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

Comparing OpenCL/OpenCL.xs (file contents):
Revision 1.2 by root, Tue Nov 15 09:24:40 2011 UTC vs.
Revision 1.79 by root, Sat May 5 15:43:02 2012 UTC

1#include "EXTERN.h" 1#include "EXTERN.h"
2#include "perl.h" 2#include "perl.h"
3#include "XSUB.h" 3#include "XSUB.h"
4 4
5#include "ecb.h"//D
6
7#define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really
8#include "xthread.h"
9#include "schmorp.h"
10
11#ifdef I_DLFCN
12 #include <dlfcn.h>
13#endif
14
15// how stupid is that, the 1.2 header files define CL_VERSION_1_1,
16// but then fail to define the api functions unless you ALSO define
17// this. This breaks 100% of the opencl 1.1 apps, for what reason?
18// after all, the functions are deprecated, not removed.
19// in addition, you cannot test for this in any future-proof way.
20// each time a new opencl version comes out, you need to make a new
21// release.
22#define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */
23
24#ifndef PREFER_1_1
25 #define PREFER_1_1 1
26#endif
27
28#if PREFER_1_1
29 #define CL_USE_DEPRECATED_OPENCL_1_1_APIS
30#endif
31
32#ifdef __APPLE__
33 #define CLHDR(name) <OpenCL/name>
34#else
35 #define CLHDR(name) <CL/name>
36#endif
37
5#include <CL/opencl.h> 38#include CLHDR(opencl.h)
6 39
40#ifndef CL_VERSION_1_2
41 #include CLHDR(cl_d3d9.h)
42#endif
43
44#if _WIN32
45 #include CLHDR(cl_d3d10.h)
46 #if CL_VERSION_1_2
47 #include CLHDR<cl_d3d11.h>
48 #endif
49 #include CLHDR<cl_dx9_media_sharing.h.h>
50#endif
51
52#ifndef CL_VERSION_1_2
53 #undef PREFER_1_1
54 #define PREFER_1_1 1
55#endif
56
57// make sure all constants we might use are actually defined
58#include "default.h"
59
7typedef cl_platform_id OpenCL__Platform; 60typedef cl_platform_id OpenCL__Platform;
8typedef cl_device_id OpenCL__Device; 61typedef cl_device_id OpenCL__Device;
62typedef cl_device_id OpenCL__SubDevice;
9typedef cl_context OpenCL__Context; 63typedef cl_context OpenCL__Context;
10typedef cl_command_queue OpenCL__Queue; 64typedef cl_command_queue OpenCL__Queue;
11typedef cl_mem OpenCL__Memory; 65typedef cl_mem OpenCL__Memory;
66typedef cl_mem OpenCL__Buffer;
67typedef cl_mem OpenCL__BufferObj;
68typedef cl_mem OpenCL__Image;
69typedef cl_mem OpenCL__Memory_ornull;
70typedef cl_mem OpenCL__Buffer_ornull;
71typedef cl_mem OpenCL__Image_ornull;
12typedef cl_sampler OpenCL__Sampler; 72typedef cl_sampler OpenCL__Sampler;
13typedef cl_program OpenCL__Program; 73typedef cl_program OpenCL__Program;
14typedef cl_kernel OpenCL__Kernel; 74typedef cl_kernel OpenCL__Kernel;
15typedef cl_event OpenCL__Event; 75typedef cl_event OpenCL__Event;
76typedef cl_event OpenCL__UserEvent;
16 77
17static const struct { 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;
104
105/*****************************************************************************/
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
152/* up to two temporary buffers */
153static void * ecb_noinline
154tmpbuf (size_t size)
155{
156 enum { buffers = 4 };
157 static int idx;
158 static void *buf [buffers];
159 static size_t len [buffers];
160
161 idx = (idx + 1) % buffers;
162
163 if (len [idx] < size)
164 {
165 free (buf [idx]);
166 len [idx] = ((size + 31) & ~4095) + 4096 - 32;
167 buf [idx] = malloc (len [idx]);
168 }
169
170 return buf [idx];
171}
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
188/*****************************************************************************/
189
190typedef struct
191{
18 IV iv; 192 IV iv;
19 const char *name; 193 const char *name;
20} cl_error[] = {
21#define def_error(name) { (IV)CL_ ## name, # name }, 194 #define const_iv(name) { (IV)CL_ ## name, # name },
22#include "invalid.h" 195} ivstr;
23}; 196
197typedef struct
198{
199 NV nv;
200 const char *name;
201 #define const_nv(name) { (NV)CL_ ## name, # name },
202} nvstr;
24 203
25static const char * 204static const char *
26clstrerror (cl_int res) 205iv2str (IV value, const ivstr *base, int count, const char *fallback)
27{ 206{
28 int i; 207 int i;
29 static char numbuf [32]; 208 static char strbuf [32];
30 209
31 for (i = sizeof (cl_error) / sizeof (cl_error [0]); i--; ) 210 for (i = count; i--; )
32 if (cl_error [i].iv == res) 211 if (base [i].iv == value)
33 return cl_error [i].name; 212 return base [i].name;
34 213
35 snprintf (numbuf, sizeof (numbuf), "ERROR(%d)", res); 214 snprintf (strbuf, sizeof (strbuf), fallback, (int)value);
36 215
37 return numbuf; 216 return strbuf;
38} 217}
39 218
219static const char *
220enum2str (cl_uint value)
221{
222 static const ivstr enumstr[] = {
223 #include "enumstr.h"
224 };
225
226 return iv2str (value, enumstr, sizeof (enumstr) / sizeof (enumstr [0]), "ENUM(0x%04x)");
227}
228
229static const char *
230err2str (cl_int err)
231{
232 static const ivstr errstr[] = {
233 #include "errstr.h"
234 };
235
236 return iv2str (err, errstr, sizeof (errstr) / sizeof (errstr [0]), "ERROR(%d)");
237}
238
239/*****************************************************************************/
240
241static cl_int res;
242
40#define FAIL(name,res) \ 243#define FAIL(name) \
41 croak ("cl" # name ": %s", clstrerror (res)); 244 croak ("cl" # name ": %s", err2str (res));
42 245
43#define NEED_SUCCESS(name,args) \ 246#define NEED_SUCCESS(name,args) \
44 do { \ 247 do { \
45 cl_int res = cl ## name args; \ 248 res = cl ## name args; \
46 \ 249 \
47 if (res) \ 250 if (res) \
48 FAIL (name, res); \ 251 FAIL (name); \
49 } while (0) 252 } while (0)
50 253
51#define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) 254#define NEED_SUCCESS_ARG(retdecl, name, args) \
52#define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) 255 retdecl = cl ## name args; \
256 if (res) \
257 FAIL (name);
53 258
54/*TODO*/ 259/*****************************************************************************/
55#define EVENT_LIST(items,count) cl_uint event_list_count = 0; cl_event *event_list_ptr = 0 260
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
273SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg)
274{
275 // sv_derived_from is quite slow :(
276 if (SvROK (sv) && sv_derived_from (sv, pkg))
277 return (void *)SvIV (SvRV (sv));
278
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));
290}
291
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 */
661
662static size_t
663img_row_pitch (cl_mem img)
664{
665 size_t res;
666 clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0);
667 return res;
668}
669
670static cl_event * ecb_noinline
671event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra)
672{
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)
679 return 0;
680
681 cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1));
682 int i = 0;
683
684 while (count--)
685 if (SvOK (items [count]))
686 list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event");
687
688 if (extra)
689 list [i++] = extra;
690
691 *rcount = i;
692
693 return i ? list : 0;
694}
695
696#define EVENT_LIST(skip) \
697 cl_uint event_list_count = items - (skip); \
698 cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0)
56 699
57#define INFO(class) \ 700#define INFO(class) \
58{ \ 701{ \
59 size_t size; \ 702 size_t size; \
60 SV *sv; \
61 \
62 NEED_SUCCESS (Get ## class ## Info, (this, name, 0, 0, &size)); \ 703 NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \
63 sv = sv_2mortal (newSV (size)); \ 704 SV *sv = sv_2mortal (newSV (size)); \
64 SvUPGRADE (sv, SVt_PV); \ 705 SvUPGRADE (sv, SVt_PV); \
65 SvPOK_only (sv); \ 706 SvPOK_only (sv); \
66 SvCUR_set (sv, size); \ 707 SvCUR_set (sv, size); \
67 NEED_SUCCESS (Get ## class ## Info, (this, name, size, SvPVX (sv), 0)); \ 708 NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \
68 XPUSHs (sv); \ 709 XPUSHs (sv); \
69} 710}
70 711
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
71MODULE = OpenCL PACKAGE = OpenCL 805MODULE = OpenCL PACKAGE = OpenCL
72 806
73PROTOTYPES: ENABLE 807PROTOTYPES: ENABLE
74 808
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
75BOOT: 820BOOT:
76{ 821{
77 HV *stash = gv_stashpv ("OpenCL", 1); 822 HV *stash = gv_stashpv ("OpenCL", 1);
78 static const struct { 823
79 const char *name; 824 static const ivstr *civ, const_iv[] = {
80 IV iv; 825 { sizeof (cl_char ), "SIZEOF_CHAR" },
81 } *civ, const_iv[] = { 826 { sizeof (cl_uchar ), "SIZEOF_UCHAR" },
82#define const_iv(name) { # name, (IV)CL_ ## name }, 827 { sizeof (cl_short ), "SIZEOF_SHORT" },
828 { sizeof (cl_ushort), "SIZEOF_USHORT" },
829 { sizeof (cl_int ), "SIZEOF_INT" },
830 { sizeof (cl_uint ), "SIZEOF_UINT" },
831 { sizeof (cl_long ), "SIZEOF_LONG" },
832 { sizeof (cl_ulong ), "SIZEOF_ULONG" },
833 { sizeof (cl_half ), "SIZEOF_HALF" },
834 { sizeof (cl_float ), "SIZEOF_FLOAT" },
835 { sizeof (cl_double), "SIZEOF_DOUBLE" },
836 { PREFER_1_1 , "PREFER_1_1" },
83#include "constiv.h" 837#include "constiv.h"
84 }; 838 };
839
85 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--)
86 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);
87} 877}
878
879cl_int
880errno ()
881 CODE:
882 RETVAL = res;
883 OUTPUT:
884 RETVAL
885
886const char *
887err2str (cl_int err = res)
888
889const char *
890enum2str (cl_uint value)
88 891
89void 892void
90platforms () 893platforms ()
91 PPCODE: 894 PPCODE:
92{
93 cl_platform_id *list; 895 cl_platform_id *list;
94 cl_uint count; 896 cl_uint count;
95 int i; 897 int i;
96 898
97 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count)); 899 NEED_SUCCESS (GetPlatformIDs, (0, 0, &count));
98 Newx (list, count, cl_platform_id); 900 list = tmpbuf (sizeof (*list) * count);
99 NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); 901 NEED_SUCCESS (GetPlatformIDs, (count, list, 0));
100 902
101 EXTEND (SP, count); 903 EXTEND (SP, count);
102 for (i = 0; i < count; ++i) 904 for (i = 0; i < count; ++i)
103 PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); 905 PUSH_CLOBJ (stash_platform, list [i]);
104 906
105 Safefree (list);
106}
107
108void 907void
109context_from_type_simple (cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 908context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
110 PPCODE: 909 PPCODE:
111{ 910 CONTEXT_NOTIFY_CALLBACK;
112 cl_int res; 911 NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res));
113 cl_context ctx = clCreateContextFromType (0, type, 0, 0, &res); 912 XPUSH_CLOBJ_CONTEXT;
114 913
115 if (res) 914void
116 FAIL (CreateContextFromType, res); 915context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef)
916 PPCODE:
917 cl_uint device_count;
918 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
117 919
118 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 920 CONTEXT_NOTIFY_CALLBACK;
119} 921 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res));
922 XPUSH_CLOBJ_CONTEXT;
120 923
121void 924void
122wait_for_events (...) 925wait_for_events (...)
123 CODE: 926 CODE:
124{
125 EVENT_LIST (0, items); 927 EVENT_LIST (0);
126 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); 928 NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr));
127}
128 929
129PROTOTYPES: DISABLE 930PROTOTYPES: DISABLE
130 931
131MODULE = OpenCL PACKAGE = OpenCL::Platform 932MODULE = OpenCL PACKAGE = OpenCL::Platform
132 933
133void 934void
134info (OpenCL::Platform this, cl_platform_info name) 935info (OpenCL::Platform self, cl_platform_info name)
135 PPCODE: 936 PPCODE:
136 INFO (Platform) 937 INFO (Platform)
137 938
138void 939void
940unload_compiler (OpenCL::Platform self)
941 CODE:
942#if CL_VERSION_1_2
943 clUnloadPlatformCompiler (self);
944#endif
945
946#BEGIN:platform
947
948void
949profile (OpenCL::Platform self)
950 ALIAS:
951 profile = CL_PLATFORM_PROFILE
952 version = CL_PLATFORM_VERSION
953 name = CL_PLATFORM_NAME
954 vendor = CL_PLATFORM_VENDOR
955 extensions = CL_PLATFORM_EXTENSIONS
956 PPCODE:
957 size_t size;
958 NEED_SUCCESS (GetPlatformInfo, (self, ix, 0, 0, &size));
959 char *value = tmpbuf (size);
960 NEED_SUCCESS (GetPlatformInfo, (self, ix, size, value, 0));
961 EXTEND (SP, 1);
962 const int i = 0;
963 PUSHs (sv_2mortal (newSVpv (value, 0)));
964
965#END:platform
966
967void
139devices (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_ALL) 968devices (OpenCL::Platform self, cl_device_type type = CL_DEVICE_TYPE_ALL)
140 PPCODE: 969 PPCODE:
141{
142 cl_device_id *list; 970 cl_device_id *list;
143 cl_uint count; 971 cl_uint count;
144 int i; 972 int i;
145 973
146 NEED_SUCCESS (GetDeviceIDs, (this, type, 0, 0, &count)); 974 NEED_SUCCESS (GetDeviceIDs, (self, type, 0, 0, &count));
147 Newx (list, count, cl_device_id); 975 list = tmpbuf (sizeof (*list) * count);
148 NEED_SUCCESS (GetDeviceIDs, (this, type, count, list, 0)); 976 NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0));
149 977
150 EXTEND (SP, count); 978 EXTEND (SP, count);
151 for (i = 0; i < count; ++i) 979 for (i = 0; i < count; ++i)
152 PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); 980 PUSH_CLOBJ (stash_device, list [i]);
153 981
154 Safefree (list);
155}
156
157void 982void
158context_from_type_simple (OpenCL::Platform this, cl_device_type type = CL_DEVICE_TYPE_DEFAULT) 983context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef)
159 PPCODE: 984 PPCODE:
160{
161 cl_int res;
162 cl_context_properties props[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)this, 0 }; 985 cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self };
163 cl_context ctx = clCreateContextFromType (props, type, 0, 0, &res); 986 cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2);
164 987
165 if (res) 988 cl_uint device_count;
166 FAIL (CreateContextFromType, res); 989 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
167 990
168 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 991 CONTEXT_NOTIFY_CALLBACK;
169} 992 NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res));
993 XPUSH_CLOBJ_CONTEXT;
170 994
171void 995void
172unload_compiler () 996context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef)
173 CODE: 997 PPCODE:
174 NEED_SUCCESS (UnloadCompiler, ()); 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;
175 1004
176MODULE = OpenCL PACKAGE = OpenCL::Device 1005MODULE = OpenCL PACKAGE = OpenCL::Device
177 1006
178void 1007void
179info (OpenCL::Device this, cl_device_info name) 1008info (OpenCL::Device self, cl_device_info name)
180 PPCODE: 1009 PPCODE:
181 INFO (Device) 1010 INFO (Device)
182 1011
1012#if CL_VERSION_1_2
1013
183void 1014void
184context_simple (OpenCL::Device this) 1015sub_devices (OpenCL::Device self, SV *properties)
1016 PPCODE:
1017 if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV)
1018 croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs");
1019
1020 properties = SvRV (properties);
1021
1022 cl_uint count = av_len ((AV *)properties) + 1;
1023 cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1);
1024
1025 int i;
1026 for (i = 0; i < count; ++i)
1027 props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0));
1028
1029 props [count] = 0;
1030
1031 cl_uint num_devices;
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));
1035
1036 EXTEND (SP, num_devices);
1037 for (i = 0; i < count; ++i)
1038 PUSH_CLOBJ (stash_subdevice, list [i]);
1039
1040#endif
1041
1042#BEGIN:device
1043
1044void
1045type (OpenCL::Device self)
1046 ALIAS:
1047 type = CL_DEVICE_TYPE
1048 address_bits = CL_DEVICE_ADDRESS_BITS
1049 max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE
1050 single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG
1051 global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE
1052 global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE
1053 max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE
1054 local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE
1055 execution_capabilities = CL_DEVICE_EXECUTION_CAPABILITIES
1056 properties = CL_DEVICE_QUEUE_PROPERTIES
1057 double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG
1058 half_fp_config = CL_DEVICE_HALF_FP_CONFIG
185 PPCODE: 1059 PPCODE:
186{ 1060 cl_ulong value [1];
187 cl_int res; 1061 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
188 cl_context ctx = clCreateContext (0, 1, &this, 0, 0, &res); 1062 EXTEND (SP, 1);
1063 const int i = 0;
1064 PUSHs (sv_2mortal (newSVuv (value [i])));
189 1065
190 if (res) 1066void
191 FAIL (CreateContext, res); 1067vendor_id (OpenCL::Device self)
1068 ALIAS:
1069 vendor_id = CL_DEVICE_VENDOR_ID
1070 max_compute_units = CL_DEVICE_MAX_COMPUTE_UNITS
1071 max_work_item_dimensions = CL_DEVICE_MAX_WORK_ITEM_DIMENSIONS
1072 preferred_vector_width_char = CL_DEVICE_PREFERRED_VECTOR_WIDTH_CHAR
1073 preferred_vector_width_short = CL_DEVICE_PREFERRED_VECTOR_WIDTH_SHORT
1074 preferred_vector_width_int = CL_DEVICE_PREFERRED_VECTOR_WIDTH_INT
1075 preferred_vector_width_long = CL_DEVICE_PREFERRED_VECTOR_WIDTH_LONG
1076 preferred_vector_width_float = CL_DEVICE_PREFERRED_VECTOR_WIDTH_FLOAT
1077 preferred_vector_width_double = CL_DEVICE_PREFERRED_VECTOR_WIDTH_DOUBLE
1078 max_clock_frequency = CL_DEVICE_MAX_CLOCK_FREQUENCY
1079 max_read_image_args = CL_DEVICE_MAX_READ_IMAGE_ARGS
1080 max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS
1081 image_support = CL_DEVICE_IMAGE_SUPPORT
1082 max_samplers = CL_DEVICE_MAX_SAMPLERS
1083 mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN
1084 min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE
1085 global_mem_cache_type = CL_DEVICE_GLOBAL_MEM_CACHE_TYPE
1086 global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE
1087 max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS
1088 local_mem_type = CL_DEVICE_LOCAL_MEM_TYPE
1089 preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF
1090 native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR
1091 native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT
1092 native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT
1093 native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG
1094 native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT
1095 native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE
1096 native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF
1097 reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT
1098 PPCODE:
1099 cl_uint value [1];
1100 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1101 EXTEND (SP, 1);
1102 const int i = 0;
1103 PUSHs (sv_2mortal (newSVuv (value [i])));
192 1104
193 XPUSH_NEW_OBJ ("OpenCL::Context", ctx); 1105void
194} 1106max_work_group_size (OpenCL::Device self)
1107 ALIAS:
1108 max_work_group_size = CL_DEVICE_MAX_WORK_GROUP_SIZE
1109 image2d_max_width = CL_DEVICE_IMAGE2D_MAX_WIDTH
1110 image2d_max_height = CL_DEVICE_IMAGE2D_MAX_HEIGHT
1111 image3d_max_width = CL_DEVICE_IMAGE3D_MAX_WIDTH
1112 image3d_max_height = CL_DEVICE_IMAGE3D_MAX_HEIGHT
1113 image3d_max_depth = CL_DEVICE_IMAGE3D_MAX_DEPTH
1114 max_parameter_size = CL_DEVICE_MAX_PARAMETER_SIZE
1115 profiling_timer_resolution = CL_DEVICE_PROFILING_TIMER_RESOLUTION
1116 PPCODE:
1117 size_t value [1];
1118 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1119 EXTEND (SP, 1);
1120 const int i = 0;
1121 PUSHs (sv_2mortal (newSVuv (value [i])));
1122
1123void
1124max_work_item_sizes (OpenCL::Device self)
1125 PPCODE:
1126 size_t size;
1127 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, 0, 0, &size));
1128 size_t *value = tmpbuf (size);
1129 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_MAX_WORK_ITEM_SIZES, size, value, 0));
1130 int i, n = size / sizeof (*value);
1131 EXTEND (SP, n);
1132 for (i = 0; i < n; ++i)
1133 PUSHs (sv_2mortal (newSVuv (value [i])));
1134
1135void
1136error_correction_support (OpenCL::Device self)
1137 ALIAS:
1138 error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT
1139 endian_little = CL_DEVICE_ENDIAN_LITTLE
1140 available = CL_DEVICE_AVAILABLE
1141 compiler_available = CL_DEVICE_COMPILER_AVAILABLE
1142 host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY
1143 PPCODE:
1144 cl_uint value [1];
1145 NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0));
1146 EXTEND (SP, 1);
1147 const int i = 0;
1148 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
1149
1150void
1151platform (OpenCL::Device self)
1152 PPCODE:
1153 cl_platform_id value [1];
1154 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0));
1155 EXTEND (SP, 1);
1156 const int i = 0;
1157 PUSH_CLOBJ (stash_platform, value [i]);
1158
1159void
1160name (OpenCL::Device self)
1161 ALIAS:
1162 name = CL_DEVICE_NAME
1163 vendor = CL_DEVICE_VENDOR
1164 driver_version = CL_DRIVER_VERSION
1165 profile = CL_DEVICE_PROFILE
1166 version = CL_DEVICE_VERSION
1167 extensions = CL_DEVICE_EXTENSIONS
1168 PPCODE:
1169 size_t size;
1170 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1171 char *value = tmpbuf (size);
1172 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1173 EXTEND (SP, 1);
1174 const int i = 0;
1175 PUSHs (sv_2mortal (newSVpv (value, 0)));
1176
1177void
1178parent_device_ext (OpenCL::Device self)
1179 PPCODE:
1180 cl_device_id value [1];
1181 NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0));
1182 EXTEND (SP, 1);
1183 const int i = 0;
1184 PUSH_CLOBJ (stash_device, value [i]);
1185
1186void
1187partition_types_ext (OpenCL::Device self)
1188 ALIAS:
1189 partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT
1190 affinity_domains_ext = CL_DEVICE_AFFINITY_DOMAINS_EXT
1191 partition_style_ext = CL_DEVICE_PARTITION_STYLE_EXT
1192 PPCODE:
1193 size_t size;
1194 NEED_SUCCESS (GetDeviceInfo, (self, ix, 0, 0, &size));
1195 cl_device_partition_property_ext *value = tmpbuf (size);
1196 NEED_SUCCESS (GetDeviceInfo, (self, ix, size, value, 0));
1197 int i, n = size / sizeof (*value);
1198 EXTEND (SP, n);
1199 for (i = 0; i < n; ++i)
1200 PUSHs (sv_2mortal (newSVuv (value [i])));
1201
1202#END:device
1203
1204MODULE = OpenCL PACKAGE = OpenCL::SubDevice
1205
1206#if CL_VERSION_1_2
1207
1208void
1209DESTROY (OpenCL::SubDevice self)
1210 CODE:
1211 clReleaseDevice (self);
1212
1213#endif
195 1214
196MODULE = OpenCL PACKAGE = OpenCL::Context 1215MODULE = OpenCL PACKAGE = OpenCL::Context
197 1216
198void 1217void
199DESTROY (OpenCL::Context context) 1218DESTROY (OpenCL::Context self)
200 CODE: 1219 CODE:
201 clReleaseContext (context); 1220 clReleaseContext (self);
202 1221
203void 1222void
204info (OpenCL::Context this, cl_context_info name) 1223info (OpenCL::Context self, cl_context_info name)
205 PPCODE: 1224 PPCODE:
206 INFO (Context) 1225 INFO (Context)
207 1226
208void 1227void
209command_queue_simple (OpenCL::Context this, OpenCL::Device device) 1228queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0)
210 PPCODE: 1229 PPCODE:
211{ 1230 NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res));
212 cl_int res; 1231 XPUSH_CLOBJ (stash_queue, queue);
213 cl_command_queue queue = clCreateCommandQueue (this, device, 0, &res);
214 1232
215 if (res)
216 FAIL (CreateCommandQueue, res);
217
218 XPUSH_NEW_OBJ ("OpenCL::Queue", queue);
219}
220
221void 1233void
1234user_event (OpenCL::Context self)
1235 PPCODE:
1236 NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res));
1237 XPUSH_CLOBJ (stash_userevent, ev);
1238
1239void
222buffer (OpenCL::Context this, cl_mem_flags flags, size_t len) 1240buffer (OpenCL::Context self, cl_mem_flags flags, size_t len)
223 PPCODE: 1241 PPCODE:
224{ 1242 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))
225 cl_int res; 1243 croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?");
226 cl_mem mem = clCreateBuffer (this, flags, len, 0, &res); 1244
1245 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res));
1246 XPUSH_CLOBJ (stash_bufferobj, mem);
227 1247
228 if (res)
229 FAIL (CreateBuffer, res);
230
231 XPUSH_NEW_OBJ ("OpenCL::Memory", mem);
232}
233
234void 1248void
235buffer_sv (OpenCL::Context this, cl_mem_flags flags, SV *data) 1249buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data)
236 PPCODE: 1250 PPCODE:
237{
238 STRLEN len; 1251 STRLEN len;
239 char *ptr = SvPVbyte (data, len); 1252 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
240 cl_int res; 1253 if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)))
241 cl_mem mem = clCreateBuffer (this, flags, len, ptr, &res); 1254 croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?");
1255 NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res));
1256 XPUSH_CLOBJ (stash_bufferobj, mem);
242 1257
243 if (res) 1258#if CL_VERSION_1_2
244 FAIL (CreateBuffer, res);
245 1259
246 XPUSH_NEW_OBJ ("OpenCL::Memory", mem);
247}
248
249void 1260void
1261image (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)
1262 PPCODE:
1263 STRLEN len;
1264 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1265 const cl_image_format format = { channel_order, channel_type };
1266 const cl_image_desc desc = {
1267 type,
1268 width, height, depth,
1269 array_size, row_pitch, slice_pitch,
1270 num_mip_level, num_samples,
1271 type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0
1272 };
1273 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1274 HV *stash = stash_image;
1275 switch (type)
1276 {
1277 case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break;
1278 case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break;
1279 case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break;
1280 case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break;
1281 case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break;
1282 case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break;
1283 }
1284 XPUSH_CLOBJ (stash, mem);
1285
1286#endif
1287
1288void
1289image2d (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)
1290 PPCODE:
1291 STRLEN len;
1292 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1293 const cl_image_format format = { channel_order, channel_type };
1294#if PREFER_1_1
1295 NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res));
1296#else
1297 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 };
1298 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1299#endif
1300 XPUSH_CLOBJ (stash_image2d, mem);
1301
1302void
1303image3d (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)
1304 PPCODE:
1305 STRLEN len;
1306 char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0;
1307 const cl_image_format format = { channel_order, channel_type };
1308#if PREFER_1_1
1309 NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res));
1310#else
1311 const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 };
1312 NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res));
1313#endif
1314 XPUSH_CLOBJ (stash_image3d, mem);
1315
1316#if cl_apple_gl_sharing || cl_khr_gl_sharing
1317
1318void
1319gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj)
1320 PPCODE:
1321 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res));
1322 XPUSH_CLOBJ (stash_bufferobj, mem);
1323
1324void
1325gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer)
1326 PPCODE:
1327 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res));
1328 XPUSH_CLOBJ (stash_image2d, mem);
1329
1330#if CL_VERSION_1_2
1331
1332void
1333gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1334 ALIAS:
1335 PPCODE:
1336 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res));
1337 cl_gl_object_type type;
1338 NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead?
1339 HV *stash = stash_memory;
1340 switch (type)
1341 {
1342 case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break;
1343 case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break;
1344 case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break;
1345 case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break;
1346 case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break;
1347 case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break;
1348 }
1349 XPUSH_CLOBJ (stash, mem);
1350
1351#endif
1352
1353void
1354gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1355 PPCODE:
1356#if PREFER_1_1
1357 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res));
1358#else
1359 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1360#endif
1361 XPUSH_CLOBJ (stash_image2d, mem);
1362
1363void
1364gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture)
1365 PPCODE:
1366#if PREFER_1_1
1367 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res));
1368#else
1369 NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res));
1370#endif
1371 XPUSH_CLOBJ (stash_image3d, mem);
1372
1373#endif
1374
1375void
1376supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type)
1377 PPCODE:
1378{
1379 cl_uint count;
1380 cl_image_format *list;
1381 int i;
1382
1383 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, 0, 0, &count));
1384 Newx (list, count, cl_image_format);
1385 NEED_SUCCESS (GetSupportedImageFormats, (self, flags, image_type, count, list, 0));
1386
1387 EXTEND (SP, count);
1388 for (i = 0; i < count; ++i)
1389 {
1390 AV *av = newAV ();
1391 av_store (av, 1, newSVuv (list [i].image_channel_data_type));
1392 av_store (av, 0, newSVuv (list [i].image_channel_order));
1393 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1394 }
1395}
1396
1397void
250sampler (OpenCL::Context this, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) 1398sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode)
251 PPCODE: 1399 PPCODE:
252{
253 cl_int res;
254 cl_sampler sampler = clCreateSampler (this, normalized_coords, addressing_mode, filter_mode, &res); 1400 NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res));
255
256 if (res)
257 FAIL (CreateSampler, res);
258
259 XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); 1401 XPUSH_CLOBJ (stash_sampler, sampler);
260}
261 1402
262void 1403void
263program_with_source (OpenCL::Context this, SV *program) 1404program_with_source (OpenCL::Context self, SV *program)
264 PPCODE: 1405 PPCODE:
265{
266 STRLEN len; 1406 STRLEN len;
267 size_t len2; 1407 size_t len2;
268 const char *ptr = SvPVbyte (program, len); 1408 const char *ptr = SvPVbyte (program, len);
269 cl_int res;
270 cl_program prog;
271 1409
272 len2 = len; 1410 len2 = len;
273 prog = clCreateProgramWithSource (this, 1, &ptr, &len2, &res); 1411 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res));
274
275 if (res)
276 FAIL (CreateProgramWithSource, res);
277
278 XPUSH_NEW_OBJ ("OpenCL::Program", prog); 1412 XPUSH_CLOBJ (stash_program, prog);
279} 1413
1414void
1415program_with_binary (OpenCL::Context self, SV *devices, SV *binaries)
1416 PPCODE:
1417 cl_uint device_count;
1418 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1419
1420 if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV)
1421 croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings");
1422
1423 binaries = SvRV (binaries);
1424
1425 if (device_count != av_len ((AV *)binaries) + 1)
1426 croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed");
1427
1428 size_t *length_list = tmpbuf (sizeof (*length_list) * device_count);
1429 const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count);
1430 cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count);
1431
1432 int i;
1433 for (i = 0; i < device_count; ++i)
1434 {
1435 STRLEN len;
1436 binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len);
1437 length_list [i] = len;
1438 }
1439
1440 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list,
1441 length_list, binary_list,
1442 GIMME_V == G_ARRAY ? status_list : 0, &res));
1443
1444 EXTEND (SP, 2);
1445 PUSH_CLOBJ (stash_program, prog);
1446
1447 if (GIMME_V == G_ARRAY)
1448 {
1449 AV *av = newAV ();
1450 PUSHs (sv_2mortal (newRV_noinc ((SV *)av)));
1451
1452 for (i = device_count; i--; )
1453 av_store (av, i, newSViv (status_list [i]));
1454 }
1455
1456#if CL_VERSION_1_2
1457
1458void
1459program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names)
1460 PPCODE:
1461 cl_uint device_count;
1462 cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count);
1463
1464 NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res));
1465
1466 XPUSH_CLOBJ (stash_program, prog);
1467
1468void
1469link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef)
1470 CODE:
1471 cl_uint device_count = 0;
1472 cl_device_id *device_list = 0;
1473
1474 if (SvOK (devices))
1475 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
1476
1477 cl_uint program_count;
1478 cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count);
1479
1480 void *user_data;
1481 program_callback pfn_notify = make_program_callback (notify, &user_data);
1482
1483 NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
1484 program_count, program_list, pfn_notify, user_data, &res));
1485
1486 XPUSH_CLOBJ (stash_program, prog);
1487
1488#endif
1489
1490#BEGIN:context
1491
1492void
1493reference_count (OpenCL::Context self)
1494 ALIAS:
1495 reference_count = CL_CONTEXT_REFERENCE_COUNT
1496 num_devices = CL_CONTEXT_NUM_DEVICES
1497 PPCODE:
1498 cl_uint value [1];
1499 NEED_SUCCESS (GetContextInfo, (self, ix, sizeof (value), value, 0));
1500 EXTEND (SP, 1);
1501 const int i = 0;
1502 PUSHs (sv_2mortal (newSVuv (value [i])));
1503
1504void
1505devices (OpenCL::Context self)
1506 PPCODE:
1507 size_t size;
1508 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, 0, 0, &size));
1509 cl_device_id *value = tmpbuf (size);
1510 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0));
1511 int i, n = size / sizeof (*value);
1512 EXTEND (SP, n);
1513 for (i = 0; i < n; ++i)
1514 PUSH_CLOBJ (stash_device, value [i]);
1515
1516void
1517properties (OpenCL::Context self)
1518 PPCODE:
1519 size_t size;
1520 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, 0, 0, &size));
1521 cl_context_properties *value = tmpbuf (size);
1522 NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_PROPERTIES, size, value, 0));
1523 int i, n = size / sizeof (*value);
1524 EXTEND (SP, n);
1525 for (i = 0; i < n; ++i)
1526 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
1527
1528#END:context
280 1529
281MODULE = OpenCL PACKAGE = OpenCL::Queue 1530MODULE = OpenCL PACKAGE = OpenCL::Queue
282 1531
283void 1532void
284DESTROY (OpenCL::Queue this) 1533DESTROY (OpenCL::Queue self)
285 CODE: 1534 CODE:
286 clReleaseCommandQueue (this); 1535 clReleaseCommandQueue (self);
287 1536
288void 1537void
289info (OpenCL::Queue this, cl_command_queue_info name)
290 PPCODE:
291 INFO (CommandQueue)
292
293void
294enqueue_read_buffer (OpenCL::Queue this, OpenCL::Memory mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) 1538read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...)
1539 ALIAS:
1540 enqueue_read_buffer = 0
295 PPCODE: 1541 PPCODE:
296{
297 cl_event ev = 0;
298 EVENT_LIST (6, items - 6); 1542 EVENT_LIST (6);
299 1543
300 SvUPGRADE (data, SVt_PV); 1544 SvUPGRADE (data, SVt_PV);
301 SvGROW (data, len); 1545 SvGROW (data, len);
302 SvPOK_only (data); 1546 SvPOK_only (data);
303 SvCUR_set (data, len); 1547 SvCUR_set (data, len);
1548
1549 cl_event ev = 0;
304 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1550 NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
305 1551
306 if (ev) 1552 if (ev)
307 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1553 XPUSH_CLOBJ (stash_event, ev);
308}
309 1554
310void 1555void
311enqueue_write_buffer (OpenCL::Queue this, OpenCL::Memory mem, cl_bool blocking, size_t offset, SV *data, ...) 1556write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...)
1557 ALIAS:
1558 enqueue_write_buffer = 0
312 PPCODE: 1559 PPCODE:
313{ 1560 EVENT_LIST (5);
314 cl_event ev = 0; 1561
315 STRLEN len; 1562 STRLEN len;
316 char *ptr = SvPVbyte (data, len); 1563 char *ptr = SvPVbyte (data, len);
317 EVENT_LIST (5, items - 5);
318 1564
1565 cl_event ev = 0;
319 NEED_SUCCESS (EnqueueReadBuffer, (this, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1566 NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
320 1567
321 if (ev) 1568 if (ev)
322 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1569 XPUSH_CLOBJ (stash_event, ev);
323}
324 1570
1571#if CL_VERSION_1_2
1572
325void 1573void
326enqueue_copy_buffer (OpenCL::Queue this, OpenCL::Memory src, OpenCL::Memory dst, size_t src_offset, size_t dst_offset, size_t len, ...) 1574fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...)
1575 ALIAS:
1576 enqueue_fill_buffer = 0
327 PPCODE: 1577 PPCODE:
328{ 1578 EVENT_LIST (5);
1579
1580 STRLEN len;
1581 char *ptr = SvPVbyte (data, len);
1582
329 cl_event ev = 0; 1583 cl_event ev = 0;
330 EVENT_LIST (6, items - 6);
331
332 NEED_SUCCESS (EnqueueCopyBuffer, (this, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); 1584 NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
333 1585
334 if (ev) 1586 if (ev)
335 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1587 XPUSH_CLOBJ (stash_event, ev);
336}
337 1588
338void 1589void
339enqueue_marker (OpenCL::Queue this) 1590fill_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, ...)
1591 ALIAS:
1592 enqueue_fill_image = 0
340 PPCODE: 1593 PPCODE:
341{ 1594 EVENT_LIST (12);
1595
1596 const size_t origin [3] = { x, y, z };
1597 const size_t region [3] = { width, height, depth };
1598
1599 const cl_float c_f [4] = { r, g, b, a };
1600 const cl_uint c_u [4] = { r, g, b, a };
1601 const cl_int c_s [4] = { r, g, b, a };
1602 const void *c_fus [3] = { &c_f, &c_u, &c_s };
1603 static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 };
1604 cl_image_format format;
1605 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
1606 assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8);
1607 if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type)
1608 croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported.");
1609
1610 cl_event ev = 0;
1611 NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]],
1612 origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1613
1614 if (ev)
1615 XPUSH_CLOBJ (stash_event, ev);
1616
1617#endif
1618
1619void
1620copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...)
1621 ALIAS:
1622 enqueue_copy_buffer = 0
1623 PPCODE:
1624 EVENT_LIST (6);
1625
1626 cl_event ev = 0;
1627 NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1628
1629 if (ev)
1630 XPUSH_CLOBJ (stash_event, ev);
1631
1632void
1633read_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, ...)
1634 ALIAS:
1635 enqueue_read_buffer_rect = 0
1636 PPCODE:
1637 EVENT_LIST (17);
1638
1639 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1640 const size_t host_origin[3] = { host_x, host_y, host_z };
1641 const size_t region[3] = { width, height, depth };
1642
1643 if (!buf_row_pitch)
1644 buf_row_pitch = region [0];
1645
1646 if (!buf_slice_pitch)
1647 buf_slice_pitch = region [1] * buf_row_pitch;
1648
1649 if (!host_row_pitch)
1650 host_row_pitch = region [0];
1651
1652 if (!host_slice_pitch)
1653 host_slice_pitch = region [1] * host_row_pitch;
1654
1655 size_t len = host_row_pitch * host_slice_pitch * region [2];
1656
1657 SvUPGRADE (data, SVt_PV);
1658 SvGROW (data, len);
1659 SvPOK_only (data);
1660 SvCUR_set (data, len);
1661
1662 cl_event ev = 0;
1663 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));
1664
1665 if (ev)
1666 XPUSH_CLOBJ (stash_event, ev);
1667
1668void
1669write_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, ...)
1670 ALIAS:
1671 enqueue_write_buffer_rect = 0
1672 PPCODE:
1673 EVENT_LIST (17);
1674
1675 const size_t buf_origin [3] = { buf_x , buf_y , buf_z };
1676 const size_t host_origin[3] = { host_x, host_y, host_z };
1677 const size_t region[3] = { width, height, depth };
1678 STRLEN len;
1679 char *ptr = SvPVbyte (data, len);
1680
1681 if (!buf_row_pitch)
1682 buf_row_pitch = region [0];
1683
1684 if (!buf_slice_pitch)
1685 buf_slice_pitch = region [1] * buf_row_pitch;
1686
1687 if (!host_row_pitch)
1688 host_row_pitch = region [0];
1689
1690 if (!host_slice_pitch)
1691 host_slice_pitch = region [1] * host_row_pitch;
1692
1693 size_t min_len = host_row_pitch * host_slice_pitch * region [2];
1694
1695 if (len < min_len)
1696 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1697
1698 cl_event ev = 0;
1699 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));
1700
1701 if (ev)
1702 XPUSH_CLOBJ (stash_event, ev);
1703
1704void
1705copy_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, ...)
1706 ALIAS:
1707 enqueue_copy_buffer_rect = 0
1708 PPCODE:
1709 EVENT_LIST (16);
1710
1711 const size_t src_origin[3] = { src_x, src_y, src_z };
1712 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1713 const size_t region[3] = { width, height, depth };
1714
1715 cl_event ev = 0;
1716 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));
1717
1718 if (ev)
1719 XPUSH_CLOBJ (stash_event, ev);
1720
1721void
1722read_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, ...)
1723 ALIAS:
1724 enqueue_read_image = 0
1725 PPCODE:
1726 EVENT_LIST (12);
1727
1728 const size_t src_origin[3] = { src_x, src_y, src_z };
1729 const size_t region[3] = { width, height, depth };
1730
1731 if (!row_pitch)
1732 row_pitch = img_row_pitch (src);
1733
1734 if (depth > 1 && !slice_pitch)
1735 slice_pitch = row_pitch * height;
1736
1737 size_t len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1738
1739 SvUPGRADE (data, SVt_PV);
1740 SvGROW (data, len);
1741 SvPOK_only (data);
1742 SvCUR_set (data, len);
1743
1744 cl_event ev = 0;
1745 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));
1746
1747 if (ev)
1748 XPUSH_CLOBJ (stash_event, ev);
1749
1750void
1751write_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, ...)
1752 ALIAS:
1753 enqueue_write_image = 0
1754 PPCODE:
1755 EVENT_LIST (12);
1756
1757 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1758 const size_t region[3] = { width, height, depth };
1759 STRLEN len;
1760 char *ptr = SvPVbyte (data, len);
1761
1762 if (!row_pitch)
1763 row_pitch = img_row_pitch (dst);
1764
1765 if (depth > 1 && !slice_pitch)
1766 slice_pitch = row_pitch * height;
1767
1768 size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height;
1769
1770 if (len < min_len)
1771 croak ("clEnqueueWriteImage: data string is shorter than what would be transferred");
1772
1773 cl_event ev = 0;
1774 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));
1775
1776 if (ev)
1777 XPUSH_CLOBJ (stash_event, ev);
1778
1779void
1780copy_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, ...)
1781 ALIAS:
1782 enqueue_copy_image = 0
1783 PPCODE:
1784 EVENT_LIST (12);
1785
1786 const size_t src_origin[3] = { src_x, src_y, src_z };
1787 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1788 const size_t region[3] = { width, height, depth };
1789
1790 cl_event ev = 0;
1791 NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1792
1793 if (ev)
1794 XPUSH_CLOBJ (stash_event, ev);
1795
1796void
1797copy_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, ...)
1798 ALIAS:
1799 enqueue_copy_image_to_buffer = 0
1800 PPCODE:
1801 EVENT_LIST (10);
1802
1803 const size_t src_origin[3] = { src_x, src_y, src_z };
1804 const size_t region [3] = { width, height, depth };
1805
1806 cl_event ev = 0;
1807 NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1808
1809 if (ev)
1810 XPUSH_CLOBJ (stash_event, ev);
1811
1812void
1813copy_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, ...)
1814 ALIAS:
1815 enqueue_copy_buffer_to_image = 0
1816 PPCODE:
1817 EVENT_LIST (10);
1818
1819 const size_t dst_origin[3] = { dst_x, dst_y, dst_z };
1820 const size_t region [3] = { width, height, depth };
1821
1822 cl_event ev = 0;
1823 NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1824
1825 if (ev)
1826 XPUSH_CLOBJ (stash_event, ev);
1827
1828void
1829map_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, ...)
1830 ALIAS:
1831 enqueue_map_buffer = 0
1832 PPCODE:
1833 EVENT_LIST (6);
1834
1835 size_t cb = SvIV (cb_);
1836
1837 if (!SvOK (cb_))
1838 {
1839 NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0));
1840 cb -= offset;
1841 }
1842
342 cl_event ev; 1843 cl_event ev;
343 NEED_SUCCESS (EnqueueMarker, (this, &ev)); 1844 NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res));
344 XPUSH_NEW_OBJ ("OpenCL::Event", ev); 1845 XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0));
345}
346 1846
347void 1847void
348enqueue_wait_for_events (OpenCL::Queue this, ...) 1848map_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, ...)
1849 ALIAS:
1850 enqueue_map_image = 0
1851 PPCODE:
1852 size_t width = SvIV (width_);
1853 if (!SvOK (width_))
1854 {
1855 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0));
1856 width -= x;
1857 }
1858
1859 size_t height = SvIV (width_);
1860 if (!SvOK (height_))
1861 {
1862 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0));
1863 height -= y;
1864 }
1865
1866 size_t depth = SvIV (width_);
1867 if (!SvOK (depth_))
1868 {
1869 NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0));
1870 depth -= z;
1871
1872 // stupid opencl returns 0 for depth, but requires 1 for 2d images
1873 if (!depth)
1874 depth = 1;
1875 }
1876
1877 const size_t origin[3] = { x, y, z };
1878 const size_t region[3] = { width, height, depth };
1879 size_t row_pitch, slice_pitch;
1880 EVENT_LIST (10);
1881
1882 cl_event ev;
1883 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));
1884
1885 size_t cb = slice_pitch ? slice_pitch * region [2]
1886 : row_pitch ? row_pitch * region [1]
1887 : region [0];
1888
1889 XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch));
1890
1891void
1892unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...)
349 CODE: 1893 PPCODE:
350{ 1894 mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2);
1895 if (GIMME_V != G_VOID)
1896 {
1897 clRetainEvent (mapped->event);
1898 XPUSH_CLOBJ (stash_event, mapped->event);
1899 }
1900
1901void
1902task (OpenCL::Queue self, OpenCL::Kernel kernel, ...)
1903 ALIAS:
1904 enqueue_task = 0
1905 PPCODE:
1906 EVENT_LIST (2);
1907
1908 cl_event ev = 0;
1909 NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1910
1911 if (ev)
1912 XPUSH_CLOBJ (stash_event, ev);
1913
1914void
1915nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...)
1916 ALIAS:
1917 enqueue_nd_range_kernel = 0
1918 PPCODE:
1919 EVENT_LIST (5);
1920
1921 size_t *gwo = 0, *gws, *lws = 0;
1922 int gws_len;
1923 size_t *lists;
1924 int i;
1925
1926 if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV)
1927 croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference");
1928
1929 gws_len = AvFILLp (SvRV (global_work_size)) + 1;
1930
1931 lists = tmpbuf (sizeof (size_t) * 3 * gws_len);
1932
1933 gws = lists + gws_len * 0;
1934 for (i = 0; i < gws_len; ++i)
1935 {
1936 gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]);
1937 // at least nvidia crashes for 0-sized work group sizes, work around
1938 if (!gws [i])
1939 croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i);
1940 }
1941
1942 if (SvOK (global_work_offset))
1943 {
1944 if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV)
1945 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference");
1946
1947 if (AvFILLp (SvRV (global_work_size)) + 1 != gws_len)
1948 croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array of same size as global_work_size");
1949
1950 gwo = lists + gws_len * 1;
1951 for (i = 0; i < gws_len; ++i)
1952 gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]);
1953 }
1954
1955 if (SvOK (local_work_size))
1956 {
1957 if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV)
1958 croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference");
1959
1960 if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len)
1961 croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size");
1962
1963 lws = lists + gws_len * 2;
1964 for (i = 0; i < gws_len; ++i)
1965 {
1966 lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]);
1967 // at least nvidia crashes for 0-sized work group sizes, work around
1968 if (!lws [i])
1969 croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i);
1970 }
1971 }
1972
1973 cl_event ev = 0;
1974 NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1975
1976 if (ev)
1977 XPUSH_CLOBJ (stash_event, ev);
1978
1979#if CL_VERSION_1_2
1980
1981void
1982migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...)
1983 ALIAS:
1984 enqueue_migrate_mem_objects = 0
1985 PPCODE:
1986 EVENT_LIST (3);
1987
1988 cl_uint obj_count;
1989 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
1990
1991 cl_event ev = 0;
1992 NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
1993
1994 if (ev)
1995 XPUSH_CLOBJ (stash_event, ev);
1996
1997#endif
1998
1999#if cl_apple_gl_sharing || cl_khr_gl_sharing
2000
2001void
2002acquire_gl_objects (OpenCL::Queue self, SV *objects, ...)
2003 ALIAS:
2004 release_gl_objects = 1
2005 enqueue_acquire_gl_objects = 0
2006 enqueue_release_gl_objects = 1
2007 PPCODE:
2008 EVENT_LIST (2);
2009
2010 cl_uint obj_count;
2011 cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count);
2012
2013 cl_event ev = 0;
2014
2015 if (ix)
2016 NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2017 else
2018 NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2019
2020 if (ev)
2021 XPUSH_CLOBJ (stash_event, ev);
2022
2023#endif
2024
2025void
2026wait_for_events (OpenCL::Queue self, ...)
2027 ALIAS:
2028 enqueue_wait_for_events = 0
2029 CODE:
351 EVENT_LIST (1, items - 1); 2030 EVENT_LIST (1);
2031#if PREFER_1_1
352 NEED_SUCCESS (EnqueueWaitForEvents, (this, event_list_count, event_list_ptr)); 2032 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
353} 2033#else
2034 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0));
2035#endif
354 2036
355void 2037void
356enqueue_barrier (OpenCL::Queue this) 2038marker (OpenCL::Queue self, ...)
2039 ALIAS:
2040 enqueue_marker = 0
357 CODE: 2041 PPCODE:
2042 EVENT_LIST (1);
2043 cl_event ev = 0;
2044#if PREFER_1_1
2045 if (!event_list_count)
2046 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2047 else
2048#if CL_VERSION_1_2
2049 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2050#else
2051 {
2052 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier
2053 NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0));
2054 }
2055#endif
2056#else
2057 NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2058#endif
2059 if (ev)
2060 XPUSH_CLOBJ (stash_event, ev);
2061
2062void
2063barrier (OpenCL::Queue self, ...)
2064 ALIAS:
2065 enqueue_barrier = 0
2066 PPCODE:
2067 EVENT_LIST (1);
2068 cl_event ev = 0;
2069#if PREFER_1_1
2070 if (!event_list_count && GIMME_V == G_VOID)
358 NEED_SUCCESS (EnqueueBarrier, (this)); 2071 NEED_SUCCESS (EnqueueBarrier, (self));
2072 else
2073#if CL_VERSION_1_2
2074 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2075#else
2076 {
2077 if (event_list_count)
2078 NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr));
2079
2080 if (GIMME_V != G_VOID)
2081 NEED_SUCCESS (EnqueueMarker, (self, &ev));
2082 }
2083#endif
2084#else
2085 NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0));
2086#endif
2087 if (ev)
2088 XPUSH_CLOBJ (stash_event, ev);
2089
2090void
2091flush (OpenCL::Queue self)
2092 CODE:
2093 NEED_SUCCESS (Flush, (self));
2094
2095void
2096finish (OpenCL::Queue self)
2097 CODE:
2098 NEED_SUCCESS (Finish, (self));
2099
2100void
2101info (OpenCL::Queue self, cl_command_queue_info name)
2102 PPCODE:
2103 INFO (CommandQueue)
2104
2105#BEGIN:command_queue
2106
2107void
2108context (OpenCL::Queue self)
2109 PPCODE:
2110 cl_context value [1];
2111 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0));
2112 EXTEND (SP, 1);
2113 const int i = 0;
2114 NEED_SUCCESS (RetainContext, (value [i]));
2115 PUSH_CLOBJ (stash_context, value [i]);
2116
2117void
2118device (OpenCL::Queue self)
2119 PPCODE:
2120 cl_device_id value [1];
2121 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0));
2122 EXTEND (SP, 1);
2123 const int i = 0;
2124 PUSH_CLOBJ (stash_device, value [i]);
2125
2126void
2127reference_count (OpenCL::Queue self)
2128 PPCODE:
2129 cl_uint value [1];
2130 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_REFERENCE_COUNT, sizeof (value), value, 0));
2131 EXTEND (SP, 1);
2132 const int i = 0;
2133 PUSHs (sv_2mortal (newSVuv (value [i])));
2134
2135void
2136properties (OpenCL::Queue self)
2137 PPCODE:
2138 cl_ulong value [1];
2139 NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0));
2140 EXTEND (SP, 1);
2141 const int i = 0;
2142 PUSHs (sv_2mortal (newSVuv (value [i])));
2143
2144#END:command_queue
359 2145
360MODULE = OpenCL PACKAGE = OpenCL::Memory 2146MODULE = OpenCL PACKAGE = OpenCL::Memory
361 2147
362void 2148void
363DESTROY (OpenCL::Memory this) 2149DESTROY (OpenCL::Memory self)
364 CODE: 2150 CODE:
365 clReleaseMemObject (this); 2151 clReleaseMemObject (self);
366 2152
367void 2153void
368info (OpenCL::Memory this, cl_mem_info name) 2154info (OpenCL::Memory self, cl_mem_info name)
369 PPCODE: 2155 PPCODE:
370 INFO (MemObject) 2156 INFO (MemObject)
371 2157
2158void
2159destructor_callback (OpenCL::Memory self, SV *notify)
2160 PPCODE:
2161 clSetMemObjectDestructorCallback (self, eq_destructor_notify, SvREFCNT_inc (s_get_cv (notify)));
2162
2163#BEGIN:mem
2164
2165void
2166type (OpenCL::Memory self)
2167 ALIAS:
2168 type = CL_MEM_TYPE
2169 map_count = CL_MEM_MAP_COUNT
2170 reference_count = CL_MEM_REFERENCE_COUNT
2171 PPCODE:
2172 cl_uint value [1];
2173 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
2174 EXTEND (SP, 1);
2175 const int i = 0;
2176 PUSHs (sv_2mortal (newSVuv (value [i])));
2177
2178void
2179flags (OpenCL::Memory self)
2180 PPCODE:
2181 cl_ulong value [1];
2182 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0));
2183 EXTEND (SP, 1);
2184 const int i = 0;
2185 PUSHs (sv_2mortal (newSVuv (value [i])));
2186
2187void
2188size (OpenCL::Memory self)
2189 ALIAS:
2190 size = CL_MEM_SIZE
2191 offset = CL_MEM_OFFSET
2192 PPCODE:
2193 size_t value [1];
2194 NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0));
2195 EXTEND (SP, 1);
2196 const int i = 0;
2197 PUSHs (sv_2mortal (newSVuv (value [i])));
2198
2199void
2200host_ptr (OpenCL::Memory self)
2201 PPCODE:
2202 void * value [1];
2203 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_HOST_PTR, sizeof (value), value, 0));
2204 EXTEND (SP, 1);
2205 const int i = 0;
2206 PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i])));
2207
2208void
2209context (OpenCL::Memory self)
2210 PPCODE:
2211 cl_context value [1];
2212 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0));
2213 EXTEND (SP, 1);
2214 const int i = 0;
2215 NEED_SUCCESS (RetainContext, (value [i]));
2216 PUSH_CLOBJ (stash_context, value [i]);
2217
2218void
2219associated_memobject (OpenCL::Memory self)
2220 PPCODE:
2221 cl_mem value [1];
2222 NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0));
2223 EXTEND (SP, 1);
2224 const int i = 0;
2225 NEED_SUCCESS (RetainMemObject, (value [i]));
2226 PUSH_CLOBJ (stash_memory, value [i]);
2227
2228#END:mem
2229
2230#if cl_apple_gl_sharing || cl_khr_gl_sharing
2231
2232void
2233gl_object_info (OpenCL::Memory self)
2234 PPCODE:
2235 cl_gl_object_type type;
2236 cl_GLuint name;
2237 NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name));
2238 EXTEND (SP, 2);
2239 PUSHs (sv_2mortal (newSVuv (type)));
2240 PUSHs (sv_2mortal (newSVuv (name)));
2241
2242#endif
2243
2244MODULE = OpenCL PACKAGE = OpenCL::BufferObj
2245
2246void
2247sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size)
2248 PPCODE:
2249 if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR | CL_MEM_ALLOC_HOST_PTR))
2250 croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!");
2251
2252 cl_buffer_region crdata = { origin, size };
2253
2254 NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res));
2255 XPUSH_CLOBJ (stash_buffer, mem);
2256
2257MODULE = OpenCL PACKAGE = OpenCL::Image
2258
2259void
2260image_info (OpenCL::Image self, cl_image_info name)
2261 PPCODE:
2262 INFO (Image)
2263
2264void
2265format (OpenCL::Image self)
2266 PPCODE:
2267 cl_image_format format;
2268 NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0));
2269 EXTEND (SP, 2);
2270 PUSHs (sv_2mortal (newSVuv (format.image_channel_order)));
2271 PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type)));
2272
2273#BEGIN:image
2274
2275void
2276element_size (OpenCL::Image self)
2277 ALIAS:
2278 element_size = CL_IMAGE_ELEMENT_SIZE
2279 row_pitch = CL_IMAGE_ROW_PITCH
2280 slice_pitch = CL_IMAGE_SLICE_PITCH
2281 width = CL_IMAGE_WIDTH
2282 height = CL_IMAGE_HEIGHT
2283 depth = CL_IMAGE_DEPTH
2284 PPCODE:
2285 size_t value [1];
2286 NEED_SUCCESS (GetImageInfo, (self, ix, sizeof (value), value, 0));
2287 EXTEND (SP, 1);
2288 const int i = 0;
2289 PUSHs (sv_2mortal (newSVuv (value [i])));
2290
2291#END:image
2292
2293#if cl_apple_gl_sharing || cl_khr_gl_sharing
2294
2295#BEGIN:gl_texture
2296
2297void
2298target (OpenCL::Image self)
2299 PPCODE:
2300 cl_GLenum value [1];
2301 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0));
2302 EXTEND (SP, 1);
2303 const int i = 0;
2304 PUSHs (sv_2mortal (newSVuv (value [i])));
2305
2306void
2307gl_mipmap_level (OpenCL::Image self)
2308 PPCODE:
2309 cl_GLint value [1];
2310 NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0));
2311 EXTEND (SP, 1);
2312 const int i = 0;
2313 PUSHs (sv_2mortal (newSViv (value [i])));
2314
2315#END:gl_texture
2316
2317#endif
2318
372MODULE = OpenCL PACKAGE = OpenCL::Sampler 2319MODULE = OpenCL PACKAGE = OpenCL::Sampler
373 2320
374void 2321void
375DESTROY (OpenCL::Sampler this) 2322DESTROY (OpenCL::Sampler self)
376 CODE: 2323 CODE:
377 clReleaseSampler (this); 2324 clReleaseSampler (self);
378 2325
379void 2326void
380info (OpenCL::Sampler this, cl_sampler_info name) 2327info (OpenCL::Sampler self, cl_sampler_info name)
381 PPCODE: 2328 PPCODE:
382 INFO (Sampler) 2329 INFO (Sampler)
383 2330
2331#BEGIN:sampler
2332
2333void
2334reference_count (OpenCL::Sampler self)
2335 ALIAS:
2336 reference_count = CL_SAMPLER_REFERENCE_COUNT
2337 normalized_coords = CL_SAMPLER_NORMALIZED_COORDS
2338 addressing_mode = CL_SAMPLER_ADDRESSING_MODE
2339 PPCODE:
2340 cl_uint value [1];
2341 NEED_SUCCESS (GetSamplerInfo, (self, ix, sizeof (value), value, 0));
2342 EXTEND (SP, 1);
2343 const int i = 0;
2344 PUSHs (sv_2mortal (newSVuv (value [i])));
2345
2346void
2347context (OpenCL::Sampler self)
2348 PPCODE:
2349 cl_context value [1];
2350 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0));
2351 EXTEND (SP, 1);
2352 const int i = 0;
2353 NEED_SUCCESS (RetainContext, (value [i]));
2354 PUSH_CLOBJ (stash_context, value [i]);
2355
2356void
2357filter_mode (OpenCL::Sampler self)
2358 PPCODE:
2359 cl_uint value [1];
2360 NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0));
2361 EXTEND (SP, 1);
2362 const int i = 0;
2363 PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no));
2364
2365#END:sampler
2366
384MODULE = OpenCL PACKAGE = OpenCL::Program 2367MODULE = OpenCL PACKAGE = OpenCL::Program
385 2368
386void 2369void
387DESTROY (OpenCL::Program this) 2370DESTROY (OpenCL::Program self)
388 CODE: 2371 CODE:
389 clReleaseProgram (this); 2372 clReleaseProgram (self);
390 2373
391void 2374void
392info (OpenCL::Program this, cl_program_info name) 2375build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef)
2376 ALIAS:
2377 build_async = 1
393 PPCODE: 2378 CODE:
394 INFO (Program) 2379 cl_uint device_count = 0;
2380 cl_device_id *device_list = 0;
395 2381
2382 if (SvOK (devices))
2383 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2384
2385 void *user_data;
2386 program_callback pfn_notify = make_program_callback (notify, &user_data);
2387
2388 if (ix)
2389 build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data);
2390 else
2391 NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data));
2392
2393#if CL_VERSION_1_2
2394
396void 2395void
397build (OpenCL::Program this, OpenCL::Device device, SV *options = &PL_sv_undef) 2396compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef)
398 CODE: 2397 CODE:
2398 cl_uint device_count = 0;
2399 cl_device_id *device_list = 0;
2400
2401 if (SvOK (devices))
2402 device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count);
2403
2404 cl_uint header_count = 0;
2405 cl_program *header_list = 0;
2406 const char **header_name = 0;
2407
2408 if (SvOK (headers))
2409 {
2410 if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV)
2411 croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs");
2412
2413 HV *hv = (HV *)SvRV (devices);
2414
2415 header_count = hv_iterinit (hv);
2416 header_list = tmpbuf (sizeof (*header_list) * header_count);
2417 header_name = tmpbuf (sizeof (*header_name) * header_count);
2418
2419 HE *he;
2420 int i = 0;
2421 while (he = hv_iternext (hv))
2422 {
2423 header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he));
2424 header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program");
2425 ++i;
2426 }
2427 }
2428
2429 void *user_data;
2430 program_callback pfn_notify = make_program_callback (notify, &user_data);
2431
399 NEED_SUCCESS (BuildProgram, (this, 1, &device, SvPVbyte_nolen (options), 0, 0)); 2432 NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options),
2433 header_count, header_list, header_name, pfn_notify, user_data));
400 2434
2435#endif
2436
401void 2437void
402build_info (OpenCL::Program this, OpenCL::Device device, cl_program_build_info name) 2438build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name)
403 PPCODE: 2439 PPCODE:
404{
405 size_t size; 2440 size_t size;
406 SV *sv;
407
408 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, 0, 0, &size)); 2441 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, 0, 0, &size));
409 sv = sv_2mortal (newSV (size)); 2442 SV *sv = sv_2mortal (newSV (size));
410 SvUPGRADE (sv, SVt_PV); 2443 SvUPGRADE (sv, SVt_PV);
411 SvPOK_only (sv); 2444 SvPOK_only (sv);
412 SvCUR_set (sv, size); 2445 SvCUR_set (sv, size);
413 NEED_SUCCESS (GetProgramBuildInfo, (this, device, name, size, SvPVX (sv), 0)); 2446 NEED_SUCCESS (GetProgramBuildInfo, (self, device, name, size, SvPVX (sv), 0));
414 XPUSHs (sv); 2447 XPUSHs (sv);
415} 2448
2449#BEGIN:program_build
2450
2451void
2452build_status (OpenCL::Program self, OpenCL::Device device)
2453 PPCODE:
2454 cl_int value [1];
2455 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0));
2456 EXTEND (SP, 1);
2457 const int i = 0;
2458 PUSHs (sv_2mortal (newSViv (value [i])));
2459
2460void
2461build_options (OpenCL::Program self, OpenCL::Device device)
2462 ALIAS:
2463 build_options = CL_PROGRAM_BUILD_OPTIONS
2464 build_log = CL_PROGRAM_BUILD_LOG
2465 PPCODE:
2466 size_t size;
2467 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, 0, 0, &size));
2468 char *value = tmpbuf (size);
2469 NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0));
2470 EXTEND (SP, 1);
2471 const int i = 0;
2472 PUSHs (sv_2mortal (newSVpv (value, 0)));
2473
2474void
2475binary_type (OpenCL::Program self, OpenCL::Device device)
2476 PPCODE:
2477 cl_uint value [1];
2478 NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0));
2479 EXTEND (SP, 1);
2480 const int i = 0;
2481 PUSHs (sv_2mortal (newSVuv ((UV)value [i])));
2482
2483#END:program_build
416 2484
417void 2485void
418kernel (OpenCL::Program program, SV *function) 2486kernel (OpenCL::Program program, SV *function)
419 PPCODE: 2487 PPCODE:
420{
421 cl_int res;
422 cl_kernel kernel = clCreateKernel (program, SvPVbyte_nolen (function), &res); 2488 NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res));
2489 XPUSH_CLOBJ (stash_kernel, kernel);
423 2490
424 if (res) 2491void
425 FAIL (CreateKernel, res); 2492kernels_in_program (OpenCL::Program program)
2493 PPCODE:
2494 cl_uint num_kernels;
2495 NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels));
2496 cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels);
2497 NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0));
426 2498
427 XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); 2499 int i;
428} 2500 EXTEND (SP, num_kernels);
2501 for (i = 0; i < num_kernels; ++i)
2502 PUSH_CLOBJ (stash_kernel, kernels [i]);
2503
2504void
2505info (OpenCL::Program self, cl_program_info name)
2506 PPCODE:
2507 INFO (Program)
2508
2509void
2510binaries (OpenCL::Program self)
2511 PPCODE:
2512 cl_uint n, i;
2513 size_t size;
2514
2515 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_NUM_DEVICES , sizeof (n) , &n , 0));
2516 if (!n) XSRETURN_EMPTY;
2517
2518 size_t *sizes = tmpbuf (sizeof (*sizes) * n);
2519 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, sizeof (*sizes) * n, sizes, &size));
2520 if (size != sizeof (*sizes) * n) XSRETURN_EMPTY;
2521 unsigned char **ptrs = tmpbuf (sizeof (*ptrs) * n);
2522
2523 EXTEND (SP, n);
2524 for (i = 0; i < n; ++i)
2525 {
2526 SV *sv = sv_2mortal (newSV (sizes [i]));
2527 SvUPGRADE (sv, SVt_PV);
2528 SvPOK_only (sv);
2529 SvCUR_set (sv, sizes [i]);
2530 ptrs [i] = (void *)SvPVX (sv);
2531 PUSHs (sv);
2532 }
2533
2534 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size));
2535 if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY;
2536
2537#BEGIN:program
2538
2539void
2540reference_count (OpenCL::Program self)
2541 ALIAS:
2542 reference_count = CL_PROGRAM_REFERENCE_COUNT
2543 num_devices = CL_PROGRAM_NUM_DEVICES
2544 PPCODE:
2545 cl_uint value [1];
2546 NEED_SUCCESS (GetProgramInfo, (self, ix, sizeof (value), value, 0));
2547 EXTEND (SP, 1);
2548 const int i = 0;
2549 PUSHs (sv_2mortal (newSVuv (value [i])));
2550
2551void
2552context (OpenCL::Program self)
2553 PPCODE:
2554 cl_context value [1];
2555 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0));
2556 EXTEND (SP, 1);
2557 const int i = 0;
2558 NEED_SUCCESS (RetainContext, (value [i]));
2559 PUSH_CLOBJ (stash_context, value [i]);
2560
2561void
2562devices (OpenCL::Program self)
2563 PPCODE:
2564 size_t size;
2565 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, 0, 0, &size));
2566 cl_device_id *value = tmpbuf (size);
2567 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0));
2568 int i, n = size / sizeof (*value);
2569 EXTEND (SP, n);
2570 for (i = 0; i < n; ++i)
2571 PUSH_CLOBJ (stash_device, value [i]);
2572
2573void
2574source (OpenCL::Program self)
2575 PPCODE:
2576 size_t size;
2577 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, 0, 0, &size));
2578 char *value = tmpbuf (size);
2579 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_SOURCE, size, value, 0));
2580 EXTEND (SP, 1);
2581 const int i = 0;
2582 PUSHs (sv_2mortal (newSVpv (value, 0)));
2583
2584void
2585binary_sizes (OpenCL::Program self)
2586 PPCODE:
2587 size_t size;
2588 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, 0, 0, &size));
2589 size_t *value = tmpbuf (size);
2590 NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARY_SIZES, size, value, 0));
2591 int i, n = size / sizeof (*value);
2592 EXTEND (SP, n);
2593 for (i = 0; i < n; ++i)
2594 PUSHs (sv_2mortal (newSVuv (value [i])));
2595
2596#END:program
429 2597
430MODULE = OpenCL PACKAGE = OpenCL::Kernel 2598MODULE = OpenCL PACKAGE = OpenCL::Kernel
431 2599
432void 2600void
433DESTROY (OpenCL::Kernel this) 2601DESTROY (OpenCL::Kernel self)
434 CODE: 2602 CODE:
435 clReleaseKernel (this); 2603 clReleaseKernel (self);
436 2604
437void 2605void
2606setf (OpenCL::Kernel self, const char *format, ...)
2607 CODE:
2608 int i;
2609 for (i = 2; ; ++i)
2610 {
2611 while (*format == ' ')
2612 ++format;
2613
2614 char type = *format++;
2615
2616 if (!type)
2617 break;
2618
2619 if (i >= items)
2620 croak ("OpenCL::Kernel::setf format string too long (not enough arguments)");
2621
2622 SV *sv = ST (i);
2623
2624 union
2625 {
2626 cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS;
2627 cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL;
2628 cl_half ch; cl_float cf; cl_double cd;
2629 cl_mem cm;
2630 cl_sampler ca;
2631 size_t cz;
2632 cl_event ce;
2633 } arg;
2634 size_t size;
2635 int nullarg = 0;
2636
2637 switch (type)
2638 {
2639 case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break;
2640 case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break;
2641 case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break;
2642 case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break;
2643 case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break;
2644 case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break;
2645 case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break;
2646 case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break;
2647
2648 case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break;
2649 case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break;
2650 case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break;
2651
2652 case 'z': nullarg = 1; size = SvIV (sv); break;
2653
2654 case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break;
2655 case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break;
2656 case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break;
2657
2658 default:
2659 croak ("OpenCL::Kernel::setf format character '%c' not supported", type);
2660 }
2661
2662 res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg);
2663 if (res)
2664 croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res));
2665 }
2666
2667 if (i != items)
2668 croak ("OpenCL::Kernel::setf format string too short (too many arguments)");
2669
2670void
2671set_char (OpenCL::Kernel self, cl_uint idx, cl_char value)
2672 CODE:
2673 clSetKernelArg (self, idx, sizeof (value), &value);
2674
2675void
2676set_uchar (OpenCL::Kernel self, cl_uint idx, cl_uchar value)
2677 CODE:
2678 clSetKernelArg (self, idx, sizeof (value), &value);
2679
2680void
2681set_short (OpenCL::Kernel self, cl_uint idx, cl_short value)
2682 CODE:
2683 clSetKernelArg (self, idx, sizeof (value), &value);
2684
2685void
2686set_ushort (OpenCL::Kernel self, cl_uint idx, cl_ushort value)
2687 CODE:
2688 clSetKernelArg (self, idx, sizeof (value), &value);
2689
2690void
2691set_int (OpenCL::Kernel self, cl_uint idx, cl_int value)
2692 CODE:
2693 clSetKernelArg (self, idx, sizeof (value), &value);
2694
2695void
2696set_uint (OpenCL::Kernel self, cl_uint idx, cl_uint value)
2697 CODE:
2698 clSetKernelArg (self, idx, sizeof (value), &value);
2699
2700void
2701set_long (OpenCL::Kernel self, cl_uint idx, cl_long value)
2702 CODE:
2703 clSetKernelArg (self, idx, sizeof (value), &value);
2704
2705void
2706set_ulong (OpenCL::Kernel self, cl_uint idx, cl_ulong value)
2707 CODE:
2708 clSetKernelArg (self, idx, sizeof (value), &value);
2709
2710void
2711set_half (OpenCL::Kernel self, cl_uint idx, cl_half value)
2712 CODE:
2713 clSetKernelArg (self, idx, sizeof (value), &value);
2714
2715void
2716set_float (OpenCL::Kernel self, cl_uint idx, cl_float value)
2717 CODE:
2718 clSetKernelArg (self, idx, sizeof (value), &value);
2719
2720void
2721set_double (OpenCL::Kernel self, cl_uint idx, cl_double value)
2722 CODE:
2723 clSetKernelArg (self, idx, sizeof (value), &value);
2724
2725void
2726set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value)
2727 CODE:
2728 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2729
2730void
2731set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value)
2732 CODE:
2733 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2734
2735void
2736set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value)
2737 CODE:
2738 clSetKernelArg (self, idx, sizeof (value), value ? &value : 0);
2739
2740void
2741set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value)
2742 CODE:
2743 clSetKernelArg (self, idx, sizeof (value), &value);
2744
2745void
2746set_local (OpenCL::Kernel self, cl_uint idx, size_t size)
2747 CODE:
2748 clSetKernelArg (self, idx, size, 0);
2749
2750void
2751set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value)
2752 CODE:
2753 clSetKernelArg (self, idx, sizeof (value), &value);
2754
2755void
438info (OpenCL::Kernel this, cl_kernel_info name) 2756info (OpenCL::Kernel self, cl_kernel_info name)
439 PPCODE: 2757 PPCODE:
440 INFO (Kernel) 2758 INFO (Kernel)
441 2759
2760#BEGIN:kernel
2761
442void 2762void
443set_bool (OpenCL::Kernel this, cl_uint idx, cl_bool value) 2763function_name (OpenCL::Kernel self)
2764 PPCODE:
2765 size_t size;
2766 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, 0, 0, &size));
2767 char *value = tmpbuf (size);
2768 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_FUNCTION_NAME, size, value, 0));
2769 EXTEND (SP, 1);
2770 const int i = 0;
2771 PUSHs (sv_2mortal (newSVpv (value, 0)));
2772
2773void
2774num_args (OpenCL::Kernel self)
2775 ALIAS:
2776 num_args = CL_KERNEL_NUM_ARGS
2777 reference_count = CL_KERNEL_REFERENCE_COUNT
2778 PPCODE:
2779 cl_uint value [1];
2780 NEED_SUCCESS (GetKernelInfo, (self, ix, sizeof (value), value, 0));
2781 EXTEND (SP, 1);
2782 const int i = 0;
2783 PUSHs (sv_2mortal (newSVuv (value [i])));
2784
2785void
2786context (OpenCL::Kernel self)
2787 PPCODE:
2788 cl_context value [1];
2789 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0));
2790 EXTEND (SP, 1);
2791 const int i = 0;
2792 NEED_SUCCESS (RetainContext, (value [i]));
2793 PUSH_CLOBJ (stash_context, value [i]);
2794
2795void
2796program (OpenCL::Kernel self)
2797 PPCODE:
2798 cl_program value [1];
2799 NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0));
2800 EXTEND (SP, 1);
2801 const int i = 0;
2802 NEED_SUCCESS (RetainProgram, (value [i]));
2803 PUSH_CLOBJ (stash_program, value [i]);
2804
2805#END:kernel
2806
2807void
2808work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name)
444 CODE: 2809 PPCODE:
445 clKernelSetArg (this, idx, sizeof (value), &value); 2810 size_t size;
2811 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, 0, 0, &size));
2812 SV *sv = sv_2mortal (newSV (size));
2813 SvUPGRADE (sv, SVt_PV);
2814 SvPOK_only (sv);
2815 SvCUR_set (sv, size);
2816 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, name, size, SvPVX (sv), 0));
2817 XPUSHs (sv);
2818
2819#BEGIN:kernel_work_group
2820
2821void
2822work_group_size (OpenCL::Kernel self, OpenCL::Device device)
2823 ALIAS:
2824 work_group_size = CL_KERNEL_WORK_GROUP_SIZE
2825 preferred_work_group_size_multiple = CL_KERNEL_PREFERRED_WORK_GROUP_SIZE_MULTIPLE
2826 PPCODE:
2827 size_t value [1];
2828 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2829 EXTEND (SP, 1);
2830 const int i = 0;
2831 PUSHs (sv_2mortal (newSVuv (value [i])));
2832
2833void
2834compile_work_group_size (OpenCL::Kernel self, OpenCL::Device device)
2835 PPCODE:
2836 size_t size;
2837 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, 0, 0, &size));
2838 size_t *value = tmpbuf (size);
2839 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, CL_KERNEL_COMPILE_WORK_GROUP_SIZE, size, value, 0));
2840 int i, n = size / sizeof (*value);
2841 EXTEND (SP, n);
2842 for (i = 0; i < n; ++i)
2843 PUSHs (sv_2mortal (newSVuv (value [i])));
2844
2845void
2846local_mem_size (OpenCL::Kernel self, OpenCL::Device device)
2847 ALIAS:
2848 local_mem_size = CL_KERNEL_LOCAL_MEM_SIZE
2849 private_mem_size = CL_KERNEL_PRIVATE_MEM_SIZE
2850 PPCODE:
2851 cl_ulong value [1];
2852 NEED_SUCCESS (GetKernelWorkGroupInfo, (self, device, ix, sizeof (value), value, 0));
2853 EXTEND (SP, 1);
2854 const int i = 0;
2855 PUSHs (sv_2mortal (newSVuv (value [i])));
2856
2857#END:kernel_work_group
2858
2859#if CL_VERSION_1_2
2860
2861void
2862arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name)
2863 PPCODE:
2864 size_t size;
2865 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size));
2866 SV *sv = sv_2mortal (newSV (size));
2867 SvUPGRADE (sv, SVt_PV);
2868 SvPOK_only (sv);
2869 SvCUR_set (sv, size);
2870 NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0));
2871 XPUSHs (sv);
2872
2873#BEGIN:kernel_arg
2874
2875void
2876arg_address_qualifier (OpenCL::Kernel self, cl_uint idx)
2877 ALIAS:
2878 arg_address_qualifier = CL_KERNEL_ARG_ADDRESS_QUALIFIER
2879 arg_access_qualifier = CL_KERNEL_ARG_ACCESS_QUALIFIER
2880 PPCODE:
2881 cl_uint value [1];
2882 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, sizeof (value), value, 0));
2883 EXTEND (SP, 1);
2884 const int i = 0;
2885 PUSHs (sv_2mortal (newSVuv (value [i])));
2886
2887void
2888arg_type_name (OpenCL::Kernel self, cl_uint idx)
2889 ALIAS:
2890 arg_type_name = CL_KERNEL_ARG_TYPE_NAME
2891 arg_name = CL_KERNEL_ARG_NAME
2892 PPCODE:
2893 size_t size;
2894 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size));
2895 char *value = tmpbuf (size);
2896 NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0));
2897 EXTEND (SP, 1);
2898 const int i = 0;
2899 PUSHs (sv_2mortal (newSVpv (value, 0)));
2900
2901void
2902arg_type_qualifier (OpenCL::Kernel self, cl_uint idx)
2903 PPCODE:
2904 cl_ulong value [1];
2905 NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0));
2906 EXTEND (SP, 1);
2907 const int i = 0;
2908 PUSHs (sv_2mortal (newSVuv (value [i])));
2909
2910#END:kernel_arg
2911
2912#endif
446 2913
447MODULE = OpenCL PACKAGE = OpenCL::Event 2914MODULE = OpenCL PACKAGE = OpenCL::Event
448 2915
449void 2916void
450DESTROY (OpenCL::Event this) 2917DESTROY (OpenCL::Event self)
451 CODE: 2918 CODE:
452 clReleaseEvent (this); 2919 clReleaseEvent (self);
453 2920
454void 2921void
2922wait (OpenCL::Event self)
2923 CODE:
2924 clWaitForEvents (1, &self);
2925
2926void
2927cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb)
2928 CODE:
2929 clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb)));
2930
2931void
455info (OpenCL::Event this, cl_event_info name) 2932info (OpenCL::Event self, cl_event_info name)
456 PPCODE: 2933 PPCODE:
457 INFO (Event) 2934 INFO (Event)
458 2935
2936#BEGIN:event
2937
459void 2938void
460wait (OpenCL::Event this) 2939command_queue (OpenCL::Event self)
2940 PPCODE:
2941 cl_command_queue value [1];
2942 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0));
2943 EXTEND (SP, 1);
2944 const int i = 0;
2945 NEED_SUCCESS (RetainCommandQueue, (value [i]));
2946 PUSH_CLOBJ (stash_queue, value [i]);
2947
2948void
2949command_type (OpenCL::Event self)
2950 ALIAS:
2951 command_type = CL_EVENT_COMMAND_TYPE
2952 reference_count = CL_EVENT_REFERENCE_COUNT
2953 command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS
2954 PPCODE:
2955 cl_uint value [1];
2956 NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0));
2957 EXTEND (SP, 1);
2958 const int i = 0;
2959 PUSHs (sv_2mortal (newSVuv (value [i])));
2960
2961void
2962context (OpenCL::Event self)
2963 PPCODE:
2964 cl_context value [1];
2965 NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0));
2966 EXTEND (SP, 1);
2967 const int i = 0;
2968 NEED_SUCCESS (RetainContext, (value [i]));
2969 PUSH_CLOBJ (stash_context, value [i]);
2970
2971#END:event
2972
2973void
2974profiling_info (OpenCL::Event self, cl_profiling_info name)
461 CODE: 2975 PPCODE:
462 clWaitForEvents (1, &this); 2976 INFO (EventProfiling)
463 2977
2978#BEGIN:profiling
2979
2980void
2981profiling_command_queued (OpenCL::Event self)
2982 ALIAS:
2983 profiling_command_queued = CL_PROFILING_COMMAND_QUEUED
2984 profiling_command_submit = CL_PROFILING_COMMAND_SUBMIT
2985 profiling_command_start = CL_PROFILING_COMMAND_START
2986 profiling_command_end = CL_PROFILING_COMMAND_END
2987 PPCODE:
2988 cl_ulong value [1];
2989 NEED_SUCCESS (GetEventProfilingInfo, (self, ix, sizeof (value), value, 0));
2990 EXTEND (SP, 1);
2991 const int i = 0;
2992 PUSHs (sv_2mortal (newSVuv (value [i])));
2993
2994#END:profiling
2995
2996MODULE = OpenCL PACKAGE = OpenCL::UserEvent
2997
2998void
2999set_status (OpenCL::UserEvent self, cl_int execution_status)
3000 CODE:
3001 clSetUserEventStatus (self, execution_status);
3002
3003MODULE = OpenCL PACKAGE = OpenCL::Mapped
3004
3005void
3006DESTROY (SV *self)
3007 CODE:
3008 OpenCL__Mapped mapped = SvMAPPED (self);
3009
3010 clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0);
3011 mapped_detach (self, mapped);
3012
3013 clReleaseCommandQueue (mapped->queue);
3014 clReleaseEvent (mapped->event);
3015 Safefree (mapped);
3016
3017void
3018unmap (OpenCL::Mapped self, ...)
3019 CODE:
3020 mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1);
3021
3022bool
3023mapped (OpenCL::Mapped self)
3024 CODE:
3025 RETVAL = !!self->ptr;
3026 OUTPUT:
3027 RETVAL
3028
3029void
3030wait (OpenCL::Mapped self)
3031 PPCODE:
3032 if (self->event)
3033 NEED_SUCCESS (WaitForEvents, (1, &self->event));
3034
3035void
3036event (OpenCL::Mapped self)
3037 PPCODE:
3038 if (!self->event)
3039 XSRETURN_UNDEF;
3040
3041 clRetainEvent (self->event);
3042 XPUSH_CLOBJ (stash_event, self->event);
3043
3044size_t
3045size (OpenCL::Mapped self)
3046 CODE:
3047 RETVAL = self->cb;
3048 OUTPUT:
3049 RETVAL
3050
3051IV
3052ptr (OpenCL::Mapped self)
3053 CODE:
3054 RETVAL = PTR2IV (self->ptr);
3055 OUTPUT:
3056 RETVAL
3057
3058void
3059set (OpenCL::Mapped self, size_t offset, SV *data)
3060 CODE:
3061 STRLEN len;
3062 const char *ptr = SvPVbyte (data, len);
3063
3064 if (offset + len > self->cb)
3065 croak ("OpenCL::Mapped::set out of bound condition detected");
3066
3067 memcpy (offset + (char *)self->ptr, ptr, len);
3068
3069MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer
3070
3071MODULE = OpenCL PACKAGE = OpenCL::MappedImage
3072
3073IV
3074row_pitch (OpenCL::Mapped self)
3075 ALIAS:
3076 slice_pitch = 1
3077 CODE:
3078 RETVAL = ix ? self->slice_pitch : self->row_pitch;
3079 OUTPUT:
3080 RETVAL
3081
3082
3083

Diff Legend

Removed lines
+ Added lines
< Changed lines
> Changed lines