel self)
2825 PPCODE:
2826 cl_program value [1];
2827 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0));
2828 EXTEND (SP, 1);
2829 const int i = 0;
2830 NEED_SUCCESS (RetainProgram, (value [i]));
2831 PUSH_CLOBJ (stash_program, value [i]);
2832
2833#END:kernel
2834
2835void
2836work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name)
793 CODE: 2837 PPCODE:
794 clSetKernelArg (this, idx, sizeof (value), &value); 2838 size_t size;
2839 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size));
2840 SV *sv = sv_2mortal (newSV (size));
2841 SvUPGRADE (sv, SVt_PV);
2842 SvPOK_only (sv);
2843 SvCUR_set (sv, size);
2844 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0));
2845 XPUSHs (sv);
795 2846
2847#BEGIN:kernel_work_group
2848
796void 2849void
797set_uchar (OpenCL::Kernel this, cl_uint idx, cl_uchar value) 2850work_group_size (OpenCL::Kernel self, OpenCL::Device device)
2851 ALIAS:
2852 work_group_size = CL_KERNEL_WORK_GROUP_SIZE
2853 preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE
2854 PPCODE:
2855 size_t value [1];
2856 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2857 EXTEND (SP, 1);
2858 const int i = 0;
2859 PUSHs (sv_2mortal (newSVuv (value [i])));
2860
2861void
2862compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device)
2863 PPCODE:
2864 size_t size;
2865 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size));
2866 size_t *value = tmpbuf (size);
2867 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0));
2868 int i, n = size / sizeof (*value);
2869 EXTEND (SP, n);
2870 for (i = 0; i < n; ++i)
2871 PUSHs (sv_2mortal (newSVuv (value [i])));
2872
2873void
2874local_mem_size (OpenCL::Kernel self, OpenCL::Device device)
2875 ALIAS:
2876 local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE
2877 private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE
2878 PPCODE:
2879 cl_ulong value [1];
2880 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2881 EXTEND (SP, 1);
2882 const int i = 0;
2883 PUSHs (sv_2mortal (newSVuv (value [i])));
2884
2885#END:kernel_work_group
2886
2887#if CL_VERSION_1_2
2888
2889void
2890arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name)
798 CODE: 2891 PPCODE:
799 clSetKernelArg (this, idx, sizeof (value), &value); 2892 size_t size;
2893 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size));
2894 SV *sv = sv_2mortal (newSV (size));
2895 SvUPGRADE (sv, SVt_PV);
2896 SvPOK_only (sv);
2897 SvCUR_set (sv, size);
2898 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0));
2899 XPUSHs (sv);
800 2900
801void 2901#BEGIN:kernel_arg
802set_short (OpenCL::Kernel this, cl_uint idx, cl_short value)
803 CODE:
804 clSetKernelArg (this, idx, sizeof (value), &value);
805 2902
806void 2903void
807set_ushort (OpenCL::Kernel this, cl_uint idx, cl_ushort value) 2904arg_address_qualifier (OpenCL::Kernel self, cl_uint idx)
808 CODE: 2905 ALIAS:
809 clSetKernelArg (this, idx, sizeof (value), &value); 2906 arg_address_qualifier = CL_KERNEL_ARG_ADDRESS_QUALIFIER
2907 arg_access_qualifier = CL_KERNEL_ARG_ACCESS_QUALIFIER
2908 PPCODE:
2909 cl_uint value [1];
2910 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, sizeof (value), value, 0));
2911 EXTEND (SP, 1);
2912 const int i = 0;
2913 PUSHs (sv_2mortal (newSVuv (value [i])));
810 2914
811void 2915void
812set_int (OpenCL::Kernel this, cl_uint idx, cl_int value) 2916arg_type_name (OpenCL::Kernel self, cl_uint idx)
813 CODE: 2917 ALIAS:
814 clSetKernelArg (this, idx, sizeof (value), &value); 2918 arg_type_name = CL_KERNEL_ARG_TYPE_NAME
2919 arg_name = CL_KERNEL_ARG_NAME
2920 PPCODE:
2921 size_t size;
2922 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size));
2923 char *value = tmpbuf (size);
2924 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0));
2925 EXTEND (SP, 1);
2926 const int i = 0;
2927 PUSHs (sv_2mortal (newSVpv (value, 0)));
815 2928
816void 2929void
817set_uint (OpenCL::Kernel this, cl_uint idx, cl_uint value) 2930arg_type_qualifier (OpenCL::Kernel self, cl_uint idx)
818 CODE: 2931 PPCODE:
819 clSetKernelArg (this, idx, sizeof (value), &value); 2932 cl_ulong value [1];
2933 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0));
2934 EXTEND (SP, 1);
2935 const int i = 0;
2936 PUSHs (sv_2mortal (newSVuv (value [i])));
820 2937
821void 2938#END:kernel_arg
822set_long (OpenCL::Kernel this, cl_uint idx, cl_long value)
823 CODE:
824 clSetKernelArg (this, idx, sizeof (value), &value);
825 2939
826void 2940#endif
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 2941
876MODULE = OpenCL PACKAGE = OpenCL::Event 2942MODULE = OpenCL PACKAGE = OpenCL::Event
877 2943
878void 2944void
879DESTROY (OpenCL::Event this) 2945DESTROY (OpenCL::Event self)
880 CODE: 2946 CODE:
881 clReleaseEvent (this); 2947 clReleaseEvent (self);
882 2948
883void 2949void
2950wait (OpenCL::Event self)
2951 CODE:
2952 clWaitForEvents (1, &self);
2953
2954void
2955cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb)
2956 CODE:
2957 clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb)));
2958
2959void
884info (OpenCL::Event this, cl_event_info name) 2960info (OpenCL::Event self, cl_event_info name)
885 PPCODE: 2961 PPCODE:
886 INFO (Event) 2962 INFO (Event)
887 2963
2964#BEGIN:event
2965
888void 2966void
889wait (OpenCL::Event this) 2967command_queue (OpenCL::Event self)
2968 PPCODE:
2969 cl_command_queue value [1];
2970 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0));
2971 EXTEND (SP, 1);
2972 const int i = 0;
2973 NEED_SUCCESS (RetainCommandQueue, (value [i]));
2974 PUSH_CLOBJ (stash_queue, value [i]);
2975
2976void
2977command_type (OpenCL::Event self)
2978 ALIAS:
2979 command_type = CL_EVENT_COMMAND_TYPE
2980 reference_count = CL_EVENT_REFERENCE_COUNT
2981 command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS
2982 PPCODE:
2983 cl_uint value [1];
2984 NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0));
2985 EXTEND (SP, 1);
2986 const int i = 0;
2987 PUSHs (sv_2mortal (newSVuv (value [i])));
2988
2989void
2990context (OpenCL::Event self)
2991 PPCODE:
2992 cl_context value [1];
2993 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0));
2994 EXTEND (SP, 1);
2995 const int i = 0;
2996 NEED_SUCCESS (RetainContext, (value [i]));
2997 PUSH_CLOBJ (stash_context, value [i]);
2998
2999#END:event
3000
3001void
3002profiling_info (OpenCL::Event self, cl_profiling_info name)
890 CODE: 3003 PPCODE:
891 clWaitForEvents (1, &this); 3004 INFO (EventProfiling)
3005
3006#BEGIN:profiling
3007
3008void
3009profiling_command_queued (OpenCL::Event self)
3010 ALIAS:
3011 profiling_command_queued = CL_PROFILING_COMMAND_QUEUED
3012 profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT
3013 profiling_command_start = CL_PROFILING_COMMAND_START
3014 profiling_command_end = CL_PROFILING_COMMAND_END
3015 PPCODE:
3016 cl_ulong value [1];
3017 NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0));
3018 EXTEND (SP, 1);
3019 const int i = 0;
3020 PUSHs (sv_2mortal (newSVuv (value [i])));
3021
3022#END:profiling
892 3023
893MODULE = OpenCL PACKAGE = OpenCL::UserEvent 3024MODULE = OpenCL PACKAGE = OpenCL::UserEvent
894 3025
895void 3026void
896set_status (OpenCL::UserEvent this, cl_int execution_status) 3027set_status (OpenCL::UserEvent self, cl_int execution_status)
897 CODE: 3028 CODE:
898 clSetUserEventStatus (this, execution_status); 3029 clSetUserEventStatus (self, execution_status);
899 3030
3031MODULE = OpenCL PACKAGE = OpenCL::Mapped
3032
3033void
3034DESTROY (SV *self)
3035 CODE:
3036 OpenCL__Mapped mapped = SvMAPPED (self);
3037
3038 clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0);
3039 mapped_detach (self, mapped);
3040
3041 clReleaseCommandQueue (mapped->queue);
3042 clReleaseEvent (mapped->event);
3043 Safefree (mapped);
3044
3045void
3046unmap (OpenCL::Mapped self, ...)
3047 CODE:
3048 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
3049
3050bool
3051mapped (OpenCL::Mapped self)
3052 CODE:
3053 RETVAL = !!self->ptr;
3054 OUTPUT:
3055 RETVAL
3056
3057void
3058wait (OpenCL::Mapped self)
3059 PPCODE:
3060 if (self->event)
3061 NEED_SUCCESS (WaitForEvents, (1, &self->event));
3062
3063void
3064event (OpenCL::Mapped self)
3065 PPCODE:
3066 if (!self->event)
3067 XSRETURN_UNDEF;
3068
3069 clRetainEvent (self->event);
3070 XPUSH_CLOBJ (stash_event, self->event);
3071
3072#define MAPPED_OFFSET_CB offsetof (struct mapped, cb)
3073#define MAPPED_OFFSET_ROW_PITCH offsetof (struct mapped, row_pitch)
3074#define MAPPED_OFFSET_SLICE_PITCH offsetof (struct mapped, slice_pitch)
3075#define MAPPED_OFFSET_WIDTH offsetof (struct mapped, width)
3076#define MAPPED_OFFSET_HEIGHT offsetof (struct mapped, height)
3077#define MAPPED_OFFSET_DEPTH offsetof (struct mapped, depth)
3078
3079IV
3080size (OpenCL::Mapped self)
3081 ALIAS:
3082 size = MAPPED_OFFSET_CB
3083 row_pitch = MAPPED_OFFSET_ROW_PITCH
3084 slice_pitch = MAPPED_OFFSET_SLICE_PITCH
3085 width = MAPPED_OFFSET_WIDTH
3086 height = MAPPED_OFFSET_HEIGHT
3087 depth = MAPPED_OFFSET_DEPTH
3088 CODE:
3089 RETVAL = *(size_t *)((char *)self + ix);
3090 OUTPUT:
3091 RETVAL
3092
3093IV
3094ptr (OpenCL::Mapped self)
3095 CODE:
3096 RETVAL = PTR2IV (self->ptr);
3097 OUTPUT:
3098 RETVAL
3099
3100void
3101set (OpenCL::Mapped self, size_t offset, SV *data)
3102 CODE:
3103 STRLEN len;
3104 const char *ptr = SvPVbyte (data, len);
3105
3106 if (offset + len > self->cb)
3107 croak ("OpenCL::Mapped::set out of bound condition detected");
3108
3109 memcpy (offset + (char *)self->ptr, ptr, len);
3110
3111void
3112get_row (OpenCL::Mapped self, size_t count, size_t x = 0, size_t y = 0, size_t z = 0)
3113 PPCODE:
3114 if (!SvOK (ST (1)))
3115 count = self->width - x;
3116
3117 if (x + count > self->width)
3118 croak ("OpenCL::Mapped::get: x + count crosses a row boundary");
3119
3120 if (y >= self->height)
3121 croak ("OpenCL::Mapped::get: y coordinate out of bounds");
3122
3123 if (z >= self->depth)
3124 croak ("OpenCL::Mapped::get: z coordinate out of bounds");
3125
3126 size_t element = mapped_element_size (self);
3127
3128 count *= element;
3129 x *= element;
3130
3131 char *ptr = (char *)self->ptr + x + y * self->row_pitch + z * self->slice_pitch;
3132 XPUSHs (sv_2mortal (newSVpvn (ptr, count)));
3133
3134void
3135set_row (OpenCL::Mapped self, SV *data, size_t x = 0, size_t y = 0, size_t z = 0)
3136 PPCODE:
3137 STRLEN count;
3138 char *dataptr = SvPVbyte (data, count);
3139 size_t element = mapped_element_size (self);
3140
3141 x *= element;
3142
3143 if (x + count > self->width * element)
3144 croak ("OpenCL::Mapped::set: x + data size crosses a row boundary");
3145
3146 if (y >= self->height)
3147 croak ("OpenCL::Mapped::set: y coordinate out of bounds");
3148
3149 if (z >= self->depth)
3150 croak ("OpenCL::Mapped::set: z coordinate out of bounds");
3151
3152 char *ptr = (char *)self->ptr + x + y * self->row_pitch + z * self->slice_pitch;
3153 memcpy (ptr, dataptr, count);
3154
3155MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer
3156
3157MODULE = OpenCL PACKAGE = OpenCL::MappedImage
3158
3159IV
3160element_size (OpenCL::Mapped self)
3161 CODE:
3162 RETVAL = mapped_element_size (self);
3163 OUTPUT:
3164 RETVAL
3165

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines