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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.23 by root, Mon Apr 16 09:43:30 2012 UTC vs.
Revision 1.78 by root, Sat May 5 15:11:19 2012 UTC

1#include "EXTERN.h" 1#include "EXTERN.h"
2#include "perl.h" 2#include "perl.h"
3#include "XSUB.h" 3#include "XSUB.h"
4 4
5#include "ecb.h"//D
6
7#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
8#include "xthread.h"
9#include "schmorp.h"
10
11#ifdef I_DLFCN
12 #include <dlfcn.h>
13#endif
14
15// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
16// but then fail to define the api functions unless you ALSO define
17// this. This breaks 100% of the opencl 1.1 apps, for what reason?
18// after all, the functions are deprecated, not removed.
19// in addition, you cannot test for this in any future-proof way.
20// each time a new opencl version comes out, you need to make a new
21// release.
22#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
23
24#ifndef PREFER_1_1
25 #define PREFER_1_1 1
26#endif
27
28#if PREFER_1_1
29 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
30#endif
31
5#ifdef __APPLE__ 32#ifdef __APPLE__
6 #include <OpenCL/opencl.h> 33 #define CLHDR(name) <OpenCL/name>
7#else 34#else
8 #include <CL/opencl.h> 35 #define CLHDR(name) <CL/name>
9#endif 36#endif
37
38#include CLHDR(opencl.h)
39
40#ifndef CL_VERSION_1_2
41 #include CLHDR(cl_d3d9.h)
42#endif
43
44#if _WIN32
45 #include CLHDR(cl_d3d10.h)
46 #if CL_VERSION_1_2
47 #include CLHDR<cl_d3d11.h>
48 #endif
49 #include CLHDR<cl_dx9_media_sharing.h.h>
50#endif
51
52#ifndef CL_VERSION_1_2
53 #undef PREFER_1_1
54 #define PREFER_1_1 1
55#endif
56
57// make sure all constants we might use are actually defined
58#include "default.h"
10 59
11typedef cl_platform_id OpenCL__Platform; 60typedef cl_platform_id OpenCL__Platform;
12typedef cl_device_id OpenCL__Device; 61typedef cl_device_id OpenCL__Device;
62typedef cl_device_id OpenCL__SubDevice;
13typedef cl_context OpenCL__Context; 63typedef cl_context OpenCL__Context;
14typedef cl_command_queue OpenCL__Queue; 64typedef cl_command_queue OpenCL__Queue;
15typedef cl_mem OpenCL__Memory; 65typedef cl_mem OpenCL__Memory;
16typedef cl_mem OpenCL__Buffer; 66typedef cl_mem OpenCL__Buffer;
17typedef cl_mem OpenCL__BufferObj; 67typedef cl_mem OpenCL__BufferObj;
18typedef cl_mem OpenCL__Image; 68typedef cl_mem OpenCL__Image;
19typedef cl_mem OpenCL__Image2D;
20typedef cl_mem OpenCL__Image3D;
21typedef cl_mem OpenCL__Memory_ornull; 69typedef cl_mem OpenCL__Memory_ornull;
22typedef cl_mem OpenCL__Buffer_ornull; 70typedef cl_mem OpenCL__Buffer_ornull;
23typedef cl_mem OpenCL__Image_ornull; 71typedef cl_mem OpenCL__Image_ornull;
24typedef cl_mem OpenCL__Image2D_ornull;
25typedef cl_mem OpenCL__Image3D_ornull;
26typedef cl_sampler OpenCL__Sampler; 72typedef cl_sampler OpenCL__Sampler;
27typedef cl_program OpenCL__Program; 73typedef cl_program OpenCL__Program;
28typedef cl_kernel OpenCL__Kernel; 74typedef cl_kernel OpenCL__Kernel;
29typedef cl_event OpenCL__Event; 75typedef cl_event OpenCL__Event;
30typedef cl_event OpenCL__UserEvent; 76typedef cl_event OpenCL__UserEvent;
31 77
32typedef SV *FUTURE; 78typedef struct mapped * OpenCL__Mapped;
79
80static HV
81 *stash_platform,
82 *stash_device,
83 *stash_subdevice,
84 *stash_context,
85 *stash_queue,
86 *stash_program,
87 *stash_kernel,
88 *stash_sampler,
89 *stash_event,
90 *stash_userevent,
91 *stash_memory,
92 *stash_buffer,
93 *stash_bufferobj,
94 *stash_image,
95 *stash_image1d,
96 *stash_image1darray,
97 *stash_image1dbuffer,
98 *stash_image2d,
99 *stash_image2darray,
100 *stash_image3d,
101 *stash_mapped,
102 *stash_mappedbuffer,
103 *stash_mappedimage;
33 104
34/*****************************************************************************/ 105/*****************************************************************************/
35 106
107// name must include a leading underscore
108// all of this horrors would be unneceesary if somebody wrote a proper OpenGL module
109// for perl. doh.
110static void *
111glsym (const char *name)
112{
113 void *fun = 0;
114
115 #if defined I_DLFCN && defined RTLD_DEFAULT
116 fun = dlsym (RTLD_DEFAULT, name + 1);
117 if (!fun) fun = dlsym (RTLD_DEFAULT, name);
118
119 if (!fun)
120 {
121 static void *libgl;
122 static const char *glso[] = {
123 "libGL.so.1",
124 "libGL.so.3",
125 "libGL.so.4.0",
126 "libGL.so",
127 "/usr/lib/libGL.so",
128 "/usr/X11R6/lib/libGL.1.dylib"
129 };
130 int i;
131
132 for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i)
133 {
134 libgl = dlopen (glso [i], RTLD_LAZY);
135 if (libgl)
136 break;
137 }
138
139 if (libgl)
140 {
141 fun = dlsym (libgl, name + 1);
142 if (!fun) fun = dlsym (libgl, name);
143 }
144 }
145 #endif
146
147 return fun;
148}
149
150/*****************************************************************************/
151
36/* up to two temporary buffers */ 152/* up to two temporary buffers */
37static void * 153static void * ecb_noinline
38tmpbuf (size_t size) 154tmpbuf (size_t size)
39{ 155{
156 enum { buffers = 4 };
40 static int idx; 157 static int idx;
41 static void *buf [2]; 158 static void *buf [buffers];
42 static size_t len [2]; 159 static size_t len [buffers];
43 160
44 idx ^= 1; 161 idx = (idx + 1) % buffers;
45 162
46 if (len [idx] < size) 163 if (len [idx] < size)
47 { 164 {
48 free (buf [idx]); 165 free (buf [idx]);
49 len [idx] = ((size + 31) & ~4095) + 4096 - 32; 166 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
51 } 168 }
52 169
53 return buf [idx]; 170 return buf [idx];
54} 171}
55 172
173static const char * ecb_noinline
174cv_get_name (CV *cv)
175{
176 static char fullname [256];
177
178 GV *gv = CvGV (cv); // gv better be nonzero
179
180 HV *stash = GvSTASH (gv);
181 const char *hvname = HvNAME_get (stash); // stash also better be nonzero
182 const char *gvname = GvNAME (gv);
183
184 snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname);
185 return fullname;
186}
187
56/*****************************************************************************/ 188/*****************************************************************************/
57 189
58typedef struct 190typedef struct
59{ 191{
60 IV iv; 192 IV iv;
61 const char *name; 193 const char *name;
62 #define const_iv(name) { (IV)CL_ ## name, # name }, 194 #define const_iv(name) { (IV)CL_ ## name, # name },
63} ivstr; 195} ivstr;
196
197typedef struct
198{
199 NV nv;
200 const char *name;
201 #define const_nv(name) { (NV)CL_ ## name, # name },
202} nvstr;
64 203
65static const char * 204static const char *
66iv2str (IV value, const ivstr *base, int count, const char *fallback) 205iv2str (IV value, const ivstr *base, int count, const char *fallback)
67{ 206{
68 int i; 207 int i;
117 if (res) \ 256 if (res) \
118 FAIL (name); 257 FAIL (name);
119 258
120/*****************************************************************************/ 259/*****************************************************************************/
121 260
122#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr)
123#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr))
124
125static void * 261static SV *
262new_clobj (HV *stash, IV id)
263{
264 return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash));
265}
266
267#define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id)))
268#define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id)))
269
270/* cl objects are either \$iv, or [$iv, ...] */
271/* they can be upgraded at runtime to the array form */
272static void * ecb_noinline
126SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) 273SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg)
127{ 274{
275 // sv_derived_from is quite slow :(
128 if (SvROK (sv) && sv_derived_from (sv, pkg)) 276 if (SvROK (sv) && sv_derived_from (sv, pkg))
129 return (void *)SvIV (SvRV (sv)); 277 return (void *)SvIV (SvRV (sv));
130 278
131 croak ("%s: %s is not of type %s", func, svname, pkg); 279 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg);
280}
281
282// the "no-inherit" version of the above
283static void * ecb_noinline
284SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash)
285{
286 if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash)
287 return (void *)SvIV (SvRV (sv));
288
289 croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash));
132} 290}
133 291
134/*****************************************************************************/ 292/*****************************************************************************/
293
294static cl_context_properties * ecb_noinline
295SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount)
296{
297 if (!sv || !SvOK (sv))
298 if (extra)
299 sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully
300 else
301 return 0;
302
303 if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV)
304 {
305 AV *av = (AV *)SvRV (sv);
306 int i, len = av_len (av) + 1;
307 cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1));
308 cl_context_properties *l = p;
309
310 if (len & 1)
311 croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname);
312
313 while (extracount--)
314 *l++ = *extra++;
315
316 for (i = 0; i < len; i += 2)
317 {
318 cl_context_properties t = SvIV (*av_fetch (av, i , 0));
319 SV *p_sv = *av_fetch (av, i + 1, 0);
320 cl_context_properties v = SvIV (p_sv); // code below can override
321
322 switch (t)
323 {
324 case CL_CONTEXT_PLATFORM:
325 if (SvROK (p_sv))
326 v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform");
327 break;
328
329 case CL_GLX_DISPLAY_KHR:
330 if (!SvOK (p_sv))
331 {
332 void *func = glsym ("_glXGetCurrentDisplay");
333 if (func)
334 v = (cl_context_properties)((void *(*)(void))func)();
335 }
336 break;
337
338 case CL_GL_CONTEXT_KHR:
339 if (!SvOK (p_sv))
340 {
341 void *func = glsym ("_glXGetCurrentContext");
342 if (func)
343 v = (cl_context_properties)((void *(*)(void))func)();
344 }
345 break;
346
347 default:
348 /* unknown property, treat as int */
349 break;
350 }
351
352 *l++ = t;
353 *l++ = v;
354 }
355
356 *l = 0;
357
358 return p;
359 }
360
361 croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname);
362}
363
364// parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation
365// is a pointer (and only on well-behaved systems).
366static void * ecb_noinline
367object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount)
368{
369 if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV)
370 croak ("%s: '%s' parameter must be %sa reference to an array of %s objects",
371 cv_get_name (cv), argname, or_undef ? "undef or " : "", klass);
372
373 AV *av = (AV *)SvRV (arg);
374 void **list = 0;
375 cl_uint count = av_len (av) + 1;
376
377 if (count)
378 {
379 list = tmpbuf (sizeof (*list) * count);
380 int i;
381 for (i = 0; i < count; ++i)
382 list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass);
383 }
384
385 if (!count && !or_undef)
386 croak ("%s: '%s' must contain at least one %s object",
387 cv_get_name (cv), argname, klass);
388
389 *rcount = count;
390 return (void *)list;
391}
392
393/*****************************************************************************/
394/* callback stuff */
395
396/* default context callback, log to stderr */
397static void CL_CALLBACK
398context_default_notify (const char *msg, const void *info, size_t cb, void *data)
399{
400 fprintf (stderr, "OpenCL Context Notify: %s\n", msg);
401}
402
403typedef struct
404{
405 int free_cb;
406 void (*push)(void *data1, void *data2, void *data3);
407} eq_vtbl;
408
409typedef struct eq_item
410{
411 struct eq_item *next;
412 eq_vtbl *vtbl;
413 SV *cb;
414 void *data1, *data2, *data3;
415} eq_item;
416
417static void (*eq_signal_func)(void *signal_arg, int value);
418static void *eq_signal_arg;
419static xmutex_t eq_lock = X_MUTEX_INIT;
420static eq_item *eq_head, *eq_tail;
421
422static void ecb_noinline
423eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3)
424{
425 eq_item *item = malloc (sizeof (eq_item));
426
427 item->next = 0;
428 item->vtbl = vtbl;
429 item->cb = cb;
430 item->data1 = data1;
431 item->data2 = data2;
432 item->data3 = data3;
433
434 X_LOCK (eq_lock);
435
436 *(eq_head ? &eq_tail->next : &eq_head) = item;
437 eq_tail = item;
438
439 X_UNLOCK (eq_lock);
440
441 eq_signal_func (eq_signal_arg, 0);
442}
443
444static eq_item *
445eq_dec (void)
446{
447 eq_item *res;
448
449 X_LOCK (eq_lock);
450
451 res = eq_head;
452 if (res)
453 eq_head = res->next;
454
455 X_UNLOCK (eq_lock);
456
457 return res;
458}
459
460static void
461eq_poll (void)
462{
463 eq_item *item;
464
465 while ((item = eq_dec ()))
466 {
467 ENTER;
468 SAVETMPS;
469
470 dSP;
471 PUSHMARK (SP);
472 EXTEND (SP, 2);
473
474 if (item->vtbl->free_cb)
475 sv_2mortal (item->cb);
476
477 PUTBACK;
478 item->vtbl->push (item->data1, item->data2, item->data3);
479
480 SV *cb = item->cb;
481 free (item);
482
483 call_sv (cb, G_DISCARD | G_VOID);
484
485 FREETMPS;
486 LEAVE;
487 }
488}
489
490static void
491eq_poll_interrupt (pTHX_ void *c_arg, int value)
492{
493 eq_poll ();
494}
495
496/*****************************************************************************/
497/* context notify */
498
499static void ecb_noinline
500eq_context_push (void *data1, void *data2, void *data3)
501{
502 dSP;
503 PUSHs (sv_2mortal (newSVpv (data1, 0)));
504 PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3)));
505 PUTBACK;
506
507 free (data1);
508 free (data2);
509}
510
511static eq_vtbl eq_context_vtbl = { 0, eq_context_push };
512
513static void CL_CALLBACK
514eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data)
515{
516 void *pvt_copy = malloc (cb);
517 memcpy (pvt_copy, pvt, cb);
518 eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb);
519}
520
521#define CONTEXT_NOTIFY_CALLBACK \
522 void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \
523 void *user_data = 0; \
524 \
525 if (SvOK (notify)) \
526 { \
527 pfn_notify = eq_context_notify; \
528 user_data = s_get_cv (notify); \
529 }
530
531static SV * ecb_noinline
532new_clobj_context (cl_context ctx, void *user_data)
533{
534 SV *sv = new_clobj (stash_context, (IV)ctx);
535
536 if (user_data)
537 sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0);
538
539 return sv;
540}
541
542#define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data));
543
544/*****************************************************************************/
545/* build/compile/link notify */
546
547static void
548eq_program_push (void *data1, void *data2, void *data3)
549{
550 dSP;
551 PUSH_CLOBJ (stash_program, data1);
552 PUTBACK;
553}
554
555static eq_vtbl eq_program_vtbl = { 1, eq_program_push };
556
557static void CL_CALLBACK
558eq_program_notify (cl_program program, void *user_data)
559{
560 clRetainProgram (program);
561
562 eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0);
563}
564
565typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data);
566
567static program_callback ecb_noinline
568make_program_callback (SV *notify, void **ruser_data)
569{
570 if (SvOK (notify))
571 {
572 *ruser_data = SvREFCNT_inc (s_get_cv (notify));
573 return eq_program_notify;
574 }
575 else
576 {
577 *ruser_data = 0;
578 return 0;
579 }
580}
581
582struct build_args
583{
584 cl_program program;
585 char *options;
586 void *user_data;
587 cl_uint num_devices;
588};
589
590X_THREAD_PROC (build_program_thread)
591{
592 struct build_args *arg = thr_arg;
593
594 clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0);
595
596 if (arg->user_data)
597 eq_program_notify (arg->program, arg->user_data);
598 else
599 clReleaseProgram (arg->program);
600
601 free (arg->options);
602 free (arg);
603
604 return 0;
605}
606
607static void
608build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data)
609{
610 struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices);
611
612 arg->program = program;
613 arg->options = strdup (options);
614 arg->user_data = user_data;
615 arg->num_devices = num_devices;
616 memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices);
617
618 xthread_t id;
619 thread_create (&id, build_program_thread, arg);
620}
621
622/*****************************************************************************/
623/* mem object destructor notify */
624
625static void ecb_noinline
626eq_destructor_push (void *data1, void *data2, void *data3)
627{
628}
629
630static eq_vtbl eq_destructor_vtbl = { 0, eq_destructor_push };
631
632static void CL_CALLBACK
633eq_destructor_notify (cl_mem memobj, void *user_data)
634{
635 eq_enq (&eq_destructor_vtbl, (SV *)user_data, (void *)memobj, 0, 0);
636}
637
638/*****************************************************************************/
639/* event objects */
640
641static void
642eq_event_push (void *data1, void *data2, void *data3)
643{
644 dSP;
645 PUSH_CLOBJ (stash_event, data1);
646 PUSHs (sv_2mortal (newSViv ((IV)data2)));
647 PUTBACK;
648}
649
650static eq_vtbl eq_event_vtbl = { 1, eq_event_push };
651
652static void CL_CALLBACK
653eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data)
654{
655 clRetainEvent (event);
656 eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0);
657}
658
659/*****************************************************************************/
660/* utilities for XS code */
135 661
136static size_t 662static size_t
137img_row_pitch (cl_mem img) 663img_row_pitch (cl_mem img)
138{ 664{
139 size_t res; 665 size_t res;
140 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); 666 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
141 return res; 667 return res;
142} 668}
143 669
144static cl_event * 670static cl_event * ecb_noinline
145event_list (SV **items, int count) 671event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
146{ 672{
147 if (!count) 673 cl_uint count = *rcount;
674
675 if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed
676 *rcount = count = 0;
677
678 if (!count && !extra)
148 return 0; 679 return 0;
149 680
150 cl_event *list = tmpbuf (sizeof (cl_event) * count); 681 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
682 int i = 0;
151 683
152 while (count--) 684 while (count--)
685 if (SvOK (items [count]))
153 list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); 686 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
154 687
688 if (extra)
689 list [i++] = extra;
690
691 *rcount = i;
692
155 return list; 693 return i ? list : 0;
156} 694}
157 695
158#define EVENT_LIST(items,count) \ 696#define EVENT_LIST(skip) \
159 cl_uint event_list_count = (count); \ 697 cl_uint event_list_count = items - (skip); \
160 cl_event *event_list_ptr = event_list (&ST (items), event_list_count) 698 cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0)
161 699
162#define INFO(class) \ 700#define INFO(class) \
163{ \ 701{ \
164 size_t size; \ 702 size_t size; \
165 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ 703 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
169 SvCUR_set (sv, size); \ 707 SvCUR_set (sv, size); \
170 NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ 708 NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \
171 XPUSHs (sv); \ 709 XPUSHs (sv); \
172} 710}
173 711
712/*****************************************************************************/
713/* mapped_xxx */
714
715static OpenCL__Mapped
716SvMAPPED (SV *sv)
717{
718 // no typechecking atm., keep your fingers crossed
719 return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr;
720}
721
722struct mapped
723{
724 cl_command_queue queue;
725 cl_mem memobj;
726 void *ptr;
727 size_t cb;
728 cl_event event;
729 size_t row_pitch;
730 size_t slice_pitch;
731};
732
733static SV *
734mapped_new (HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags, void *ptr, size_t cb, cl_event ev, size_t row_pitch, size_t slice_pitch)
735{
736 SV *data = newSV (0);
737 SvUPGRADE (data, SVt_PVMG);
738
739 OpenCL__Mapped mapped;
740 New (0, mapped, 1, struct mapped);
741
742 clRetainCommandQueue (queue);
743
744 mapped->queue = queue;
745 mapped->memobj = memobj;
746 mapped->ptr = ptr;
747 mapped->cb = cb;
748 mapped->event = ev;
749 mapped->row_pitch = row_pitch;
750 mapped->slice_pitch = slice_pitch;
751
752 sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0);
753
754 if (SvLEN (data))
755 Safefree (data);
756
757 SvPVX (data) = (char *)ptr;
758 SvCUR_set (data, cb);
759 SvLEN_set (data, 0);
760 SvPOK_only (data);
761
762 SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash));
763
764 if (!(flags & CL_MAP_WRITE))
765 SvREADONLY_on (data);
766
767 return obj;
768}
769
770static void
771mapped_detach (SV *sv, OpenCL__Mapped mapped)
772{
773 SV *data = SvRV (sv);
774
775 // the next check checks both before AND after detach, where SvPVX should STILL be 0
776 if (SvPVX (data) != (char *)mapped->ptr)
777 warn ("FATAL: OpenCL memory mapped scalar changed location, detected");
778 else
779 {
780 SvREADONLY_off (data);
781 SvCUR_set (data, 0);
782 SvPVX (data) = 0;
783 SvOK_off (data);
784 }
785
786 mapped->ptr = 0;
787}
788
789static void
790mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count)
791{
792 cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event);
793 cl_event ev;
794
795 NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev));
796
797 clReleaseEvent (mapped->event);
798 mapped->event = ev;
799
800 mapped_detach (self, mapped);
801}
802
803/*****************************************************************************/
804
174MODULE = OpenCL PACKAGE = OpenCL 805MODULE = OpenCL PACKAGE = OpenCL
175 806
176PROTOTYPES: ENABLE 807PROTOTYPES: ENABLE
177 808
809void
810poll ()
811 CODE:
812 eq_poll ();
813
814void
815_eq_initialise (IV func, IV arg)
816 CODE:
817 eq_signal_func = (void (*)(void *, int))func;
818 eq_signal_arg = (void*)arg;
819
178BOOT: 820BOOT:
179{ 821{
180 HV *stash = gv_stashpv ("OpenCL", 1); 822 HV *stash = gv_stashpv ("OpenCL", 1);
823
181 static const ivstr *civ, const_iv[] = { 824 static const ivstr *civ, const_iv[] = {
182 { sizeof (cl_char ), "SIZEOF_CHAR" }, 825 { sizeof (cl_char ), "SIZEOF_CHAR" },
183 { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, 826 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
184 { sizeof (cl_short ), "SIZEOF_SHORT" }, 827 { sizeof (cl_short ), "SIZEOF_SHORT" },
185 { sizeof (cl_ushort), "SIZEOF_USHORT" }, 828 { sizeof (cl_ushort), "SIZEOF_USHORT" },
186 { sizeof (cl_int ), "SIZEOF_INT" }, 829 { sizeof (cl_int ), "SIZEOF_INT" },
187 { sizeof (cl_uint ), "SIZEOF_UINT" }, 830 { sizeof (cl_uint ), "SIZEOF_UINT" },
188 { sizeof (cl_long ), "SIZEOF_LONG" }, 831 { sizeof (cl_long ), "SIZEOF_LONG" },
189 { sizeof (cl_ulong ), "SIZEOF_ULONG" }, 832 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
190 { sizeof (cl_half ), "SIZEOF_HALF" }, 833 { sizeof (cl_half ), "SIZEOF_HALF" },
191 { sizeof (cl_float ), "SIZEOF_FLOAT" }, 834 { sizeof (cl_float ), "SIZEOF_FLOAT" },
192 { sizeof (cl_double), "SIZEOF_DOUBLE" }, 835 { sizeof (cl_double), "SIZEOF_DOUBLE" },
836 { PREFER_1_1 , "PREFER_1_1" },
193#include "constiv.h" 837#include "constiv.h"
194 }; 838 };
839
195 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) 840 for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--)
196 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); 841 newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv));
842
843 static const nvstr *cnv, const_nv[] = {
844#include "constnv.h"
845 };
846
847 for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--)
848 newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv));
849
850 newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call
851
852 stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD);
853 stash_device = gv_stashpv ("OpenCL::Device", GV_ADD);
854 stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD);
855 stash_context = gv_stashpv ("OpenCL::Context", GV_ADD);
856 stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD);
857 stash_program = gv_stashpv ("OpenCL::Program", GV_ADD);
858 stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD);
859 stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD);
860 stash_event = gv_stashpv ("OpenCL::Event", GV_ADD);
861 stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD);
862 stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD);
863 stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD);
864 stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD);
865 stash_image = gv_stashpv ("OpenCL::Image", GV_ADD);
866 stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD);
867 stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD);
868 stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD);
869 stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD);
870 stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD);
871 stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD);
872 stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD);
873 stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD);
874 stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD);
875
876 sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt);
197} 877}
198 878
199cl_int 879cl_int
200errno () 880errno ()
201 CODE: 881 CODE:
202 errno = res; 882 RETVAL = res;
883 OUTPUT:
884 RETVAL
203 885
204const char * 886const char *
205err2str (cl_int err) 887err2str (cl_int err = res)
206 888
207const char * 889const char *
208enum2str (cl_uint value) 890enum2str (cl_uint value)
209 891
210void 892void
218 list = tmpbuf (sizeof (*list) * count); 900 list = tmpbuf (sizeof (*list) * count);
219 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 901 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
220 902
221 EXTEND (SP, count); 903 EXTEND (SP, count);
222 for (i = 0; i < count; ++i) 904 for (i = 0; i < count; ++i)
223 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 905 PUSH_CLOBJ (stash_platform, list [i]);
224 906
225void 907void
226context_from_type (FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 908context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
227 PPCODE: 909 PPCODE:
910 CONTEXT_NOTIFY_CALLBACK;
228 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (0, type, 0, 0, &res)); 911 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
229 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 912 XPUSH_CLOBJ_CONTEXT;
230 913
231void 914void
232context (FUTURE properties, FUTURE devices, FUTURE notify = 0) 915context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
233 PPCODE: 916 PPCODE:
234 /* der Gipfel der Kunst */ 917 cl_uint device_count;
918 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
919
920 CONTEXT_NOTIFY_CALLBACK;
921 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res));
922 XPUSH_CLOBJ_CONTEXT;
235 923
236void 924void
237wait_for_events (...) 925wait_for_events (...)
238 CODE: 926 CODE:
239 EVENT_LIST (0, items); 927 EVENT_LIST (0);
240 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); 928 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr));
241 929
242PROTOTYPES: DISABLE 930PROTOTYPES: DISABLE
243 931
244MODULE = OpenCL PACKAGE = OpenCL::Platform 932MODULE = OpenCL PACKAGE = OpenCL::Platform
245 933
246void 934void
247info (OpenCL::Platform self, cl_platform_info name) 935info (OpenCL::Platform self, cl_platform_info name)
248 PPCODE: 936 PPCODE:
249 INFO (Platform) 937 INFO (Platform)
938
939void
940unload_compiler (OpenCL::Platform self)
941 CODE:
942#if CL_VERSION_1_2
943 clUnloadPlatformCompiler (self);
944#endif
250 945
251#BEGIN:platform 946#BEGIN:platform
252 947
253void 948void
254profile (OpenCL::Platform self) 949profile (OpenCL::Platform self)
280 list = tmpbuf (sizeof (*list) * count); 975 list = tmpbuf (sizeof (*list) * count);
281 NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); 976 NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0));
282 977
283 EXTEND (SP, count); 978 EXTEND (SP, count);
284 for (i = 0; i < count; ++i) 979 for (i = 0; i < count; ++i)
285 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 980 PUSH_CLOBJ (stash_device, list [i]);
286 981
287void 982void
288context (OpenCL::Platform self, FUTURE properties, SV *devices, FUTURE notify = 0) 983context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
289 PPCODE: 984 PPCODE:
985 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
986 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
987
988 cl_uint device_count;
989 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
990
991 CONTEXT_NOTIFY_CALLBACK;
992 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
993 XPUSH_CLOBJ_CONTEXT;
994
995void
996context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
997 PPCODE:
998 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
999 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
1000
1001 CONTEXT_NOTIFY_CALLBACK;
1002 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res));
1003 XPUSH_CLOBJ_CONTEXT;
1004
1005MODULE = OpenCL PACKAGE = OpenCL::Device
1006
1007void
1008info (OpenCL::Device self, cl_device_info name)
1009 PPCODE:
1010 INFO (Device)
1011
1012#if CL_VERSION_1_2
1013
1014void
1015sub_devices (OpenCL::Device self, SV *properties)
1016 PPCODE:
290 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) 1017 if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV)
291 croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); 1018 croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs");
292 1019
293 AV *av = (AV *)SvRV (devices); 1020 properties = SvRV (properties);
294 cl_uint num_devices = av_len (av) + 1; 1021
295 cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); 1022 cl_uint count = av_len ((AV *)properties) + 1;
1023 cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1);
1024
296 int i; 1025 int i;
1026 for (i = 0; i < count; ++i)
1027 props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0));
297 1028
298 for (i = num_devices; i--; ) 1029 props [count] = 0;
299 device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device");
300 1030
301 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (0, num_devices, device_list, 0, 0, &res)); 1031 cl_uint num_devices;
302 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 1032 NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices));
1033 cl_device_id *list = tmpbuf (sizeof (*list) * num_devices);
1034 NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0));
303 1035
304void 1036 EXTEND (SP, num_devices);
305context_from_type (OpenCL::Platform self, FUTURE properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) 1037 for (i = 0; i < count; ++i)
306 PPCODE: 1038 PUSH_CLOBJ (stash_subdevice, list [i]);
307 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self, 0 };
308 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res));
309 XPUSH_NEW_OBJ ("OpenCL::Context", ctx);
310 1039
311MODULE = OpenCL PACKAGE = OpenCL::Device 1040#endif
312
313void
314info (OpenCL::Device self, cl_device_info name)
315 PPCODE:
316 INFO (Device)
317 1041
318#BEGIN:device 1042#BEGIN:device
319 1043
320void 1044void
321type (OpenCL::Device self) 1045type (OpenCL::Device self)
322 PPCODE: 1046 PPCODE:
323 cl_device_type value [1]; 1047 cl_device_type value [1];
324 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); 1048 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0));
325 EXTEND (SP, 1); 1049 EXTEND (SP, 1);
326 const int i = 0; 1050 const int i = 0;
327 PUSHs (sv_2mortal (newSViv (value [i]))); 1051 PUSHs (sv_2mortal (newSVuv (value [i])));
328 1052
329void 1053void
330vendor_id (OpenCL::Device self) 1054vendor_id (OpenCL::Device self)
331 ALIAS: 1055 ALIAS:
332 vendor_id = CL_DEVICE_VENDOR_ID 1056 vendor_id = CL_DEVICE_VENDOR_ID
353 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT 1077 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
354 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG 1078 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
355 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT 1079 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
356 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE 1080 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
357 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF 1081 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
358 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT 1082 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
359 PPCODE: 1083 PPCODE:
360 cl_uint value [1]; 1084 cl_uint value [1];
361 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); 1085 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
362 EXTEND (SP, 1); 1086 EXTEND (SP, 1);
363 const int i = 0; 1087 const int i = 0;
477 PPCODE: 1201 PPCODE:
478 cl_command_queue_properties value [1]; 1202 cl_command_queue_properties value [1];
479 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); 1203 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0));
480 EXTEND (SP, 1); 1204 EXTEND (SP, 1);
481 const int i = 0; 1205 const int i = 0;
482 PUSHs (sv_2mortal (newSViv (value [i]))); 1206 PUSHs (sv_2mortal (newSVuv (value [i])));
483 1207
484void 1208void
485platform (OpenCL::Device self) 1209platform (OpenCL::Device self)
486 PPCODE: 1210 PPCODE:
487 cl_platform_id value [1]; 1211 cl_platform_id value [1];
488 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); 1212 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0));
489 EXTEND (SP, 1); 1213 EXTEND (SP, 1);
490 const int i = 0; 1214 const int i = 0;
491 { 1215 {
492 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); 1216 PUSH_CLOBJ (stash_platform, value [i]);
493 } 1217 }
494 1218
495void 1219void
496name (OpenCL::Device self) 1220name (OpenCL::Device self)
497 ALIAS: 1221 ALIAS:
516 cl_device_id value [1]; 1240 cl_device_id value [1];
517 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); 1241 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0));
518 EXTEND (SP, 1); 1242 EXTEND (SP, 1);
519 const int i = 0; 1243 const int i = 0;
520 { 1244 {
521 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); 1245 PUSH_CLOBJ (stash_device, value [i]);
522 } 1246 }
523 1247
524void 1248void
525partition_types_ext (OpenCL::Device self) 1249partition_types_ext (OpenCL::Device self)
526 ALIAS: 1250 ALIAS:
537 for (i = 0; i < n; ++i) 1261 for (i = 0; i < n; ++i)
538 PUSHs (sv_2mortal (newSVuv (value [i]))); 1262 PUSHs (sv_2mortal (newSVuv (value [i])));
539 1263
540#END:device 1264#END:device
541 1265
1266MODULE = OpenCL PACKAGE = OpenCL::SubDevice
1267
1268#if CL_VERSION_1_2
1269
1270void
1271DESTROY (OpenCL::SubDevice self)
1272 CODE:
1273 clReleaseDevice (self);
1274
1275#endif
1276
542MODULE = OpenCL PACKAGE = OpenCL::Context 1277MODULE = OpenCL PACKAGE = OpenCL::Context
543 1278
544void 1279void
545DESTROY (OpenCL::Context context) 1280DESTROY (OpenCL::Context self)
546 CODE: 1281 CODE:
547 clReleaseContext (context); 1282 clReleaseContext (self);
548 1283
549void 1284void
550info (OpenCL::Context self, cl_context_info name) 1285info (OpenCL::Context self, cl_context_info name)
551 PPCODE: 1286 PPCODE:
552 INFO (Context) 1287 INFO (Context)
553 1288
554void 1289void
555queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) 1290queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0)
556 PPCODE: 1291 PPCODE:
557 NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); 1292 NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res));
558 XPUSH_NEW_OBJ ("OpenCL::Queue", queue); 1293 XPUSH_CLOBJ (stash_queue, queue);
559 1294
560void 1295void
561user_event (OpenCL::Context self) 1296user_event (OpenCL::Context self)
562 PPCODE: 1297 PPCODE:
563 NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); 1298 NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res));
564 XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); 1299 XPUSH_CLOBJ (stash_userevent, ev);
565 1300
566void 1301void
567buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) 1302buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
568 PPCODE: 1303 PPCODE:
569 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) 1304 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))
570 croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); 1305 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
571 1306
572 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); 1307 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
573 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 1308 XPUSH_CLOBJ (stash_bufferobj, mem);
574 1309
575void 1310void
576buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) 1311buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
577 PPCODE: 1312 PPCODE:
578 STRLEN len; 1313 STRLEN len;
579 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 1314 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
580 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) 1315 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
581 croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); 1316 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
582 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); 1317 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
583 XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); 1318 XPUSH_CLOBJ (stash_bufferobj, mem);
584 1319
1320#if CL_VERSION_1_2
1321
585void 1322void
586image2d (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) 1323image (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)
587 PPCODE: 1324 PPCODE:
588 STRLEN len; 1325 STRLEN len;
589 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 1326 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
590 const cl_image_format format = { channel_order, channel_type }; 1327 const cl_image_format format = { channel_order, channel_type };
1328 const cl_image_desc desc = {
1329 type,
1330 width, height, depth,
1331 array_size, row_pitch, slice_pitch,
1332 num_mip_level, num_samples,
1333 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0
1334 };
591 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); 1335 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
592 XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); 1336 HV *stash = stash_image;
1337 switch (type)
1338 {
1339 case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break;
1340 case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break;
1341 case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break;
1342 case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break;
1343 case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break;
1344 case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break;
1345 }
1346 XPUSH_CLOBJ (stash, mem);
593 1347
1348#endif
1349
594void 1350void
595image3d (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) 1351image2d (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)
596 PPCODE: 1352 PPCODE:
597 STRLEN len; 1353 STRLEN len;
598 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; 1354 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
599 const cl_image_format format = { channel_order, channel_type }; 1355 const cl_image_format format = { channel_order, channel_type };
1356#if PREFER_1_1
1357 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
1358#else
1359 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
1360 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1361#endif
1362 XPUSH_CLOBJ (stash_image2d, mem);
1363
1364void
1365image3d (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)
1366 PPCODE:
1367 STRLEN len;
1368 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1369 const cl_image_format format = { channel_order, channel_type };
1370#if PREFER_1_1
600 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); 1371 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
601 XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); 1372#else
1373 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
1374 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1375#endif
1376 XPUSH_CLOBJ (stash_image3d, mem);
1377
1378#if cl_apple_gl_sharing || cl_khr_gl_sharing
1379
1380void
1381gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
1382 PPCODE:
1383 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
1384 XPUSH_CLOBJ (stash_bufferobj, mem);
1385
1386void
1387gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
1388 PPCODE:
1389 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
1390 XPUSH_CLOBJ (stash_image2d, mem);
1391
1392#if CL_VERSION_1_2
1393
1394void
1395gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1396 ALIAS:
1397 PPCODE:
1398 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
1399 cl_gl_object_type type;
1400 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
1401 HV *stash = stash_memory;
1402 switch (type)
1403 {
1404 case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break;
1405 case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break;
1406 case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break;
1407 case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break;
1408 case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break;
1409 case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break;
1410 }
1411 XPUSH_CLOBJ (stash, mem);
1412
1413#endif
1414
1415void
1416gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1417 PPCODE:
1418#if PREFER_1_1
1419 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
1420#else
1421 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1422#endif
1423 XPUSH_CLOBJ (stash_image2d, mem);
1424
1425void
1426gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1427 PPCODE:
1428#if PREFER_1_1
1429 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
1430#else
1431 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1432#endif
1433 XPUSH_CLOBJ (stash_image3d, mem);
1434
1435#endif
602 1436
603void 1437void
604supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) 1438supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
605 PPCODE: 1439 PPCODE:
606{ 1440{
624 1458
625void 1459void
626sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) 1460sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode)
627 PPCODE: 1461 PPCODE:
628 NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); 1462 NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res));
629 XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); 1463 XPUSH_CLOBJ (stash_sampler, sampler);
630 1464
631void 1465void
632program_with_source (OpenCL::Context self, SV *program) 1466program_with_source (OpenCL::Context self, SV *program)
633 PPCODE: 1467 PPCODE:
634 STRLEN len; 1468 STRLEN len;
635 size_t len2; 1469 size_t len2;
636 const char *ptr = SvPVbyte (program, len); 1470 const char *ptr = SvPVbyte (program, len);
637 1471
638 len2 = len; 1472 len2 = len;
639 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); 1473 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res));
640 XPUSH_NEW_OBJ ("OpenCL::Program", prog); 1474 XPUSH_CLOBJ (stash_program, prog);
1475
1476void
1477program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1478 PPCODE:
1479 cl_uint device_count;
1480 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1481
1482 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1483 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1484
1485 binaries = SvRV (binaries);
1486
1487 if (device_count != av_len ((AV *)binaries) + 1)
1488 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1489
1490 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1491 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1492 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1493
1494 int i;
1495 for (i = 0; i < device_count; ++i)
1496 {
1497 STRLEN len;
1498 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1499 length_list [i] = len;
1500 }
1501
1502 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list,
1503 length_list, binary_list,
1504 GIMME_V == G_ARRAY ? status_list : 0, &res));
1505
1506 EXTEND (SP, 2);
1507 PUSH_CLOBJ (stash_program, prog);
1508
1509 if (GIMME_V == G_ARRAY)
1510 {
1511 AV *av = newAV ();
1512 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1513
1514 for (i = device_count; i--; )
1515 av_store (av, i, newSViv (status_list [i]));
1516 }
1517
1518#if CL_VERSION_1_2
1519
1520void
1521program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names)
1522 PPCODE:
1523 cl_uint device_count;
1524 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1525
1526 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res));
1527
1528 XPUSH_CLOBJ (stash_program, prog);
1529
1530void
1531link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef)
1532 CODE:
1533 cl_uint device_count = 0;
1534 cl_device_id *device_list = 0;
1535
1536 if (SvOK (devices))
1537 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
1538
1539 cl_uint program_count;
1540 cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count);
1541
1542 void *user_data;
1543 program_callback pfn_notify = make_program_callback (notify, &user_data);
1544
1545 NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
1546 program_count, program_list, pfn_notify, user_data, &res));
1547
1548 XPUSH_CLOBJ (stash_program, prog);
1549
1550#endif
641 1551
642#BEGIN:context 1552#BEGIN:context
643 1553
644void 1554void
645reference_count (OpenCL::Context self) 1555reference_count (OpenCL::Context self)
662 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); 1572 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0));
663 int i, n = size / sizeof (*value); 1573 int i, n = size / sizeof (*value);
664 EXTEND (SP, n); 1574 EXTEND (SP, n);
665 for (i = 0; i < n; ++i) 1575 for (i = 0; i < n; ++i)
666 { 1576 {
667 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); 1577 PUSH_CLOBJ (stash_device, value [i]);
668 } 1578 }
669 1579
670void 1580void
671properties (OpenCL::Context self) 1581properties (OpenCL::Context self)
672 PPCODE: 1582 PPCODE:
687DESTROY (OpenCL::Queue self) 1597DESTROY (OpenCL::Queue self)
688 CODE: 1598 CODE:
689 clReleaseCommandQueue (self); 1599 clReleaseCommandQueue (self);
690 1600
691void 1601void
692enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) 1602read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1603 ALIAS:
1604 enqueue_read_buffer = 0
693 PPCODE: 1605 PPCODE:
694 cl_event ev = 0;
695 EVENT_LIST (6, items - 6); 1606 EVENT_LIST (6);
696 1607
697 SvUPGRADE (data, SVt_PV); 1608 SvUPGRADE (data, SVt_PV);
698 SvGROW (data, len); 1609 SvGROW (data, len);
699 SvPOK_only (data); 1610 SvPOK_only (data);
700 SvCUR_set (data, len); 1611 SvCUR_set (data, len);
1612
1613 cl_event ev = 0;
701 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1614 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
702 1615
703 if (ev) 1616 if (ev)
704 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1617 XPUSH_CLOBJ (stash_event, ev);
705 1618
706void 1619void
707enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) 1620write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1621 ALIAS:
1622 enqueue_write_buffer = 0
708 PPCODE: 1623 PPCODE:
709 cl_event ev = 0; 1624 EVENT_LIST (5);
1625
710 STRLEN len; 1626 STRLEN len;
711 char *ptr = SvPVbyte (data, len); 1627 char *ptr = SvPVbyte (data, len);
712 EVENT_LIST (5, items - 5);
713 1628
1629 cl_event ev = 0;
714 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1630 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
715 1631
716 if (ev) 1632 if (ev)
717 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1633 XPUSH_CLOBJ (stash_event, ev);
718 1634
1635#if CL_VERSION_1_2
1636
719void 1637void
720enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1638fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1639 ALIAS:
1640 enqueue_fill_buffer = 0
721 PPCODE: 1641 PPCODE:
1642 EVENT_LIST (5);
1643
1644 STRLEN len;
1645 char *ptr = SvPVbyte (data, len);
1646
722 cl_event ev = 0; 1647 cl_event ev = 0;
1648 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1649
1650 if (ev)
1651 XPUSH_CLOBJ (stash_event, ev);
1652
1653void
1654fill_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, ...)
1655 ALIAS:
1656 enqueue_fill_image = 0
1657 PPCODE:
1658 EVENT_LIST (12);
1659
1660 const size_t origin [3] = { x, y, z };
1661 const size_t region [3] = { width, height, depth };
1662
1663 const cl_float c_f [4] = { r, g, b, a };
1664 const cl_uint c_u [4] = { r, g, b, a };
1665 const cl_int c_s [4] = { r, g, b, a };
1666 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1667 static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 };
1668 cl_image_format format;
1669 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1670 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1671 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1672 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1673
1674 cl_event ev = 0;
1675 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1676 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1677
1678 if (ev)
1679 XPUSH_CLOBJ (stash_event, ev);
1680
1681#endif
1682
1683void
1684copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1685 ALIAS:
1686 enqueue_copy_buffer = 0
1687 PPCODE:
723 EVENT_LIST (6, items - 6); 1688 EVENT_LIST (6);
724 1689
1690 cl_event ev = 0;
725 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1691 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
726 1692
727 if (ev) 1693 if (ev)
728 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1694 XPUSH_CLOBJ (stash_event, ev);
729 1695
730void 1696void
731enqueue_read_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, ...) 1697read_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, ...)
1698 ALIAS:
1699 enqueue_read_buffer_rect = 0
732 PPCODE: 1700 PPCODE:
733 cl_event ev = 0; 1701 EVENT_LIST (17);
1702
734 const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; 1703 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
735 const size_t host_origin[3] = { host_x, host_y, host_z }; 1704 const size_t host_origin[3] = { host_x, host_y, host_z };
736 const size_t region[3] = { width, height, depth }; 1705 const size_t region[3] = { width, height, depth };
737 EVENT_LIST (17, items - 17);
738 1706
739 if (!buf_row_pitch) 1707 if (!buf_row_pitch)
740 buf_row_pitch = region [0]; 1708 buf_row_pitch = region [0];
741 1709
742 if (!buf_slice_pitch) 1710 if (!buf_slice_pitch)
752 1720
753 SvUPGRADE (data, SVt_PV); 1721 SvUPGRADE (data, SVt_PV);
754 SvGROW (data, len); 1722 SvGROW (data, len);
755 SvPOK_only (data); 1723 SvPOK_only (data);
756 SvCUR_set (data, len); 1724 SvCUR_set (data, len);
1725
1726 cl_event ev = 0;
757 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)); 1727 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));
758 1728
759 if (ev) 1729 if (ev)
760 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1730 XPUSH_CLOBJ (stash_event, ev);
761 1731
762void 1732void
763enqueue_write_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, ...) 1733write_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, ...)
1734 ALIAS:
1735 enqueue_write_buffer_rect = 0
764 PPCODE: 1736 PPCODE:
765 cl_event ev = 0; 1737 EVENT_LIST (17);
1738
766 const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; 1739 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
767 const size_t host_origin[3] = { host_x, host_y, host_z }; 1740 const size_t host_origin[3] = { host_x, host_y, host_z };
768 const size_t region[3] = { width, height, depth }; 1741 const size_t region[3] = { width, height, depth };
769 STRLEN len; 1742 STRLEN len;
770 char *ptr = SvPVbyte (data, len); 1743 char *ptr = SvPVbyte (data, len);
771 EVENT_LIST (17, items - 17);
772 1744
773 if (!buf_row_pitch) 1745 if (!buf_row_pitch)
774 buf_row_pitch = region [0]; 1746 buf_row_pitch = region [0];
775 1747
776 if (!buf_slice_pitch) 1748 if (!buf_slice_pitch)
785 size_t min_len = host_row_pitch * host_slice_pitch * region [2]; 1757 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
786 1758
787 if (len < min_len) 1759 if (len < min_len)
788 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1760 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
789 1761
1762 cl_event ev = 0;
790 NEED_SUCCESS (EnqueueWriteBufferRect, (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)); 1763 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));
791 1764
792 if (ev) 1765 if (ev)
793 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1766 XPUSH_CLOBJ (stash_event, ev);
794 1767
795void 1768void
796enqueue_copy_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, ...) 1769copy_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, ...)
1770 ALIAS:
1771 enqueue_copy_buffer_rect = 0
797 PPCODE: 1772 PPCODE:
798 cl_event ev = 0; 1773 EVENT_LIST (16);
1774
799 const size_t src_origin[3] = { src_x, src_y, src_z }; 1775 const size_t src_origin[3] = { src_x, src_y, src_z };
800 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1776 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
801 const size_t region[3] = { width, height, depth }; 1777 const size_t region[3] = { width, height, depth };
802 EVENT_LIST (16, items - 16);
803 1778
1779 cl_event ev = 0;
804 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)); 1780 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));
805 1781
806 if (ev) 1782 if (ev)
807 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1783 XPUSH_CLOBJ (stash_event, ev);
808 1784
809void 1785void
810enqueue_read_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, ...) 1786read_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, ...)
1787 ALIAS:
1788 enqueue_read_image = 0
811 PPCODE: 1789 PPCODE:
812 cl_event ev = 0; 1790 EVENT_LIST (12);
1791
813 const size_t src_origin[3] = { src_x, src_y, src_z }; 1792 const size_t src_origin[3] = { src_x, src_y, src_z };
814 const size_t region[3] = { width, height, depth }; 1793 const size_t region[3] = { width, height, depth };
815 EVENT_LIST (12, items - 12);
816 1794
817 if (!row_pitch) 1795 if (!row_pitch)
818 row_pitch = img_row_pitch (src); 1796 row_pitch = img_row_pitch (src);
819 1797
820 if (depth > 1 && !slice_pitch) 1798 if (depth > 1 && !slice_pitch)
824 1802
825 SvUPGRADE (data, SVt_PV); 1803 SvUPGRADE (data, SVt_PV);
826 SvGROW (data, len); 1804 SvGROW (data, len);
827 SvPOK_only (data); 1805 SvPOK_only (data);
828 SvCUR_set (data, len); 1806 SvCUR_set (data, len);
1807
1808 cl_event ev = 0;
829 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)); 1809 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));
830 1810
831 if (ev) 1811 if (ev)
832 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1812 XPUSH_CLOBJ (stash_event, ev);
833 1813
834void 1814void
835enqueue_write_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, ...) 1815write_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, ...)
1816 ALIAS:
1817 enqueue_write_image = 0
836 PPCODE: 1818 PPCODE:
837 cl_event ev = 0; 1819 EVENT_LIST (12);
1820
838 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1821 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
839 const size_t region[3] = { width, height, depth }; 1822 const size_t region[3] = { width, height, depth };
840 STRLEN len; 1823 STRLEN len;
841 char *ptr = SvPVbyte (data, len); 1824 char *ptr = SvPVbyte (data, len);
842 EVENT_LIST (12, items - 12);
843 1825
844 if (!row_pitch) 1826 if (!row_pitch)
845 row_pitch = img_row_pitch (dst); 1827 row_pitch = img_row_pitch (dst);
846 1828
847 if (depth > 1 && !slice_pitch) 1829 if (depth > 1 && !slice_pitch)
850 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; 1832 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
851 1833
852 if (len < min_len) 1834 if (len < min_len)
853 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); 1835 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
854 1836
1837 cl_event ev = 0;
855 NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1838 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));
856 1839
857 if (ev) 1840 if (ev)
858 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1841 XPUSH_CLOBJ (stash_event, ev);
859 1842
860void 1843void
861enqueue_copy_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, ...) 1844copy_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, ...)
1845 ALIAS:
1846 enqueue_copy_image = 0
862 PPCODE: 1847 PPCODE:
863 cl_event ev = 0; 1848 EVENT_LIST (12);
1849
864 const size_t src_origin[3] = { src_x, src_y, src_z }; 1850 const size_t src_origin[3] = { src_x, src_y, src_z };
865 const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; 1851 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
866 const size_t region[3] = { width, height, depth }; 1852 const size_t region[3] = { width, height, depth };
867 EVENT_LIST (12, items - 12);
868 1853
1854 cl_event ev = 0;
869 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1855 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
870 1856
871 if (ev) 1857 if (ev)
872 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1858 XPUSH_CLOBJ (stash_event, ev);
873 1859
874void 1860void
875enqueue_copy_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, ...) 1861copy_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, ...)
1862 ALIAS:
1863 enqueue_copy_image_to_buffer = 0
876 PPCODE: 1864 PPCODE:
1865 EVENT_LIST (10);
1866
1867 const size_t src_origin[3] = { src_x, src_y, src_z };
1868 const size_t region [3] = { width, height, depth };
1869
877 cl_event ev = 0; 1870 cl_event ev = 0;
1871 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1872
1873 if (ev)
1874 XPUSH_CLOBJ (stash_event, ev);
1875
1876void
1877copy_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, ...)
1878 ALIAS:
1879 enqueue_copy_buffer_to_image = 0
1880 PPCODE:
1881 EVENT_LIST (10);
1882
878 const size_t src_origin[3] = { src_x, src_y, src_z }; 1883 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1884 const size_t region [3] = { width, height, depth };
1885
1886 cl_event ev = 0;
1887 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1888
1889 if (ev)
1890 XPUSH_CLOBJ (stash_event, ev);
1891
1892void
1893map_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, ...)
1894 ALIAS:
1895 enqueue_map_buffer = 0
1896 PPCODE:
1897 EVENT_LIST (6);
1898
1899 size_t cb = SvIV (cb_);
1900
1901 if (!SvOK (cb_))
1902 {
1903 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1904 cb -= offset;
1905 }
1906
1907 cl_event ev;
1908 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
1909 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0));
1910
1911void
1912map_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, ...)
1913 ALIAS:
1914 enqueue_map_image = 0
1915 PPCODE:
1916 size_t width = SvIV (width_);
1917 if (!SvOK (width_))
1918 {
1919 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0));
1920 width -= x;
1921 }
1922
1923 size_t height = SvIV (width_);
1924 if (!SvOK (height_))
1925 {
1926 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0));
1927 height -= y;
1928 }
1929
1930 size_t depth = SvIV (width_);
1931 if (!SvOK (depth_))
1932 {
1933 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0));
1934 depth -= z;
1935
1936 // stupid opencl returns 0 for depth, but requires 1 for 2d images
1937 if (!depth)
1938 depth = 1;
1939 }
1940
1941 const size_t origin[3] = { x, y, z };
879 const size_t region[3] = { width, height, depth }; 1942 const size_t region[3] = { width, height, depth };
1943 size_t row_pitch, slice_pitch;
880 EVENT_LIST (10, items - 10); 1944 EVENT_LIST (10);
881 1945
1946 cl_event ev;
1947 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));
1948
1949 size_t cb = slice_pitch ? slice_pitch * region [2]
1950 : row_pitch ? row_pitch * region [1]
1951 : region [0];
1952
1953 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch));
1954
1955void
1956unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
1957 PPCODE:
1958 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1959 if (GIMME_V != G_VOID)
1960 {
1961 clRetainEvent (mapped->event);
1962 XPUSH_CLOBJ (stash_event, mapped->event);
1963 }
1964
1965void
1966task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1967 ALIAS:
1968 enqueue_task = 0
1969 PPCODE:
1970 EVENT_LIST (2);
1971
1972 cl_event ev = 0;
882 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1973 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
883 1974
884 if (ev) 1975 if (ev)
885 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1976 XPUSH_CLOBJ (stash_event, ev);
886 1977
887void 1978void
888enqueue_copy_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, ...)
889 PPCODE:
890 cl_event ev = 0;
891 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
892 const size_t region[3] = { width, height, depth };
893 EVENT_LIST (10, items - 10);
894
895 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
896
897 if (ev)
898 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
899
900void
901enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
902 PPCODE:
903 cl_event ev = 0;
904 EVENT_LIST (2, items - 2);
905
906 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
907
908 if (ev)
909 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
910
911void
912enqueue_nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) 1979nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1980 ALIAS:
1981 enqueue_nd_range_kernel = 0
913 PPCODE: 1982 PPCODE:
914 cl_event ev = 0; 1983 EVENT_LIST (5);
1984
915 size_t *gwo = 0, *gws, *lws = 0; 1985 size_t *gwo = 0, *gws, *lws = 0;
916 int gws_len; 1986 int gws_len;
917 size_t *lists; 1987 size_t *lists;
918 int i; 1988 int i;
919 EVENT_LIST (5, items - 5);
920 1989
921 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) 1990 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
922 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); 1991 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
923 1992
924 gws_len = AvFILLp (SvRV (global_work_size)) + 1; 1993 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
925 1994
926 lists = tmpbuf (sizeof (size_t) * 3 * gws_len); 1995 lists = tmpbuf (sizeof (size_t) * 3 * gws_len);
927 1996
928 gws = lists + gws_len * 0; 1997 gws = lists + gws_len * 0;
929 for (i = 0; i < gws_len; ++i) 1998 for (i = 0; i < gws_len; ++i)
1999 {
930 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); 2000 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]);
2001 // at least nvidia crashes for 0-sized work group sizes, work around
2002 if (!gws [i])
2003 croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i);
2004 }
931 2005
932 if (SvOK (global_work_offset)) 2006 if (SvOK (global_work_offset))
933 { 2007 {
934 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) 2008 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV)
935 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); 2009 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference");
942 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); 2016 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
943 } 2017 }
944 2018
945 if (SvOK (local_work_size)) 2019 if (SvOK (local_work_size))
946 { 2020 {
947 if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) 2021 if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV)
948 croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); 2022 croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference");
949 2023
950 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) 2024 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
951 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); 2025 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
952 2026
953 lws = lists + gws_len * 2; 2027 lws = lists + gws_len * 2;
954 for (i = 0; i < gws_len; ++i) 2028 for (i = 0; i < gws_len; ++i)
2029 {
955 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); 2030 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]);
2031 // at least nvidia crashes for 0-sized work group sizes, work around
2032 if (!lws [i])
2033 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
2034 }
956 } 2035 }
957 2036
2037 cl_event ev = 0;
958 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 2038 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
959 2039
960 if (ev) 2040 if (ev)
961 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 2041 XPUSH_CLOBJ (stash_event, ev);
962 2042
2043#if CL_VERSION_1_2
2044
963void 2045void
964enqueue_marker (OpenCL::Queue self) 2046migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...)
2047 ALIAS:
2048 enqueue_migrate_mem_objects = 0
965 PPCODE: 2049 PPCODE:
2050 EVENT_LIST (3);
2051
2052 cl_uint obj_count;
2053 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2054
2055 cl_event ev = 0;
2056 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2057
2058 if (ev)
2059 XPUSH_CLOBJ (stash_event, ev);
2060
2061#endif
2062
2063#if cl_apple_gl_sharing || cl_khr_gl_sharing
2064
2065void
2066acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
2067 ALIAS:
2068 release_gl_objects = 1
2069 enqueue_acquire_gl_objects = 0
2070 enqueue_release_gl_objects = 1
2071 PPCODE:
2072 EVENT_LIST (2);
2073
2074 cl_uint obj_count;
2075 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2076
966 cl_event ev; 2077 cl_event ev = 0;
967 NEED_SUCCESS (EnqueueMarker, (self, &ev));
968 XPUSH_NEW_OBJ ("OpenCL::Event", ev);
969 2078
2079 if (ix)
2080 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2081 else
2082 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2083
2084 if (ev)
2085 XPUSH_CLOBJ (stash_event, ev);
2086
2087#endif
2088
970void 2089void
971enqueue_wait_for_events (OpenCL::Queue self, ...) 2090wait_for_events (OpenCL::Queue self, ...)
2091 ALIAS:
2092 enqueue_wait_for_events = 0
972 CODE: 2093 CODE:
973 EVENT_LIST (1, items - 1); 2094 EVENT_LIST (1);
2095#if PREFER_1_1
974 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); 2096 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
2097#else
2098 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0));
2099#endif
975 2100
976void 2101void
977enqueue_barrier (OpenCL::Queue self) 2102marker (OpenCL::Queue self, ...)
2103 ALIAS:
2104 enqueue_marker = 0
978 CODE: 2105 PPCODE:
2106 EVENT_LIST (1);
2107 cl_event ev = 0;
2108#if PREFER_1_1
2109 if (!event_list_count)
2110 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2111 else
2112#if CL_VERSION_1_2
2113 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2114#else
2115 {
2116 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier
2117 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2118 }
2119#endif
2120#else
2121 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2122#endif
2123 if (ev)
2124 XPUSH_CLOBJ (stash_event, ev);
2125
2126void
2127barrier (OpenCL::Queue self, ...)
2128 ALIAS:
2129 enqueue_barrier = 0
2130 PPCODE:
2131 EVENT_LIST (1);
2132 cl_event ev = 0;
2133#if PREFER_1_1
2134 if (!event_list_count && GIMME_V == G_VOID)
979 NEED_SUCCESS (EnqueueBarrier, (self)); 2135 NEED_SUCCESS (EnqueueBarrier, (self));
2136 else
2137#if CL_VERSION_1_2
2138 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2139#else
2140 {
2141 if (event_list_count)
2142 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
2143
2144 if (GIMME_V != G_VOID)
2145 NEED_SUCCESS (EnqueueMarker, (self, &ev));
2146 }
2147#endif
2148#else
2149 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2150#endif
2151 if (ev)
2152 XPUSH_CLOBJ (stash_event, ev);
980 2153
981void 2154void
982flush (OpenCL::Queue self) 2155flush (OpenCL::Queue self)
983 CODE: 2156 CODE:
984 NEED_SUCCESS (Flush, (self)); 2157 NEED_SUCCESS (Flush, (self));
1002 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); 2175 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0));
1003 EXTEND (SP, 1); 2176 EXTEND (SP, 1);
1004 const int i = 0; 2177 const int i = 0;
1005 { 2178 {
1006 NEED_SUCCESS (RetainContext, (value [i])); 2179 NEED_SUCCESS (RetainContext, (value [i]));
1007 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); 2180 PUSH_CLOBJ (stash_context, value [i]);
1008 } 2181 }
1009 2182
1010void 2183void
1011device (OpenCL::Queue self) 2184device (OpenCL::Queue self)
1012 PPCODE: 2185 PPCODE:
1013 cl_device_id value [1]; 2186 cl_device_id value [1];
1014 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); 2187 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0));
1015 EXTEND (SP, 1); 2188 EXTEND (SP, 1);
1016 const int i = 0; 2189 const int i = 0;
1017 { 2190 {
1018 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); 2191 PUSH_CLOBJ (stash_device, value [i]);
1019 } 2192 }
1020 2193
1021void 2194void
1022reference_count (OpenCL::Queue self) 2195reference_count (OpenCL::Queue self)
1023 PPCODE: 2196 PPCODE:
1032 PPCODE: 2205 PPCODE:
1033 cl_command_queue_properties value [1]; 2206 cl_command_queue_properties value [1];
1034 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); 2207 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0));
1035 EXTEND (SP, 1); 2208 EXTEND (SP, 1);
1036 const int i = 0; 2209 const int i = 0;
1037 PUSHs (sv_2mortal (newSViv (value [i]))); 2210 PUSHs (sv_2mortal (newSVuv (value [i])));
1038 2211
1039#END:command_queue 2212#END:command_queue
1040 2213
1041MODULE = OpenCL PACKAGE = OpenCL::Memory 2214MODULE = OpenCL PACKAGE = OpenCL::Memory
1042 2215
1047 2220
1048void 2221void
1049info (OpenCL::Memory self, cl_mem_info name) 2222info (OpenCL::Memory self, cl_mem_info name)
1050 PPCODE: 2223 PPCODE:
1051 INFO (MemObject) 2224 INFO (MemObject)
2225
2226void
2227destructor_callback (OpenCL::Memory self, SV *notify)
2228 PPCODE:
2229 clSetMemObjectDestructorCallback (self, eq_destructor_notify, SvREFCNT_inc (s_get_cv (notify)));
1052 2230
1053#BEGIN:mem 2231#BEGIN:mem
1054 2232
1055void 2233void
1056type (OpenCL::Memory self) 2234type (OpenCL::Memory self)
1057 PPCODE: 2235 PPCODE:
1058 cl_mem_object_type value [1]; 2236 cl_mem_object_type value [1];
1059 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); 2237 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0));
1060 EXTEND (SP, 1); 2238 EXTEND (SP, 1);
1061 const int i = 0; 2239 const int i = 0;
1062 PUSHs (sv_2mortal (newSViv (value [i]))); 2240 PUSHs (sv_2mortal (newSVuv (value [i])));
1063 2241
1064void 2242void
1065flags (OpenCL::Memory self) 2243flags (OpenCL::Memory self)
1066 PPCODE: 2244 PPCODE:
1067 cl_mem_flags value [1]; 2245 cl_mem_flags value [1];
1068 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); 2246 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0));
1069 EXTEND (SP, 1); 2247 EXTEND (SP, 1);
1070 const int i = 0; 2248 const int i = 0;
1071 PUSHs (sv_2mortal (newSViv (value [i]))); 2249 PUSHs (sv_2mortal (newSVuv (value [i])));
1072 2250
1073void 2251void
1074size (OpenCL::Memory self) 2252size (OpenCL::Memory self)
1075 ALIAS: 2253 ALIAS:
1076 size = CL_MEM_SIZE 2254 size = CL_MEM_SIZE
1110 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); 2288 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0));
1111 EXTEND (SP, 1); 2289 EXTEND (SP, 1);
1112 const int i = 0; 2290 const int i = 0;
1113 { 2291 {
1114 NEED_SUCCESS (RetainContext, (value [i])); 2292 NEED_SUCCESS (RetainContext, (value [i]));
1115 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); 2293 PUSH_CLOBJ (stash_context, value [i]);
1116 } 2294 }
1117 2295
1118void 2296void
1119associated_memobject (OpenCL::Memory self) 2297associated_memobject (OpenCL::Memory self)
1120 PPCODE: 2298 PPCODE:
1122 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); 2300 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0));
1123 EXTEND (SP, 1); 2301 EXTEND (SP, 1);
1124 const int i = 0; 2302 const int i = 0;
1125 { 2303 {
1126 NEED_SUCCESS (RetainMemObject, (value [i])); 2304 NEED_SUCCESS (RetainMemObject, (value [i]));
1127 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); 2305 PUSH_CLOBJ (stash_memory, value [i]);
1128 } 2306 }
1129 2307
1130#END:mem 2308#END:mem
2309
2310#if cl_apple_gl_sharing || cl_khr_gl_sharing
2311
2312void
2313gl_object_info (OpenCL::Memory self)
2314 PPCODE:
2315 cl_gl_object_type type;
2316 cl_GLuint name;
2317 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
2318 EXTEND (SP, 2);
2319 PUSHs (sv_2mortal (newSVuv (type)));
2320 PUSHs (sv_2mortal (newSVuv (name)));
2321
2322#endif
1131 2323
1132MODULE = OpenCL PACKAGE = OpenCL::BufferObj 2324MODULE = OpenCL PACKAGE = OpenCL::BufferObj
1133 2325
1134void 2326void
1135sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) 2327sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
1138 croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); 2330 croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!");
1139 2331
1140 cl_buffer_region crdata = { origin, size }; 2332 cl_buffer_region crdata = { origin, size };
1141 2333
1142 NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); 2334 NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res));
1143 XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); 2335 XPUSH_CLOBJ (stash_buffer, mem);
1144 2336
1145MODULE = OpenCL PACKAGE = OpenCL::Image 2337MODULE = OpenCL PACKAGE = OpenCL::Image
1146 2338
1147void 2339void
1148image_info (OpenCL::Image self, cl_image_info name) 2340image_info (OpenCL::Image self, cl_image_info name)
1149 PPCODE: 2341 PPCODE:
1150 INFO (Image) 2342 INFO (Image)
2343
2344void
2345format (OpenCL::Image self)
2346 PPCODE:
2347 cl_image_format format;
2348 NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
2349 EXTEND (SP, 2);
2350 PUSHs (sv_2mortal (newSVuv (format.image_channel_order)));
2351 PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type)));
1151 2352
1152#BEGIN:image 2353#BEGIN:image
1153 2354
1154void 2355void
1155element_size (OpenCL::Image self) 2356element_size (OpenCL::Image self)
1167 const int i = 0; 2368 const int i = 0;
1168 PUSHs (sv_2mortal (newSVuv (value [i]))); 2369 PUSHs (sv_2mortal (newSVuv (value [i])));
1169 2370
1170#END:image 2371#END:image
1171 2372
2373#if cl_apple_gl_sharing || cl_khr_gl_sharing
2374
2375#BEGIN:gl_texture
2376
2377void
2378target (OpenCL::Image self)
2379 PPCODE:
2380 cl_GLenum value [1];
2381 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
2382 EXTEND (SP, 1);
2383 const int i = 0;
2384 PUSHs (sv_2mortal (newSVuv (value [i])));
2385
2386void
2387gl_mipmap_level (OpenCL::Image self)
2388 PPCODE:
2389 cl_GLint value [1];
2390 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
2391 EXTEND (SP, 1);
2392 const int i = 0;
2393 PUSHs (sv_2mortal (newSViv (value [i])));
2394
2395#END:gl_texture
2396
2397#endif
2398
1172MODULE = OpenCL PACKAGE = OpenCL::Sampler 2399MODULE = OpenCL PACKAGE = OpenCL::Sampler
1173 2400
1174void 2401void
1175DESTROY (OpenCL::Sampler self) 2402DESTROY (OpenCL::Sampler self)
1176 CODE: 2403 CODE:
1199 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); 2426 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0));
1200 EXTEND (SP, 1); 2427 EXTEND (SP, 1);
1201 const int i = 0; 2428 const int i = 0;
1202 { 2429 {
1203 NEED_SUCCESS (RetainContext, (value [i])); 2430 NEED_SUCCESS (RetainContext, (value [i]));
1204 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); 2431 PUSH_CLOBJ (stash_context, value [i]);
1205 } 2432 }
1206 2433
1207void 2434void
1208normalized_coords (OpenCL::Sampler self) 2435normalized_coords (OpenCL::Sampler self)
1209 PPCODE: 2436 PPCODE:
1210 cl_addressing_mode value [1]; 2437 cl_addressing_mode value [1];
1211 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); 2438 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0));
1212 EXTEND (SP, 1); 2439 EXTEND (SP, 1);
1213 const int i = 0; 2440 const int i = 0;
1214 PUSHs (sv_2mortal (newSViv (value [i]))); 2441 PUSHs (sv_2mortal (newSVuv (value [i])));
1215 2442
1216void 2443void
1217addressing_mode (OpenCL::Sampler self) 2444addressing_mode (OpenCL::Sampler self)
1218 PPCODE: 2445 PPCODE:
1219 cl_filter_mode value [1]; 2446 cl_filter_mode value [1];
1220 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); 2447 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0));
1221 EXTEND (SP, 1); 2448 EXTEND (SP, 1);
1222 const int i = 0; 2449 const int i = 0;
1223 PUSHs (sv_2mortal (newSViv (value [i]))); 2450 PUSHs (sv_2mortal (newSVuv (value [i])));
1224 2451
1225void 2452void
1226filter_mode (OpenCL::Sampler self) 2453filter_mode (OpenCL::Sampler self)
1227 PPCODE: 2454 PPCODE:
1228 cl_bool value [1]; 2455 cl_bool value [1];
1239DESTROY (OpenCL::Program self) 2466DESTROY (OpenCL::Program self)
1240 CODE: 2467 CODE:
1241 clReleaseProgram (self); 2468 clReleaseProgram (self);
1242 2469
1243void 2470void
1244build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) 2471build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2472 ALIAS:
2473 build_async = 1
1245 CODE: 2474 CODE:
2475 cl_uint device_count = 0;
2476 cl_device_id *device_list = 0;
2477
2478 if (SvOK (devices))
2479 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2480
2481 void *user_data;
2482 program_callback pfn_notify = make_program_callback (notify, &user_data);
2483
2484 if (ix)
2485 build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data);
2486 else
2487 NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
2488
2489#if CL_VERSION_1_2
2490
2491void
2492compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef)
2493 CODE:
2494 cl_uint device_count = 0;
2495 cl_device_id *device_list = 0;
2496
2497 if (SvOK (devices))
2498 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2499
2500 cl_uint header_count = 0;
2501 cl_program *header_list = 0;
2502 const char **header_name = 0;
2503
2504 if (SvOK (headers))
2505 {
2506 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2507 croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs");
2508
2509 HV *hv = (HV *)SvRV (devices);
2510
2511 header_count = hv_iterinit (hv);
2512 header_list = tmpbuf (sizeof (*header_list) * header_count);
2513 header_name = tmpbuf (sizeof (*header_name) * header_count);
2514
2515 HE *he;
2516 int i = 0;
2517 while (he = hv_iternext (hv))
2518 {
2519 header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he));
2520 header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program");
2521 ++i;
2522 }
2523 }
2524
2525 void *user_data;
2526 program_callback pfn_notify = make_program_callback (notify, &user_data);
2527
1246 NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); 2528 NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
2529 header_count, header_list, header_name, pfn_notify, user_data));
2530
2531#endif
1247 2532
1248void 2533void
1249build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) 2534build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
1250 PPCODE: 2535 PPCODE:
1251 size_t size; 2536 size_t size;
1280 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); 2565 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
1281 EXTEND (SP, 1); 2566 EXTEND (SP, 1);
1282 const int i = 0; 2567 const int i = 0;
1283 PUSHs (sv_2mortal (newSVpv (value, 0))); 2568 PUSHs (sv_2mortal (newSVpv (value, 0)));
1284 2569
2570void
2571binary_type (OpenCL::Program self, OpenCL::Device device)
2572 PPCODE:
2573 cl_program_binary_type value [1];
2574 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0));
2575 EXTEND (SP, 1);
2576 const int i = 0;
2577 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
2578
1285#END:program_build 2579#END:program_build
1286 2580
1287void 2581void
1288kernel (OpenCL::Program program, SV *function) 2582kernel (OpenCL::Program program, SV *function)
1289 PPCODE: 2583 PPCODE:
1290 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); 2584 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
1291 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 2585 XPUSH_CLOBJ (stash_kernel, kernel);
2586
2587void
2588kernels_in_program (OpenCL::Program program)
2589 PPCODE:
2590 cl_uint num_kernels;
2591 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
2592 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
2593 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
2594
2595 int i;
2596 EXTEND (SP, num_kernels);
2597 for (i = 0; i < num_kernels; ++i)
2598 PUSH_CLOBJ (stash_kernel, kernels [i]);
1292 2599
1293void 2600void
1294info (OpenCL::Program self, cl_program_info name) 2601info (OpenCL::Program self, cl_program_info name)
1295 PPCODE: 2602 PPCODE:
1296 INFO (Program) 2603 INFO (Program)
1314 { 2621 {
1315 SV *sv = sv_2mortal (newSV (sizes [i])); 2622 SV *sv = sv_2mortal (newSV (sizes [i]));
1316 SvUPGRADE (sv, SVt_PV); 2623 SvUPGRADE (sv, SVt_PV);
1317 SvPOK_only (sv); 2624 SvPOK_only (sv);
1318 SvCUR_set (sv, sizes [i]); 2625 SvCUR_set (sv, sizes [i]);
1319 ptrs [i] = SvPVX (sv); 2626 ptrs [i] = (void *)SvPVX (sv);
1320 PUSHs (sv); 2627 PUSHs (sv);
1321 } 2628 }
1322 2629
1323 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); 2630 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
1324 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; 2631 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
1344 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); 2651 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0));
1345 EXTEND (SP, 1); 2652 EXTEND (SP, 1);
1346 const int i = 0; 2653 const int i = 0;
1347 { 2654 {
1348 NEED_SUCCESS (RetainContext, (value [i])); 2655 NEED_SUCCESS (RetainContext, (value [i]));
1349 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); 2656 PUSH_CLOBJ (stash_context, value [i]);
1350 } 2657 }
1351 2658
1352void 2659void
1353devices (OpenCL::Program self) 2660devices (OpenCL::Program self)
1354 PPCODE: 2661 PPCODE:
1358 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); 2665 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0));
1359 int i, n = size / sizeof (*value); 2666 int i, n = size / sizeof (*value);
1360 EXTEND (SP, n); 2667 EXTEND (SP, n);
1361 for (i = 0; i < n; ++i) 2668 for (i = 0; i < n; ++i)
1362 { 2669 {
1363 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); 2670 PUSH_CLOBJ (stash_device, value [i]);
1364 } 2671 }
1365 2672
1366void 2673void
1367source (OpenCL::Program self) 2674source (OpenCL::Program self)
1368 PPCODE: 2675 PPCODE:
1394DESTROY (OpenCL::Kernel self) 2701DESTROY (OpenCL::Kernel self)
1395 CODE: 2702 CODE:
1396 clReleaseKernel (self); 2703 clReleaseKernel (self);
1397 2704
1398void 2705void
2706setf (OpenCL::Kernel self, const char *format, ...)
2707 CODE:
2708 int i;
2709 for (i = 2; ; ++i)
2710 {
2711 while (*format == ' ')
2712 ++format;
2713
2714 char type = *format++;
2715
2716 if (!type)
2717 break;
2718
2719 if (i >= items)
2720 croak ("OpenCL::Kernel::setf format string too long (not enough arguments)");
2721
2722 SV *sv = ST (i);
2723
2724 union
2725 {
2726 cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS;
2727 cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL;
2728 cl_half ch; cl_float cf; cl_double cd;
2729 cl_mem cm;
2730 cl_sampler ca;
2731 size_t cz;
2732 cl_event ce;
2733 } arg;
2734 size_t size;
2735 int nullarg = 0;
2736
2737 switch (type)
2738 {
2739 case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break;
2740 case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break;
2741 case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break;
2742 case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break;
2743 case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break;
2744 case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break;
2745 case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break;
2746 case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break;
2747
2748 case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break;
2749 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2750 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2751
2752 case 'z': nullarg = 1; size = SvIV (sv); break;
2753
2754 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2755 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2756 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2757
2758 default:
2759 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2760 }
2761
2762 res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg);
2763 if (res)
2764 croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res));
2765 }
2766
2767 if (i != items)
2768 croak ("OpenCL::Kernel::setf format string too short (too many arguments)");
2769
2770void
1399set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) 2771set_char (OpenCL::Kernel self, cl_uint idx, cl_char value)
1400 CODE: 2772 CODE:
1401 clSetKernelArg (self, idx, sizeof (value), &value); 2773 clSetKernelArg (self, idx, sizeof (value), &value);
1402 2774
1403void 2775void
1451 clSetKernelArg (self, idx, sizeof (value), &value); 2823 clSetKernelArg (self, idx, sizeof (value), &value);
1452 2824
1453void 2825void
1454set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) 2826set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value)
1455 CODE: 2827 CODE:
2828 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2829
2830void
2831set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value)
2832 CODE:
2833 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2834
2835void
2836set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value)
2837 CODE:
2838 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2839
2840void
2841set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
2842 CODE:
1456 clSetKernelArg (self, idx, sizeof (value), &value); 2843 clSetKernelArg (self, idx, sizeof (value), &value);
1457 2844
1458void 2845void
1459set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) 2846set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
1460 CODE: 2847 CODE:
1461 clSetKernelArg (self, idx, sizeof (value), &value); 2848 clSetKernelArg (self, idx, size, 0);
1462
1463void
1464set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value)
1465 CODE:
1466 clSetKernelArg (self, idx, sizeof (value), &value);
1467
1468void
1469set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value)
1470 CODE:
1471 clSetKernelArg (self, idx, sizeof (value), &value);
1472
1473void
1474set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
1475 CODE:
1476 clSetKernelArg (self, idx, sizeof (value), &value);
1477 2849
1478void 2850void
1479set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) 2851set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
1480 CODE: 2852 CODE:
1481 clSetKernelArg (self, idx, sizeof (value), &value); 2853 clSetKernelArg (self, idx, sizeof (value), &value);
1517 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); 2889 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0));
1518 EXTEND (SP, 1); 2890 EXTEND (SP, 1);
1519 const int i = 0; 2891 const int i = 0;
1520 { 2892 {
1521 NEED_SUCCESS (RetainContext, (value [i])); 2893 NEED_SUCCESS (RetainContext, (value [i]));
1522 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); 2894 PUSH_CLOBJ (stash_context, value [i]);
1523 } 2895 }
1524 2896
1525void 2897void
1526program (OpenCL::Kernel self) 2898program (OpenCL::Kernel self)
1527 PPCODE: 2899 PPCODE:
1529 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); 2901 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0));
1530 EXTEND (SP, 1); 2902 EXTEND (SP, 1);
1531 const int i = 0; 2903 const int i = 0;
1532 { 2904 {
1533 NEED_SUCCESS (RetainProgram, (value [i])); 2905 NEED_SUCCESS (RetainProgram, (value [i]));
1534 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); 2906 PUSH_CLOBJ (stash_program, value [i]);
1535 } 2907 }
1536 2908
1537#END:kernel 2909#END:kernel
1538 2910
1539void 2911void
1586 const int i = 0; 2958 const int i = 0;
1587 PUSHs (sv_2mortal (newSVuv (value [i]))); 2959 PUSHs (sv_2mortal (newSVuv (value [i])));
1588 2960
1589#END:kernel_work_group 2961#END:kernel_work_group
1590 2962
2963#if CL_VERSION_1_2
2964
2965void
2966arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name)
2967 PPCODE:
2968 size_t size;
2969 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size));
2970 SV *sv = sv_2mortal (newSV (size));
2971 SvUPGRADE (sv, SVt_PV);
2972 SvPOK_only (sv);
2973 SvCUR_set (sv, size);
2974 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0));
2975 XPUSHs (sv);
2976
2977#BEGIN:kernel_arg
2978
2979void
2980arg_address_qualifier (OpenCL::Kernel self, cl_uint idx)
2981 PPCODE:
2982 cl_kernel_arg_address_qualifier value [1];
2983 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0));
2984 EXTEND (SP, 1);
2985 const int i = 0;
2986 PUSHs (sv_2mortal (newSVuv (value [i])));
2987
2988void
2989arg_access_qualifier (OpenCL::Kernel self, cl_uint idx)
2990 PPCODE:
2991 cl_kernel_arg_access_qualifier value [1];
2992 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0));
2993 EXTEND (SP, 1);
2994 const int i = 0;
2995 PUSHs (sv_2mortal (newSVuv (value [i])));
2996
2997void
2998arg_type_name (OpenCL::Kernel self, cl_uint idx)
2999 ALIAS:
3000 arg_type_name = CL_KERNEL_ARG_TYPE_NAME
3001 arg_name = CL_KERNEL_ARG_NAME
3002 PPCODE:
3003 size_t size;
3004 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size));
3005 char *value = tmpbuf (size);
3006 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0));
3007 EXTEND (SP, 1);
3008 const int i = 0;
3009 PUSHs (sv_2mortal (newSVpv (value, 0)));
3010
3011void
3012arg_type_qualifier (OpenCL::Kernel self, cl_uint idx)
3013 PPCODE:
3014 cl_kernel_arg_type_qualifier value [1];
3015 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0));
3016 EXTEND (SP, 1);
3017 const int i = 0;
3018 PUSHs (sv_2mortal (newSVuv (value [i])));
3019
3020#END:kernel_arg
3021
3022#endif
3023
1591MODULE = OpenCL PACKAGE = OpenCL::Event 3024MODULE = OpenCL PACKAGE = OpenCL::Event
1592 3025
1593void 3026void
1594DESTROY (OpenCL::Event self) 3027DESTROY (OpenCL::Event self)
1595 CODE: 3028 CODE:
1597 3030
1598void 3031void
1599wait (OpenCL::Event self) 3032wait (OpenCL::Event self)
1600 CODE: 3033 CODE:
1601 clWaitForEvents (1, &self); 3034 clWaitForEvents (1, &self);
3035
3036void
3037cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb)
3038 CODE:
3039 clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb)));
1602 3040
1603void 3041void
1604info (OpenCL::Event self, cl_event_info name) 3042info (OpenCL::Event self, cl_event_info name)
1605 PPCODE: 3043 PPCODE:
1606 INFO (Event) 3044 INFO (Event)
1614 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); 3052 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0));
1615 EXTEND (SP, 1); 3053 EXTEND (SP, 1);
1616 const int i = 0; 3054 const int i = 0;
1617 { 3055 {
1618 NEED_SUCCESS (RetainCommandQueue, (value [i])); 3056 NEED_SUCCESS (RetainCommandQueue, (value [i]));
1619 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); 3057 PUSH_CLOBJ (stash_queue, value [i]);
1620 } 3058 }
1621 3059
1622void 3060void
1623command_type (OpenCL::Event self) 3061command_type (OpenCL::Event self)
1624 PPCODE: 3062 PPCODE:
1647 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); 3085 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0));
1648 EXTEND (SP, 1); 3086 EXTEND (SP, 1);
1649 const int i = 0; 3087 const int i = 0;
1650 { 3088 {
1651 NEED_SUCCESS (RetainContext, (value [i])); 3089 NEED_SUCCESS (RetainContext, (value [i]));
1652 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); 3090 PUSH_CLOBJ (stash_context, value [i]);
1653 } 3091 }
1654 3092
1655#END:event 3093#END:event
1656 3094
1657void 3095void
1682void 3120void
1683set_status (OpenCL::UserEvent self, cl_int execution_status) 3121set_status (OpenCL::UserEvent self, cl_int execution_status)
1684 CODE: 3122 CODE:
1685 clSetUserEventStatus (self, execution_status); 3123 clSetUserEventStatus (self, execution_status);
1686 3124
3125MODULE = OpenCL PACKAGE = OpenCL::Mapped
3126
3127void
3128DESTROY (SV *self)
3129 CODE:
3130 OpenCL__Mapped mapped = SvMAPPED (self);
3131
3132 clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0);
3133 mapped_detach (self, mapped);
3134
3135 clReleaseCommandQueue (mapped->queue);
3136 clReleaseEvent (mapped->event);
3137 Safefree (mapped);
3138
3139void
3140unmap (OpenCL::Mapped self, ...)
3141 CODE:
3142 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
3143
3144bool
3145mapped (OpenCL::Mapped self)
3146 CODE:
3147 RETVAL = !!self->ptr;
3148 OUTPUT:
3149 RETVAL
3150
3151void
3152wait (OpenCL::Mapped self)
3153 PPCODE:
3154 if (self->event)
3155 NEED_SUCCESS (WaitForEvents, (1, &self->event));
3156
3157void
3158event (OpenCL::Mapped self)
3159 PPCODE:
3160 if (!self->event)
3161 XSRETURN_UNDEF;
3162
3163 clRetainEvent (self->event);
3164 XPUSH_CLOBJ (stash_event, self->event);
3165
3166size_t
3167size (OpenCL::Mapped self)
3168 CODE:
3169 RETVAL = self->cb;
3170 OUTPUT:
3171 RETVAL
3172
3173IV
3174ptr (OpenCL::Mapped self)
3175 CODE:
3176 RETVAL = PTR2IV (self->ptr);
3177 OUTPUT:
3178 RETVAL
3179
3180void
3181set (OpenCL::Mapped self, size_t offset, SV *data)
3182 CODE:
3183 STRLEN len;
3184 const char *ptr = SvPVbyte (data, len);
3185
3186 if (offset + len > self->cb)
3187 croak ("OpenCL::Mapped::set out of bound condition detected");
3188
3189 memcpy (offset + (char *)self->ptr, ptr, len);
3190
3191MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer
3192
3193MODULE = OpenCL PACKAGE = OpenCL::MappedImage
3194
3195IV
3196row_pitch (OpenCL::Mapped self)
3197 ALIAS:
3198 slice_pitch = 1
3199 CODE:
3200 RETVAL = ix ? self->slice_pitch : self->row_pitch;
3201 OUTPUT:
3202 RETVAL
3203
3204
3205

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines