1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
4 | |
4 | |
|
|
5 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
6 | #include "xthread.h" |
|
|
7 | #include "schmorp.h" |
|
|
8 | |
5 | #ifdef I_DLFCN |
9 | #ifdef I_DLFCN |
6 | #include <dlfcn.h> |
10 | #include <dlfcn.h> |
|
|
11 | #endif |
|
|
12 | |
|
|
13 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
14 | // but then fail to define the api functions unless you ALSO define |
|
|
15 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
16 | // after all, the functions are deprecated, not removed. |
|
|
17 | // in addition, you cannot test for this in any future-proof way. |
|
|
18 | // each time a new opencl version comes out, you need to make a new |
|
|
19 | // release. |
|
|
20 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
21 | |
|
|
22 | #ifndef PREFER_1_1 |
|
|
23 | #define PREFER_1_1 1 |
|
|
24 | #endif |
|
|
25 | |
|
|
26 | #if PREFER_1_1 |
|
|
27 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
7 | #endif |
28 | #endif |
8 | |
29 | |
9 | #ifdef __APPLE__ |
30 | #ifdef __APPLE__ |
10 | #include <OpenCL/opencl.h> |
31 | #include <OpenCL/opencl.h> |
11 | #else |
32 | #else |
12 | #include <CL/opencl.h> |
33 | #include <CL/opencl.h> |
|
|
34 | #endif |
|
|
35 | |
|
|
36 | #ifndef CL_VERSION_1_2 |
|
|
37 | #undef PREFER_1_1 |
|
|
38 | #define PREFER_1_1 1 |
13 | #endif |
39 | #endif |
14 | |
40 | |
15 | typedef cl_platform_id OpenCL__Platform; |
41 | typedef cl_platform_id OpenCL__Platform; |
16 | typedef cl_device_id OpenCL__Device; |
42 | typedef cl_device_id OpenCL__Device; |
17 | typedef cl_context OpenCL__Context; |
43 | typedef cl_context OpenCL__Context; |
… | |
… | |
31 | typedef cl_program OpenCL__Program; |
57 | typedef cl_program OpenCL__Program; |
32 | typedef cl_kernel OpenCL__Kernel; |
58 | typedef cl_kernel OpenCL__Kernel; |
33 | typedef cl_event OpenCL__Event; |
59 | typedef cl_event OpenCL__Event; |
34 | typedef cl_event OpenCL__UserEvent; |
60 | typedef cl_event OpenCL__UserEvent; |
35 | |
61 | |
|
|
62 | typedef struct mapped * OpenCL__Mapped; |
|
|
63 | |
36 | typedef SV *FUTURE; |
64 | typedef SV *FUTURE; |
|
|
65 | |
|
|
66 | static HV |
|
|
67 | *stash_platform, |
|
|
68 | *stash_device, |
|
|
69 | *stash_context, |
|
|
70 | *stash_queue, |
|
|
71 | *stash_program, |
|
|
72 | *stash_kernel, |
|
|
73 | *stash_sampler, |
|
|
74 | *stash_event, |
|
|
75 | *stash_userevent, |
|
|
76 | *stash_memory, |
|
|
77 | *stash_buffer, |
|
|
78 | *stash_bufferobj, |
|
|
79 | *stash_image, |
|
|
80 | *stash_image1d, |
|
|
81 | *stash_image1darray, |
|
|
82 | *stash_image1dbuffer, |
|
|
83 | *stash_image2d, |
|
|
84 | *stash_image2darray, |
|
|
85 | *stash_image3d, |
|
|
86 | *stash_mapped, |
|
|
87 | *stash_mappedbuffer, |
|
|
88 | *stash_mappedimage; |
37 | |
89 | |
38 | /*****************************************************************************/ |
90 | /*****************************************************************************/ |
39 | |
91 | |
40 | // name must include a leading underscore |
92 | // name must include a leading underscore |
41 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
93 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
… | |
… | |
89 | enum { buffers = 3 }; |
141 | enum { buffers = 3 }; |
90 | static int idx; |
142 | static int idx; |
91 | static void *buf [buffers]; |
143 | static void *buf [buffers]; |
92 | static size_t len [buffers]; |
144 | static size_t len [buffers]; |
93 | |
145 | |
94 | idx = ++idx % buffers; |
146 | idx = (idx + 1) % buffers; |
95 | |
147 | |
96 | if (len [idx] < size) |
148 | if (len [idx] < size) |
97 | { |
149 | { |
98 | free (buf [idx]); |
150 | free (buf [idx]); |
99 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
151 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
164 | |
216 | |
165 | #define NEED_SUCCESS_ARG(retdecl, name, args) \ |
217 | #define NEED_SUCCESS_ARG(retdecl, name, args) \ |
166 | retdecl = cl ## name args; \ |
218 | retdecl = cl ## name args; \ |
167 | if (res) \ |
219 | if (res) \ |
168 | FAIL (name); |
220 | FAIL (name); |
169 | |
|
|
170 | /*****************************************************************************/ |
|
|
171 | |
|
|
172 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
173 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
174 | |
|
|
175 | static void * |
|
|
176 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
177 | { |
|
|
178 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
179 | return (void *)SvIV (SvRV (sv)); |
|
|
180 | |
|
|
181 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
182 | } |
|
|
183 | |
221 | |
184 | /*****************************************************************************/ |
222 | /*****************************************************************************/ |
185 | |
223 | |
186 | static cl_context_properties * |
224 | static cl_context_properties * |
187 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
225 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
… | |
… | |
248 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
286 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
249 | } |
287 | } |
250 | |
288 | |
251 | /*****************************************************************************/ |
289 | /*****************************************************************************/ |
252 | |
290 | |
|
|
291 | static SV * |
|
|
292 | new_clobj (HV *stash, IV id) |
|
|
293 | { |
|
|
294 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
295 | } |
|
|
296 | |
|
|
297 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
298 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
299 | |
|
|
300 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
301 | /* they can be upgraded at runtime to the array form */ |
|
|
302 | static void * |
|
|
303 | SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
304 | { |
|
|
305 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
306 | return (void *)SvIV (SvRV (sv)); |
|
|
307 | |
|
|
308 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
309 | } |
|
|
310 | |
|
|
311 | /*****************************************************************************/ |
|
|
312 | /* callback stuff */ |
|
|
313 | |
|
|
314 | /* default context callback, log to stderr */ |
|
|
315 | static void CL_CALLBACK |
|
|
316 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
317 | { |
|
|
318 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
319 | } |
|
|
320 | |
|
|
321 | typedef struct |
|
|
322 | { |
|
|
323 | int free_cb; |
|
|
324 | void (*push)(void *data1, void *data2, void *data3); |
|
|
325 | } eq_vtbl; |
|
|
326 | |
|
|
327 | typedef struct eq_item |
|
|
328 | { |
|
|
329 | struct eq_item *next; |
|
|
330 | eq_vtbl *vtbl; |
|
|
331 | SV *cb; |
|
|
332 | void *data1, *data2, *data3; |
|
|
333 | } eq_item; |
|
|
334 | |
|
|
335 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
336 | static void *eq_signal_arg; |
|
|
337 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
338 | static eq_item *eq_head, *eq_tail; |
|
|
339 | |
|
|
340 | static void |
|
|
341 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
342 | { |
|
|
343 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
344 | |
|
|
345 | item->next = 0; |
|
|
346 | item->vtbl = vtbl; |
|
|
347 | item->cb = cb; |
|
|
348 | item->data1 = data1; |
|
|
349 | item->data2 = data2; |
|
|
350 | item->data3 = data3; |
|
|
351 | |
|
|
352 | X_LOCK (eq_lock); |
|
|
353 | |
|
|
354 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
355 | eq_tail = item; |
|
|
356 | |
|
|
357 | X_UNLOCK (eq_lock); |
|
|
358 | |
|
|
359 | eq_signal_func (eq_signal_arg, 0); |
|
|
360 | } |
|
|
361 | |
|
|
362 | static eq_item * |
|
|
363 | eq_dec (void) |
|
|
364 | { |
|
|
365 | eq_item *res; |
|
|
366 | |
|
|
367 | X_LOCK (eq_lock); |
|
|
368 | |
|
|
369 | res = eq_head; |
|
|
370 | if (res) |
|
|
371 | eq_head = res->next; |
|
|
372 | |
|
|
373 | X_UNLOCK (eq_lock); |
|
|
374 | |
|
|
375 | return res; |
|
|
376 | } |
|
|
377 | |
|
|
378 | static void |
|
|
379 | eq_poll (void) |
|
|
380 | { |
|
|
381 | eq_item *item; |
|
|
382 | |
|
|
383 | while ((item = eq_dec ())) |
|
|
384 | { |
|
|
385 | ENTER; |
|
|
386 | SAVETMPS; |
|
|
387 | |
|
|
388 | dSP; |
|
|
389 | PUSHMARK (SP); |
|
|
390 | EXTEND (SP, 2); |
|
|
391 | |
|
|
392 | if (item->vtbl->free_cb) |
|
|
393 | sv_2mortal (item->cb); |
|
|
394 | |
|
|
395 | PUTBACK; |
|
|
396 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
397 | |
|
|
398 | SV *cb = item->cb; |
|
|
399 | free (item); |
|
|
400 | |
|
|
401 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
402 | |
|
|
403 | FREETMPS; |
|
|
404 | LEAVE; |
|
|
405 | } |
|
|
406 | } |
|
|
407 | |
|
|
408 | static void |
|
|
409 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
410 | { |
|
|
411 | eq_poll (); |
|
|
412 | } |
|
|
413 | |
|
|
414 | /*****************************************************************************/ |
|
|
415 | /* context notify */ |
|
|
416 | |
|
|
417 | static void |
|
|
418 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
419 | { |
|
|
420 | dSP; |
|
|
421 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
422 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
423 | PUTBACK; |
|
|
424 | |
|
|
425 | free (data1); |
|
|
426 | free (data2); |
|
|
427 | } |
|
|
428 | |
|
|
429 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
430 | |
|
|
431 | static void CL_CALLBACK |
|
|
432 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
433 | { |
|
|
434 | void *pvt_copy = malloc (cb); |
|
|
435 | memcpy (pvt_copy, pvt, cb); |
|
|
436 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
437 | } |
|
|
438 | |
|
|
439 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
440 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
441 | void *user_data = 0; \ |
|
|
442 | \ |
|
|
443 | if (SvOK (notify)) \ |
|
|
444 | { \ |
|
|
445 | pfn_notify = eq_context_notify; \ |
|
|
446 | user_data = s_get_cv (notify); \ |
|
|
447 | } |
|
|
448 | |
|
|
449 | static SV * |
|
|
450 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
451 | { |
|
|
452 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
453 | |
|
|
454 | if (user_data) |
|
|
455 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
456 | |
|
|
457 | return sv; |
|
|
458 | } |
|
|
459 | |
|
|
460 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
461 | |
|
|
462 | /*****************************************************************************/ |
|
|
463 | /* build/compile/link notify */ |
|
|
464 | |
|
|
465 | static void |
|
|
466 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
467 | { |
|
|
468 | dSP; |
|
|
469 | PUSH_CLOBJ (stash_program, data1); |
|
|
470 | PUTBACK; |
|
|
471 | } |
|
|
472 | |
|
|
473 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
474 | |
|
|
475 | static void CL_CALLBACK |
|
|
476 | eq_program_notify (cl_program program, void *user_data) |
|
|
477 | { |
|
|
478 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
479 | } |
|
|
480 | |
|
|
481 | struct build_args |
|
|
482 | { |
|
|
483 | cl_program program; |
|
|
484 | char *options; |
|
|
485 | void *user_data; |
|
|
486 | cl_uint num_devices; |
|
|
487 | }; |
|
|
488 | |
|
|
489 | X_THREAD_PROC (build_program_thread) |
|
|
490 | { |
|
|
491 | struct build_args *arg = thr_arg; |
|
|
492 | |
|
|
493 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
494 | |
|
|
495 | if (arg->user_data) |
|
|
496 | eq_program_notify (arg->program, arg->user_data); |
|
|
497 | else |
|
|
498 | clReleaseProgram (arg->program); |
|
|
499 | |
|
|
500 | free (arg->options); |
|
|
501 | free (arg); |
|
|
502 | } |
|
|
503 | |
|
|
504 | static void |
|
|
505 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
506 | { |
|
|
507 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
508 | |
|
|
509 | arg->program = program; |
|
|
510 | arg->options = strdup (options); |
|
|
511 | arg->user_data = user_data; |
|
|
512 | arg->num_devices = num_devices; |
|
|
513 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
514 | |
|
|
515 | xthread_t id; |
|
|
516 | thread_create (&id, build_program_thread, arg); |
|
|
517 | } |
|
|
518 | |
|
|
519 | /*****************************************************************************/ |
|
|
520 | /* event objects */ |
|
|
521 | |
|
|
522 | static void |
|
|
523 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
524 | { |
|
|
525 | dSP; |
|
|
526 | PUSH_CLOBJ (stash_event, data1); |
|
|
527 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
528 | PUTBACK; |
|
|
529 | } |
|
|
530 | |
|
|
531 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
532 | |
|
|
533 | static void CL_CALLBACK |
|
|
534 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
535 | { |
|
|
536 | clRetainEvent (event); |
|
|
537 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
538 | } |
|
|
539 | |
|
|
540 | /*****************************************************************************/ |
|
|
541 | /* mapped_xxx */ |
|
|
542 | |
|
|
543 | static OpenCL__Mapped |
|
|
544 | SvMAPPED (SV *sv) |
|
|
545 | { |
|
|
546 | // no typechecking atm., keep your fingers crossed |
|
|
547 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
548 | } |
|
|
549 | |
|
|
550 | struct mapped |
|
|
551 | { |
|
|
552 | cl_command_queue queue; |
|
|
553 | cl_mem memobj; |
|
|
554 | void *ptr; |
|
|
555 | size_t cb; |
|
|
556 | cl_event event; |
|
|
557 | size_t row_pitch; |
|
|
558 | size_t slice_pitch; |
|
|
559 | }; |
|
|
560 | |
|
|
561 | static SV * |
|
|
562 | mapped_new (HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags, void *ptr, size_t cb, cl_event ev, size_t row_pitch, size_t slice_pitch) |
|
|
563 | { |
|
|
564 | SV *data = newSV (0); |
|
|
565 | SvUPGRADE (data, SVt_PVMG); |
|
|
566 | |
|
|
567 | OpenCL__Mapped mapped; |
|
|
568 | New (0, mapped, 1, struct mapped); |
|
|
569 | |
|
|
570 | clRetainCommandQueue (queue); |
|
|
571 | |
|
|
572 | mapped->queue = queue; |
|
|
573 | mapped->memobj = memobj; |
|
|
574 | mapped->ptr = ptr; |
|
|
575 | mapped->cb = cb; |
|
|
576 | mapped->event = ev; |
|
|
577 | mapped->row_pitch = row_pitch; |
|
|
578 | mapped->slice_pitch = slice_pitch; |
|
|
579 | |
|
|
580 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
581 | |
|
|
582 | if (!(flags & CL_MAP_WRITE)) |
|
|
583 | SvREADONLY_on (data); |
|
|
584 | |
|
|
585 | if (SvLEN (data)) |
|
|
586 | Safefree (data); |
|
|
587 | |
|
|
588 | SvPVX (data) = (char *)ptr; |
|
|
589 | SvCUR_set (data, cb); |
|
|
590 | SvLEN_set (data, 0); |
|
|
591 | SvPOK_only (data); |
|
|
592 | |
|
|
593 | return sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
594 | } |
|
|
595 | |
|
|
596 | static void |
|
|
597 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
598 | { |
|
|
599 | SV *data = SvRV (sv); |
|
|
600 | |
|
|
601 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
602 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
603 | else |
|
|
604 | { |
|
|
605 | SvREADONLY_off (data); |
|
|
606 | SvCUR_set (data, 0); |
|
|
607 | SvPVX (data) = 0; |
|
|
608 | SvOK_off (data); |
|
|
609 | } |
|
|
610 | |
|
|
611 | mapped->ptr = 0; |
|
|
612 | } |
|
|
613 | |
|
|
614 | /*****************************************************************************/ |
|
|
615 | |
253 | static size_t |
616 | static size_t |
254 | img_row_pitch (cl_mem img) |
617 | img_row_pitch (cl_mem img) |
255 | { |
618 | { |
256 | size_t res; |
619 | size_t res; |
257 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
620 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
258 | return res; |
621 | return res; |
259 | } |
622 | } |
260 | |
623 | |
261 | static cl_event * |
624 | static cl_event * |
262 | event_list (SV **items, int count) |
625 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
263 | { |
626 | { |
264 | if (!count) |
627 | cl_uint count = *rcount; |
|
|
628 | |
|
|
629 | if (!count && !extra) |
265 | return 0; |
630 | return 0; |
266 | |
631 | |
267 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
632 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
633 | int i = 0; |
268 | |
634 | |
269 | while (count--) |
635 | while (count) |
|
|
636 | { |
|
|
637 | --count; |
|
|
638 | if (SvOK (items [count])) |
270 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
639 | list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
|
|
640 | } |
271 | |
641 | |
|
|
642 | if (extra) |
|
|
643 | list [i++] = extra; |
|
|
644 | |
|
|
645 | *rcount = i; |
|
|
646 | |
272 | return list; |
647 | return i ? list : 0; |
273 | } |
648 | } |
274 | |
649 | |
275 | #define EVENT_LIST(items,count) \ |
650 | #define EVENT_LIST(skip) \ |
276 | cl_uint event_list_count = (count); \ |
651 | cl_uint event_list_count = items - skip; \ |
277 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
652 | cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) |
278 | |
653 | |
279 | #define INFO(class) \ |
654 | #define INFO(class) \ |
280 | { \ |
655 | { \ |
281 | size_t size; \ |
656 | size_t size; \ |
282 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
657 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
289 | } |
664 | } |
290 | |
665 | |
291 | MODULE = OpenCL PACKAGE = OpenCL |
666 | MODULE = OpenCL PACKAGE = OpenCL |
292 | |
667 | |
293 | PROTOTYPES: ENABLE |
668 | PROTOTYPES: ENABLE |
|
|
669 | |
|
|
670 | void |
|
|
671 | poll () |
|
|
672 | CODE: |
|
|
673 | eq_poll (); |
|
|
674 | |
|
|
675 | void |
|
|
676 | _eq_initialise (IV func, IV arg) |
|
|
677 | CODE: |
|
|
678 | eq_signal_func = (void (*)(void *, int))func; |
|
|
679 | eq_signal_arg = (void*)arg; |
294 | |
680 | |
295 | BOOT: |
681 | BOOT: |
296 | { |
682 | { |
297 | HV *stash = gv_stashpv ("OpenCL", 1); |
683 | HV *stash = gv_stashpv ("OpenCL", 1); |
298 | static const ivstr *civ, const_iv[] = { |
684 | static const ivstr *civ, const_iv[] = { |
… | |
… | |
307 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
693 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
308 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
694 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
309 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
695 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
310 | #include "constiv.h" |
696 | #include "constiv.h" |
311 | }; |
697 | }; |
|
|
698 | |
312 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
699 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
313 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
700 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
701 | |
|
|
702 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
703 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
704 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
705 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
706 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
707 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
708 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
709 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
710 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
711 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
712 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
713 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
714 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
715 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
716 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
717 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
718 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
719 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
720 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
721 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
722 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
723 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
724 | |
|
|
725 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
314 | } |
726 | } |
315 | |
727 | |
316 | cl_int |
728 | cl_int |
317 | errno () |
729 | errno () |
318 | CODE: |
730 | CODE: |
319 | errno = res; |
731 | RETVAL = res; |
|
|
732 | OUTPUT: |
|
|
733 | RETVAL |
320 | |
734 | |
321 | const char * |
735 | const char * |
322 | err2str (cl_int err) |
736 | err2str (cl_int err = res) |
323 | |
737 | |
324 | const char * |
738 | const char * |
325 | enum2str (cl_uint value) |
739 | enum2str (cl_uint value) |
326 | |
740 | |
327 | void |
741 | void |
… | |
… | |
335 | list = tmpbuf (sizeof (*list) * count); |
749 | list = tmpbuf (sizeof (*list) * count); |
336 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
750 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
337 | |
751 | |
338 | EXTEND (SP, count); |
752 | EXTEND (SP, count); |
339 | for (i = 0; i < count; ++i) |
753 | for (i = 0; i < count; ++i) |
340 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
754 | PUSH_CLOBJ (stash_platform, (IV)list [i]); |
341 | |
755 | |
342 | void |
756 | void |
343 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
757 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
344 | PPCODE: |
758 | PPCODE: |
|
|
759 | CONTEXT_NOTIFY_CALLBACK; |
345 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
760 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
346 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
761 | XPUSH_CLOBJ_CONTEXT; |
347 | |
762 | |
348 | void |
763 | void |
349 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
764 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
350 | PPCODE: |
765 | PPCODE: |
351 | /* der Gipfel der Kunst */ |
766 | /* der Gipfel der Kunst */ |
352 | |
767 | |
353 | void |
768 | void |
354 | wait_for_events (...) |
769 | wait_for_events (...) |
355 | CODE: |
770 | CODE: |
356 | EVENT_LIST (0, items); |
771 | EVENT_LIST (0); |
357 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
772 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
358 | |
773 | |
359 | PROTOTYPES: DISABLE |
774 | PROTOTYPES: DISABLE |
360 | |
775 | |
361 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
776 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
362 | |
777 | |
363 | void |
778 | void |
364 | info (OpenCL::Platform self, cl_platform_info name) |
779 | info (OpenCL::Platform self, cl_platform_info name) |
365 | PPCODE: |
780 | PPCODE: |
366 | INFO (Platform) |
781 | INFO (Platform) |
|
|
782 | |
|
|
783 | void |
|
|
784 | unload_compiler (OpenCL::Platform self) |
|
|
785 | CODE: |
|
|
786 | #if CL_VERSION_1_2 |
|
|
787 | clUnloadPlatformCompiler (self); |
|
|
788 | #endif |
367 | |
789 | |
368 | #BEGIN:platform |
790 | #BEGIN:platform |
369 | |
791 | |
370 | void |
792 | void |
371 | profile (OpenCL::Platform self) |
793 | profile (OpenCL::Platform self) |
… | |
… | |
397 | list = tmpbuf (sizeof (*list) * count); |
819 | list = tmpbuf (sizeof (*list) * count); |
398 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
820 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
399 | |
821 | |
400 | EXTEND (SP, count); |
822 | EXTEND (SP, count); |
401 | for (i = 0; i < count; ++i) |
823 | for (i = 0; i < count; ++i) |
402 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
824 | PUSH_CLOBJ (stash_device, list [i]); |
403 | |
825 | |
404 | void |
826 | void |
405 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
827 | context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
406 | PPCODE: |
828 | PPCODE: |
407 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
829 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
408 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
830 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
409 | |
831 | |
410 | AV *av = (AV *)SvRV (devices); |
832 | AV *av = (AV *)SvRV (devices); |
411 | cl_uint num_devices = av_len (av) + 1; |
833 | cl_uint num_devices = av_len (av) + 1; |
412 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
834 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
|
|
835 | |
413 | int i; |
836 | int i; |
414 | |
|
|
415 | for (i = num_devices; i--; ) |
837 | for (i = num_devices; i--; ) |
416 | device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
838 | device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
417 | |
839 | |
|
|
840 | CONTEXT_NOTIFY_CALLBACK; |
418 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res)); |
841 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res)); |
419 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
842 | XPUSH_CLOBJ_CONTEXT; |
420 | |
843 | |
421 | void |
844 | void |
422 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
845 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
423 | PPCODE: |
846 | PPCODE: |
424 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
847 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
425 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
848 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
849 | |
|
|
850 | CONTEXT_NOTIFY_CALLBACK; |
426 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
851 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
427 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
852 | XPUSH_CLOBJ_CONTEXT; |
428 | |
853 | |
429 | MODULE = OpenCL PACKAGE = OpenCL::Device |
854 | MODULE = OpenCL PACKAGE = OpenCL::Device |
430 | |
855 | |
431 | void |
856 | void |
432 | info (OpenCL::Device self, cl_device_info name) |
857 | info (OpenCL::Device self, cl_device_info name) |
… | |
… | |
440 | PPCODE: |
865 | PPCODE: |
441 | cl_device_type value [1]; |
866 | cl_device_type value [1]; |
442 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
867 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
443 | EXTEND (SP, 1); |
868 | EXTEND (SP, 1); |
444 | const int i = 0; |
869 | const int i = 0; |
445 | PUSHs (sv_2mortal (newSViv (value [i]))); |
870 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
446 | |
871 | |
447 | void |
872 | void |
448 | vendor_id (OpenCL::Device self) |
873 | vendor_id (OpenCL::Device self) |
449 | ALIAS: |
874 | ALIAS: |
450 | vendor_id = CL_DEVICE_VENDOR_ID |
875 | vendor_id = CL_DEVICE_VENDOR_ID |
… | |
… | |
471 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
896 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
472 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
897 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
473 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
898 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
474 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
899 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
475 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
900 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
476 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
901 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
477 | PPCODE: |
902 | PPCODE: |
478 | cl_uint value [1]; |
903 | cl_uint value [1]; |
479 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
904 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
480 | EXTEND (SP, 1); |
905 | EXTEND (SP, 1); |
481 | const int i = 0; |
906 | const int i = 0; |
… | |
… | |
595 | PPCODE: |
1020 | PPCODE: |
596 | cl_command_queue_properties value [1]; |
1021 | cl_command_queue_properties value [1]; |
597 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1022 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
598 | EXTEND (SP, 1); |
1023 | EXTEND (SP, 1); |
599 | const int i = 0; |
1024 | const int i = 0; |
600 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1025 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
601 | |
1026 | |
602 | void |
1027 | void |
603 | platform (OpenCL::Device self) |
1028 | platform (OpenCL::Device self) |
604 | PPCODE: |
1029 | PPCODE: |
605 | cl_platform_id value [1]; |
1030 | cl_platform_id value [1]; |
606 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1031 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
607 | EXTEND (SP, 1); |
1032 | EXTEND (SP, 1); |
608 | const int i = 0; |
1033 | const int i = 0; |
609 | { |
1034 | { |
610 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); |
1035 | PUSH_CLOBJ (stash_platform, value [i]); |
611 | } |
1036 | } |
612 | |
1037 | |
613 | void |
1038 | void |
614 | name (OpenCL::Device self) |
1039 | name (OpenCL::Device self) |
615 | ALIAS: |
1040 | ALIAS: |
… | |
… | |
634 | cl_device_id value [1]; |
1059 | cl_device_id value [1]; |
635 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1060 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
636 | EXTEND (SP, 1); |
1061 | EXTEND (SP, 1); |
637 | const int i = 0; |
1062 | const int i = 0; |
638 | { |
1063 | { |
639 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1064 | PUSH_CLOBJ (stash_device, value [i]); |
640 | } |
1065 | } |
641 | |
1066 | |
642 | void |
1067 | void |
643 | partition_types_ext (OpenCL::Device self) |
1068 | partition_types_ext (OpenCL::Device self) |
644 | ALIAS: |
1069 | ALIAS: |
… | |
… | |
671 | |
1096 | |
672 | void |
1097 | void |
673 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
1098 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
674 | PPCODE: |
1099 | PPCODE: |
675 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
1100 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
676 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
1101 | XPUSH_CLOBJ (stash_queue, queue); |
677 | |
1102 | |
678 | void |
1103 | void |
679 | user_event (OpenCL::Context self) |
1104 | user_event (OpenCL::Context self) |
680 | PPCODE: |
1105 | PPCODE: |
681 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
1106 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
682 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
1107 | XPUSH_CLOBJ (stash_userevent, ev); |
683 | |
1108 | |
684 | void |
1109 | void |
685 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
1110 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
686 | PPCODE: |
1111 | PPCODE: |
687 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1112 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
688 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1113 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
689 | |
1114 | |
690 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
1115 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
691 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1116 | XPUSH_CLOBJ (stash_bufferobj, mem); |
692 | |
1117 | |
693 | void |
1118 | void |
694 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
1119 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
695 | PPCODE: |
1120 | PPCODE: |
696 | STRLEN len; |
1121 | STRLEN len; |
697 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1122 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
698 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1123 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
699 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1124 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
700 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
1125 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
701 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1126 | XPUSH_CLOBJ (stash_bufferobj, mem); |
702 | |
1127 | |
|
|
1128 | #if CL_VERSION_1_2 |
|
|
1129 | |
703 | void |
1130 | void |
704 | image2d (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) |
1131 | image (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) |
705 | PPCODE: |
1132 | PPCODE: |
706 | STRLEN len; |
1133 | STRLEN len; |
707 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1134 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
708 | const cl_image_format format = { channel_order, channel_type }; |
1135 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1136 | const cl_image_desc desc = { |
|
|
1137 | type, |
|
|
1138 | width, height, depth, |
|
|
1139 | array_size, row_pitch, slice_pitch, |
|
|
1140 | num_mip_level, num_samples, |
|
|
1141 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
|
|
1142 | }; |
709 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
1143 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
710 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1144 | HV *stash = stash_image; |
|
|
1145 | switch (type) |
|
|
1146 | { |
|
|
1147 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1148 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1149 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1150 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1151 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1152 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1153 | } |
|
|
1154 | XPUSH_CLOBJ (stash, mem); |
711 | |
1155 | |
|
|
1156 | #endif |
|
|
1157 | |
712 | void |
1158 | void |
713 | image3d (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) |
1159 | image2d (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) |
714 | PPCODE: |
1160 | PPCODE: |
715 | STRLEN len; |
1161 | STRLEN len; |
716 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1162 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
717 | const cl_image_format format = { channel_order, channel_type }; |
1163 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1164 | #if PREFER_1_1 |
|
|
1165 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1166 | #else |
|
|
1167 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1168 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1169 | #endif |
|
|
1170 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1171 | |
|
|
1172 | void |
|
|
1173 | image3d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t depth, size_t row_pitch = 0, size_t slice_pitch = 0, SV *data = &PL_sv_undef) |
|
|
1174 | PPCODE: |
|
|
1175 | STRLEN len; |
|
|
1176 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1177 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1178 | #if PREFER_1_1 |
718 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
1179 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
719 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1180 | #else |
|
|
1181 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1182 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1183 | #endif |
|
|
1184 | XPUSH_CLOBJ (stash_image3d, mem); |
720 | |
1185 | |
721 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1186 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
722 | |
1187 | |
723 | void |
1188 | void |
724 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
1189 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
725 | PPCODE: |
1190 | PPCODE: |
726 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
1191 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
727 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1192 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1193 | |
|
|
1194 | void |
|
|
1195 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1196 | PPCODE: |
|
|
1197 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1198 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1199 | |
|
|
1200 | #if CL_VERSION_1_2 |
|
|
1201 | |
|
|
1202 | void |
|
|
1203 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1204 | ALIAS: |
|
|
1205 | PPCODE: |
|
|
1206 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1207 | cl_gl_object_type type; |
|
|
1208 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1209 | HV *stash = stash_memory; |
|
|
1210 | switch (type) |
|
|
1211 | { |
|
|
1212 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1213 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1214 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1215 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1216 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1217 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1218 | } |
|
|
1219 | XPUSH_CLOBJ (stash, mem); |
|
|
1220 | |
|
|
1221 | #endif |
728 | |
1222 | |
729 | void |
1223 | void |
730 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1224 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
731 | PPCODE: |
1225 | PPCODE: |
|
|
1226 | #if PREFER_1_1 |
732 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
1227 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
733 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1228 | #else |
|
|
1229 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1230 | #endif |
|
|
1231 | XPUSH_CLOBJ (stash_image2d, mem); |
734 | |
1232 | |
735 | void |
1233 | void |
736 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1234 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
737 | PPCODE: |
1235 | PPCODE: |
|
|
1236 | #if PREFER_1_1 |
738 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
1237 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
739 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1238 | #else |
740 | |
1239 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
741 | void |
1240 | #endif |
742 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
1241 | XPUSH_CLOBJ (stash_image3d, mem); |
743 | PPCODE: |
|
|
744 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
745 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
746 | |
1242 | |
747 | #endif |
1243 | #endif |
748 | |
1244 | |
749 | void |
1245 | void |
750 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
1246 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
770 | |
1266 | |
771 | void |
1267 | void |
772 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1268 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
773 | PPCODE: |
1269 | PPCODE: |
774 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
1270 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
775 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1271 | XPUSH_CLOBJ (stash_sampler, sampler); |
776 | |
1272 | |
777 | void |
1273 | void |
778 | program_with_source (OpenCL::Context self, SV *program) |
1274 | program_with_source (OpenCL::Context self, SV *program) |
779 | PPCODE: |
1275 | PPCODE: |
780 | STRLEN len; |
1276 | STRLEN len; |
781 | size_t len2; |
1277 | size_t len2; |
782 | const char *ptr = SvPVbyte (program, len); |
1278 | const char *ptr = SvPVbyte (program, len); |
783 | |
1279 | |
784 | len2 = len; |
1280 | len2 = len; |
785 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1281 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
786 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1282 | XPUSH_CLOBJ (stash_program, prog); |
787 | |
1283 | |
788 | #BEGIN:context |
1284 | #BEGIN:context |
789 | |
1285 | |
790 | void |
1286 | void |
791 | reference_count (OpenCL::Context self) |
1287 | reference_count (OpenCL::Context self) |
… | |
… | |
808 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1304 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
809 | int i, n = size / sizeof (*value); |
1305 | int i, n = size / sizeof (*value); |
810 | EXTEND (SP, n); |
1306 | EXTEND (SP, n); |
811 | for (i = 0; i < n; ++i) |
1307 | for (i = 0; i < n; ++i) |
812 | { |
1308 | { |
813 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1309 | PUSH_CLOBJ (stash_device, value [i]); |
814 | } |
1310 | } |
815 | |
1311 | |
816 | void |
1312 | void |
817 | properties (OpenCL::Context self) |
1313 | properties (OpenCL::Context self) |
818 | PPCODE: |
1314 | PPCODE: |
… | |
… | |
833 | DESTROY (OpenCL::Queue self) |
1329 | DESTROY (OpenCL::Queue self) |
834 | CODE: |
1330 | CODE: |
835 | clReleaseCommandQueue (self); |
1331 | clReleaseCommandQueue (self); |
836 | |
1332 | |
837 | void |
1333 | void |
838 | enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1334 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1335 | ALIAS: |
|
|
1336 | enqueue_read_buffer = 0 |
839 | PPCODE: |
1337 | PPCODE: |
840 | cl_event ev = 0; |
1338 | cl_event ev = 0; |
841 | EVENT_LIST (6, items - 6); |
1339 | EVENT_LIST (6); |
842 | |
1340 | |
843 | SvUPGRADE (data, SVt_PV); |
1341 | SvUPGRADE (data, SVt_PV); |
844 | SvGROW (data, len); |
1342 | SvGROW (data, len); |
845 | SvPOK_only (data); |
1343 | SvPOK_only (data); |
846 | SvCUR_set (data, len); |
1344 | SvCUR_set (data, len); |
847 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1345 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
848 | |
1346 | |
849 | if (ev) |
1347 | if (ev) |
850 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1348 | XPUSH_CLOBJ (stash_event, ev); |
851 | |
1349 | |
852 | void |
1350 | void |
853 | enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1351 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1352 | ALIAS: |
|
|
1353 | enqueue_write_buffer = 0 |
854 | PPCODE: |
1354 | PPCODE: |
855 | cl_event ev = 0; |
1355 | cl_event ev = 0; |
856 | STRLEN len; |
1356 | STRLEN len; |
857 | char *ptr = SvPVbyte (data, len); |
1357 | char *ptr = SvPVbyte (data, len); |
858 | EVENT_LIST (5, items - 5); |
1358 | EVENT_LIST (5); |
859 | |
1359 | |
860 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1360 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
861 | |
1361 | |
862 | if (ev) |
1362 | if (ev) |
863 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1363 | XPUSH_CLOBJ (stash_event, ev); |
864 | |
1364 | |
|
|
1365 | #if CL_VERSION_1_2 |
|
|
1366 | |
865 | void |
1367 | void |
866 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1368 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1369 | ALIAS: |
|
|
1370 | enqueue_fill_buffer = 0 |
867 | PPCODE: |
1371 | PPCODE: |
868 | cl_event ev = 0; |
1372 | cl_event ev = 0; |
|
|
1373 | STRLEN len; |
|
|
1374 | char *ptr = SvPVbyte (data, len); |
|
|
1375 | EVENT_LIST (5); |
|
|
1376 | |
|
|
1377 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1378 | |
|
|
1379 | if (ev) |
|
|
1380 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1381 | |
|
|
1382 | void |
|
|
1383 | fill_image (OpenCL::Queue self, OpenCL::Image img, NV r, NV g, NV b, NV a, size_t x, size_t y, size_t z, size_t width, size_t height, size_t depth, ...) |
|
|
1384 | ALIAS: |
|
|
1385 | enqueue_fill_image = 0 |
|
|
1386 | PPCODE: |
|
|
1387 | cl_event ev = 0; |
|
|
1388 | STRLEN len; |
|
|
1389 | const size_t origin [3] = { x, y, z }; |
|
|
1390 | const size_t region [3] = { width, height, depth }; |
|
|
1391 | EVENT_LIST (12); |
|
|
1392 | |
|
|
1393 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1394 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1395 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1396 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1397 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1398 | cl_image_format format; |
|
|
1399 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1400 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1401 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1402 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1403 | |
|
|
1404 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1405 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1406 | |
|
|
1407 | if (ev) |
|
|
1408 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1409 | |
|
|
1410 | #endif |
|
|
1411 | |
|
|
1412 | void |
|
|
1413 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1414 | ALIAS: |
|
|
1415 | enqueue_copy_buffer = 0 |
|
|
1416 | PPCODE: |
|
|
1417 | cl_event ev = 0; |
869 | EVENT_LIST (6, items - 6); |
1418 | EVENT_LIST (6); |
870 | |
1419 | |
871 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1420 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
872 | |
1421 | |
873 | if (ev) |
1422 | if (ev) |
874 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1423 | XPUSH_CLOBJ (stash_event, ev); |
875 | |
1424 | |
876 | void |
1425 | void |
877 | enqueue_read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1426 | read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
|
|
1427 | ALIAS: |
|
|
1428 | enqueue_read_buffer_rect = 0 |
878 | PPCODE: |
1429 | PPCODE: |
879 | cl_event ev = 0; |
1430 | cl_event ev = 0; |
880 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1431 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
881 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1432 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
882 | const size_t region[3] = { width, height, depth }; |
1433 | const size_t region[3] = { width, height, depth }; |
883 | EVENT_LIST (17, items - 17); |
1434 | EVENT_LIST (17); |
884 | |
1435 | |
885 | if (!buf_row_pitch) |
1436 | if (!buf_row_pitch) |
886 | buf_row_pitch = region [0]; |
1437 | buf_row_pitch = region [0]; |
887 | |
1438 | |
888 | if (!buf_slice_pitch) |
1439 | if (!buf_slice_pitch) |
… | |
… | |
901 | SvPOK_only (data); |
1452 | SvPOK_only (data); |
902 | SvCUR_set (data, len); |
1453 | SvCUR_set (data, len); |
903 | 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)); |
1454 | NEED_SUCCESS (EnqueueReadBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
904 | |
1455 | |
905 | if (ev) |
1456 | if (ev) |
906 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1457 | XPUSH_CLOBJ (stash_event, ev); |
907 | |
1458 | |
908 | void |
1459 | void |
909 | enqueue_write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1460 | write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
|
|
1461 | ALIAS: |
|
|
1462 | enqueue_write_buffer_rect = 0 |
910 | PPCODE: |
1463 | PPCODE: |
911 | cl_event ev = 0; |
1464 | cl_event ev = 0; |
912 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1465 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
913 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1466 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
914 | const size_t region[3] = { width, height, depth }; |
1467 | const size_t region[3] = { width, height, depth }; |
915 | STRLEN len; |
1468 | STRLEN len; |
916 | char *ptr = SvPVbyte (data, len); |
1469 | char *ptr = SvPVbyte (data, len); |
917 | EVENT_LIST (17, items - 17); |
1470 | EVENT_LIST (17); |
918 | |
1471 | |
919 | if (!buf_row_pitch) |
1472 | if (!buf_row_pitch) |
920 | buf_row_pitch = region [0]; |
1473 | buf_row_pitch = region [0]; |
921 | |
1474 | |
922 | if (!buf_slice_pitch) |
1475 | if (!buf_slice_pitch) |
… | |
… | |
931 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1484 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
932 | |
1485 | |
933 | if (len < min_len) |
1486 | if (len < min_len) |
934 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1487 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
935 | |
1488 | |
936 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1489 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
937 | |
1490 | |
938 | if (ev) |
1491 | if (ev) |
939 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1492 | XPUSH_CLOBJ (stash_event, ev); |
940 | |
1493 | |
941 | void |
1494 | void |
942 | enqueue_copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
1495 | copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
|
|
1496 | ALIAS: |
|
|
1497 | enqueue_copy_buffer_rect = 0 |
943 | PPCODE: |
1498 | PPCODE: |
944 | cl_event ev = 0; |
1499 | cl_event ev = 0; |
945 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1500 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
946 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1501 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
947 | const size_t region[3] = { width, height, depth }; |
1502 | const size_t region[3] = { width, height, depth }; |
948 | EVENT_LIST (16, items - 16); |
1503 | EVENT_LIST (16); |
949 | |
1504 | |
950 | 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)); |
1505 | NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
951 | |
1506 | |
952 | if (ev) |
1507 | if (ev) |
953 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1508 | XPUSH_CLOBJ (stash_event, ev); |
954 | |
1509 | |
955 | void |
1510 | void |
956 | enqueue_read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1511 | read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
|
|
1512 | ALIAS: |
|
|
1513 | enqueue_read_image = 0 |
957 | PPCODE: |
1514 | PPCODE: |
958 | cl_event ev = 0; |
1515 | cl_event ev = 0; |
959 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1516 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
960 | const size_t region[3] = { width, height, depth }; |
1517 | const size_t region[3] = { width, height, depth }; |
961 | EVENT_LIST (12, items - 12); |
1518 | EVENT_LIST (12); |
962 | |
1519 | |
963 | if (!row_pitch) |
1520 | if (!row_pitch) |
964 | row_pitch = img_row_pitch (src); |
1521 | row_pitch = img_row_pitch (src); |
965 | |
1522 | |
966 | if (depth > 1 && !slice_pitch) |
1523 | if (depth > 1 && !slice_pitch) |
… | |
… | |
973 | SvPOK_only (data); |
1530 | SvPOK_only (data); |
974 | SvCUR_set (data, len); |
1531 | SvCUR_set (data, len); |
975 | 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)); |
1532 | NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
976 | |
1533 | |
977 | if (ev) |
1534 | if (ev) |
978 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1535 | XPUSH_CLOBJ (stash_event, ev); |
979 | |
1536 | |
980 | void |
1537 | void |
981 | enqueue_write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1538 | write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
|
|
1539 | ALIAS: |
|
|
1540 | enqueue_write_image = 0 |
982 | PPCODE: |
1541 | PPCODE: |
983 | cl_event ev = 0; |
1542 | cl_event ev = 0; |
984 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1543 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
985 | const size_t region[3] = { width, height, depth }; |
1544 | const size_t region[3] = { width, height, depth }; |
986 | STRLEN len; |
1545 | STRLEN len; |
987 | char *ptr = SvPVbyte (data, len); |
1546 | char *ptr = SvPVbyte (data, len); |
988 | EVENT_LIST (12, items - 12); |
1547 | EVENT_LIST (12); |
989 | |
1548 | |
990 | if (!row_pitch) |
1549 | if (!row_pitch) |
991 | row_pitch = img_row_pitch (dst); |
1550 | row_pitch = img_row_pitch (dst); |
992 | |
1551 | |
993 | if (depth > 1 && !slice_pitch) |
1552 | if (depth > 1 && !slice_pitch) |
… | |
… | |
996 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1555 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
997 | |
1556 | |
998 | if (len < min_len) |
1557 | if (len < min_len) |
999 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1558 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1000 | |
1559 | |
1001 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1560 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1002 | |
1561 | |
1003 | if (ev) |
1562 | if (ev) |
1004 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1563 | XPUSH_CLOBJ (stash_event, ev); |
1005 | |
1564 | |
1006 | void |
1565 | void |
1007 | enqueue_copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
1566 | copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
1567 | ALIAS: |
|
|
1568 | enqueue_copy_image = 0 |
1008 | PPCODE: |
1569 | PPCODE: |
1009 | cl_event ev = 0; |
1570 | cl_event ev = 0; |
1010 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1571 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1011 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1572 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1012 | const size_t region[3] = { width, height, depth }; |
1573 | const size_t region[3] = { width, height, depth }; |
1013 | EVENT_LIST (12, items - 12); |
1574 | EVENT_LIST (12); |
1014 | |
1575 | |
1015 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1576 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1016 | |
1577 | |
1017 | if (ev) |
1578 | if (ev) |
1018 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1579 | XPUSH_CLOBJ (stash_event, ev); |
1019 | |
1580 | |
1020 | void |
1581 | void |
1021 | enqueue_copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
1582 | copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
|
|
1583 | ALIAS: |
|
|
1584 | enqueue_copy_image_to_buffer = 0 |
1022 | PPCODE: |
1585 | PPCODE: |
1023 | cl_event ev = 0; |
1586 | cl_event ev = 0; |
1024 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1587 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1588 | const size_t region [3] = { width, height, depth }; |
|
|
1589 | EVENT_LIST (10); |
|
|
1590 | |
|
|
1591 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1592 | |
|
|
1593 | if (ev) |
|
|
1594 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1595 | |
|
|
1596 | void |
|
|
1597 | copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
1598 | ALIAS: |
|
|
1599 | enqueue_copy_buffer_to_image = 0 |
|
|
1600 | PPCODE: |
|
|
1601 | cl_event ev = 0; |
|
|
1602 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1603 | const size_t region [3] = { width, height, depth }; |
|
|
1604 | EVENT_LIST (10); |
|
|
1605 | |
|
|
1606 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1607 | |
|
|
1608 | if (ev) |
|
|
1609 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1610 | |
|
|
1611 | void |
|
|
1612 | map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, size_t cb = 0, ...) |
|
|
1613 | ALIAS: |
|
|
1614 | enqueue_map_buffer = 0 |
|
|
1615 | PPCODE: |
|
|
1616 | cl_event ev; |
|
|
1617 | EVENT_LIST (7); |
|
|
1618 | |
|
|
1619 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1620 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1621 | |
|
|
1622 | void |
|
|
1623 | map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, size_t width = 0, size_t height = 0, size_t depth = 0, ...) |
|
|
1624 | ALIAS: |
|
|
1625 | enqueue_map_image = 0 |
|
|
1626 | PPCODE: |
|
|
1627 | cl_event ev; |
|
|
1628 | const size_t origin[3] = { x, y, z }; |
1025 | const size_t region[3] = { width, height, depth }; |
1629 | const size_t region[3] = { width, height, depth }; |
|
|
1630 | size_t row_pitch, slice_pitch; |
1026 | EVENT_LIST (10, items - 10); |
1631 | EVENT_LIST (11); |
1027 | |
1632 | |
|
|
1633 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapImage, (self, img, blocking, map_flags, origin, region, &row_pitch, &slice_pitch, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1634 | |
|
|
1635 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1636 | : row_pitch ? row_pitch * region [1] |
|
|
1637 | : region [0]; |
|
|
1638 | |
|
|
1639 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1640 | |
|
|
1641 | void |
|
|
1642 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1643 | PPCODE: |
|
|
1644 | cl_uint event_list_count = items - 2; |
|
|
1645 | cl_event *event_list_ptr = event_list (&ST (2), &event_list_count, mapped->event); |
|
|
1646 | |
|
|
1647 | NEED_SUCCESS (EnqueueUnmapMemObject, (self, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &mapped->event)); |
|
|
1648 | mapped_detach (ST (1), mapped); |
|
|
1649 | |
|
|
1650 | if (GIMME_V != G_VOID) |
|
|
1651 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1652 | |
|
|
1653 | void |
|
|
1654 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1655 | ALIAS: |
|
|
1656 | enqueue_task = 0 |
|
|
1657 | PPCODE: |
|
|
1658 | cl_event ev = 0; |
|
|
1659 | EVENT_LIST (2); |
|
|
1660 | |
1028 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1661 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1029 | |
1662 | |
1030 | if (ev) |
1663 | if (ev) |
1031 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1664 | XPUSH_CLOBJ (stash_event, ev); |
1032 | |
1665 | |
1033 | void |
1666 | void |
1034 | enqueue_copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
1035 | PPCODE: |
|
|
1036 | cl_event ev = 0; |
|
|
1037 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1038 | const size_t region[3] = { width, height, depth }; |
|
|
1039 | EVENT_LIST (10, items - 10); |
|
|
1040 | |
|
|
1041 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1042 | |
|
|
1043 | if (ev) |
|
|
1044 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1045 | |
|
|
1046 | void |
|
|
1047 | enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1048 | PPCODE: |
|
|
1049 | cl_event ev = 0; |
|
|
1050 | EVENT_LIST (2, items - 2); |
|
|
1051 | |
|
|
1052 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1053 | |
|
|
1054 | if (ev) |
|
|
1055 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1056 | |
|
|
1057 | void |
|
|
1058 | enqueue_nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1667 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1668 | ALIAS: |
|
|
1669 | enqueue_nd_range_kernel = 0 |
1059 | PPCODE: |
1670 | PPCODE: |
1060 | cl_event ev = 0; |
1671 | cl_event ev = 0; |
1061 | size_t *gwo = 0, *gws, *lws = 0; |
1672 | size_t *gwo = 0, *gws, *lws = 0; |
1062 | int gws_len; |
1673 | int gws_len; |
1063 | size_t *lists; |
1674 | size_t *lists; |
1064 | int i; |
1675 | int i; |
1065 | EVENT_LIST (5, items - 5); |
1676 | EVENT_LIST (5); |
1066 | |
1677 | |
1067 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1678 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1068 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1679 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1069 | |
1680 | |
1070 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1681 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1071 | |
1682 | |
1072 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1683 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1073 | |
1684 | |
1074 | gws = lists + gws_len * 0; |
1685 | gws = lists + gws_len * 0; |
1075 | for (i = 0; i < gws_len; ++i) |
1686 | for (i = 0; i < gws_len; ++i) |
|
|
1687 | { |
1076 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
1688 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1689 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1690 | if (!gws [i]) |
|
|
1691 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1692 | } |
1077 | |
1693 | |
1078 | if (SvOK (global_work_offset)) |
1694 | if (SvOK (global_work_offset)) |
1079 | { |
1695 | { |
1080 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1696 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1081 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
1697 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
1088 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1704 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1089 | } |
1705 | } |
1090 | |
1706 | |
1091 | if (SvOK (local_work_size)) |
1707 | if (SvOK (local_work_size)) |
1092 | { |
1708 | { |
1093 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1709 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1094 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1710 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
1095 | |
1711 | |
1096 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1712 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1097 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1713 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1098 | |
1714 | |
1099 | lws = lists + gws_len * 2; |
1715 | lws = lists + gws_len * 2; |
1100 | for (i = 0; i < gws_len; ++i) |
1716 | for (i = 0; i < gws_len; ++i) |
|
|
1717 | { |
1101 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
1718 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1719 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1720 | if (!lws [i]) |
|
|
1721 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
1722 | } |
1102 | } |
1723 | } |
1103 | |
1724 | |
1104 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1725 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1105 | |
1726 | |
1106 | if (ev) |
1727 | if (ev) |
1107 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1728 | XPUSH_CLOBJ (stash_event, ev); |
1108 | |
1729 | |
1109 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1730 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1110 | |
1731 | |
1111 | void |
1732 | void |
1112 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1733 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1113 | ALIAS: |
1734 | ALIAS: |
|
|
1735 | release_gl_objects = 1 |
|
|
1736 | enqueue_acquire_gl_objects = 0 |
1114 | enqueue_release_gl_objects = 1 |
1737 | enqueue_release_gl_objects = 1 |
1115 | CODE: |
1738 | PPCODE: |
1116 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
1739 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
1117 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
1740 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
1118 | |
1741 | |
1119 | cl_event ev = 0; |
1742 | cl_event ev = 0; |
1120 | EVENT_LIST (2, items - 2); |
1743 | EVENT_LIST (2); |
1121 | AV *av = (AV *)SvRV (objects); |
1744 | AV *av = (AV *)SvRV (objects); |
1122 | cl_uint num_objects = av_len (av) + 1; |
1745 | cl_uint num_objects = av_len (av) + 1; |
1123 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
1746 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
1124 | int i; |
1747 | int i; |
1125 | |
1748 | |
1126 | for (i = num_objects; i--; ) |
1749 | for (i = num_objects; i--; ) |
1127 | object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
1750 | object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
1128 | |
1751 | |
1129 | if (ix) |
1752 | if (ix) |
1130 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1753 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1131 | else |
1754 | else |
1132 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1755 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1133 | |
1756 | |
1134 | if (ev) |
1757 | if (ev) |
1135 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1758 | XPUSH_CLOBJ (stash_event, ev); |
1136 | |
1759 | |
1137 | #endif |
1760 | #endif |
1138 | |
1761 | |
1139 | void |
1762 | void |
1140 | enqueue_marker (OpenCL::Queue self) |
|
|
1141 | PPCODE: |
|
|
1142 | cl_event ev; |
|
|
1143 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1144 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1145 | |
|
|
1146 | void |
|
|
1147 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1763 | wait_for_events (OpenCL::Queue self, ...) |
|
|
1764 | ALIAS: |
|
|
1765 | enqueue_wait_for_events = 0 |
1148 | CODE: |
1766 | CODE: |
1149 | EVENT_LIST (1, items - 1); |
1767 | EVENT_LIST (1); |
|
|
1768 | #if PREFER_1_1 |
1150 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1769 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1770 | #else |
|
|
1771 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
1772 | #endif |
1151 | |
1773 | |
1152 | void |
1774 | void |
1153 | enqueue_barrier (OpenCL::Queue self) |
1775 | marker (OpenCL::Queue self, ...) |
|
|
1776 | ALIAS: |
|
|
1777 | enqueue_marker = 0 |
1154 | CODE: |
1778 | PPCODE: |
|
|
1779 | cl_event ev = 0; |
|
|
1780 | EVENT_LIST (1); |
|
|
1781 | #if PREFER_1_1 |
|
|
1782 | if (!event_list_count) |
|
|
1783 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1784 | else |
|
|
1785 | #if CL_VERSION_1_2 |
|
|
1786 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1787 | #else |
|
|
1788 | { |
|
|
1789 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
1790 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1791 | } |
|
|
1792 | #endif |
|
|
1793 | #else |
|
|
1794 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1795 | #endif |
|
|
1796 | if (ev) |
|
|
1797 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1798 | |
|
|
1799 | void |
|
|
1800 | barrier (OpenCL::Queue self, ...) |
|
|
1801 | ALIAS: |
|
|
1802 | enqueue_barrier = 0 |
|
|
1803 | PPCODE: |
|
|
1804 | cl_event ev = 0; |
|
|
1805 | EVENT_LIST (1); |
|
|
1806 | #if PREFER_1_1 |
|
|
1807 | if (!event_list_count && GIMME_V == G_VOID) |
1155 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1808 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1809 | else |
|
|
1810 | #if CL_VERSION_1_2 |
|
|
1811 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1812 | #else |
|
|
1813 | { |
|
|
1814 | if (event_list_count) |
|
|
1815 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1816 | |
|
|
1817 | if (GIMME_V != G_VOID) |
|
|
1818 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1819 | } |
|
|
1820 | #endif |
|
|
1821 | #else |
|
|
1822 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1823 | #endif |
|
|
1824 | if (ev) |
|
|
1825 | XPUSH_CLOBJ (stash_event, ev); |
1156 | |
1826 | |
1157 | void |
1827 | void |
1158 | flush (OpenCL::Queue self) |
1828 | flush (OpenCL::Queue self) |
1159 | CODE: |
1829 | CODE: |
1160 | NEED_SUCCESS (Flush, (self)); |
1830 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1178 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1848 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1179 | EXTEND (SP, 1); |
1849 | EXTEND (SP, 1); |
1180 | const int i = 0; |
1850 | const int i = 0; |
1181 | { |
1851 | { |
1182 | NEED_SUCCESS (RetainContext, (value [i])); |
1852 | NEED_SUCCESS (RetainContext, (value [i])); |
1183 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
1853 | PUSH_CLOBJ (stash_context, value [i]); |
1184 | } |
1854 | } |
1185 | |
1855 | |
1186 | void |
1856 | void |
1187 | device (OpenCL::Queue self) |
1857 | device (OpenCL::Queue self) |
1188 | PPCODE: |
1858 | PPCODE: |
1189 | cl_device_id value [1]; |
1859 | cl_device_id value [1]; |
1190 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1860 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1191 | EXTEND (SP, 1); |
1861 | EXTEND (SP, 1); |
1192 | const int i = 0; |
1862 | const int i = 0; |
1193 | { |
1863 | { |
1194 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1864 | PUSH_CLOBJ (stash_device, value [i]); |
1195 | } |
1865 | } |
1196 | |
1866 | |
1197 | void |
1867 | void |
1198 | reference_count (OpenCL::Queue self) |
1868 | reference_count (OpenCL::Queue self) |
1199 | PPCODE: |
1869 | PPCODE: |
… | |
… | |
1208 | PPCODE: |
1878 | PPCODE: |
1209 | cl_command_queue_properties value [1]; |
1879 | cl_command_queue_properties value [1]; |
1210 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1880 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1211 | EXTEND (SP, 1); |
1881 | EXTEND (SP, 1); |
1212 | const int i = 0; |
1882 | const int i = 0; |
1213 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1883 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1214 | |
1884 | |
1215 | #END:command_queue |
1885 | #END:command_queue |
1216 | |
1886 | |
1217 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1887 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1218 | |
1888 | |
… | |
… | |
1233 | PPCODE: |
1903 | PPCODE: |
1234 | cl_mem_object_type value [1]; |
1904 | cl_mem_object_type value [1]; |
1235 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1905 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1236 | EXTEND (SP, 1); |
1906 | EXTEND (SP, 1); |
1237 | const int i = 0; |
1907 | const int i = 0; |
1238 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1908 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1239 | |
1909 | |
1240 | void |
1910 | void |
1241 | flags (OpenCL::Memory self) |
1911 | flags (OpenCL::Memory self) |
1242 | PPCODE: |
1912 | PPCODE: |
1243 | cl_mem_flags value [1]; |
1913 | cl_mem_flags value [1]; |
1244 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1914 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1245 | EXTEND (SP, 1); |
1915 | EXTEND (SP, 1); |
1246 | const int i = 0; |
1916 | const int i = 0; |
1247 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1917 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1248 | |
1918 | |
1249 | void |
1919 | void |
1250 | size (OpenCL::Memory self) |
1920 | size (OpenCL::Memory self) |
1251 | ALIAS: |
1921 | ALIAS: |
1252 | size = CL_MEM_SIZE |
1922 | size = CL_MEM_SIZE |
… | |
… | |
1286 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1956 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1287 | EXTEND (SP, 1); |
1957 | EXTEND (SP, 1); |
1288 | const int i = 0; |
1958 | const int i = 0; |
1289 | { |
1959 | { |
1290 | NEED_SUCCESS (RetainContext, (value [i])); |
1960 | NEED_SUCCESS (RetainContext, (value [i])); |
1291 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
1961 | PUSH_CLOBJ (stash_context, value [i]); |
1292 | } |
1962 | } |
1293 | |
1963 | |
1294 | void |
1964 | void |
1295 | associated_memobject (OpenCL::Memory self) |
1965 | associated_memobject (OpenCL::Memory self) |
1296 | PPCODE: |
1966 | PPCODE: |
… | |
… | |
1298 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1968 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1299 | EXTEND (SP, 1); |
1969 | EXTEND (SP, 1); |
1300 | const int i = 0; |
1970 | const int i = 0; |
1301 | { |
1971 | { |
1302 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1972 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1303 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
1973 | PUSH_CLOBJ (stash_memory, value [i]); |
1304 | } |
1974 | } |
1305 | |
1975 | |
1306 | #END:mem |
1976 | #END:mem |
1307 | |
1977 | |
1308 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1978 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
… | |
… | |
1328 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1998 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1329 | |
1999 | |
1330 | cl_buffer_region crdata = { origin, size }; |
2000 | cl_buffer_region crdata = { origin, size }; |
1331 | |
2001 | |
1332 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
2002 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
1333 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
2003 | XPUSH_CLOBJ (stash_buffer, mem); |
1334 | |
2004 | |
1335 | MODULE = OpenCL PACKAGE = OpenCL::Image |
2005 | MODULE = OpenCL PACKAGE = OpenCL::Image |
1336 | |
2006 | |
1337 | void |
2007 | void |
1338 | image_info (OpenCL::Image self, cl_image_info name) |
2008 | image_info (OpenCL::Image self, cl_image_info name) |
1339 | PPCODE: |
2009 | PPCODE: |
1340 | INFO (Image) |
2010 | INFO (Image) |
|
|
2011 | |
|
|
2012 | void |
|
|
2013 | format (OpenCL::Image self) |
|
|
2014 | PPCODE: |
|
|
2015 | cl_image_format format; |
|
|
2016 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2017 | EXTEND (SP, 2); |
|
|
2018 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2019 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
1341 | |
2020 | |
1342 | #BEGIN:image |
2021 | #BEGIN:image |
1343 | |
2022 | |
1344 | void |
2023 | void |
1345 | element_size (OpenCL::Image self) |
2024 | element_size (OpenCL::Image self) |
… | |
… | |
1415 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2094 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
1416 | EXTEND (SP, 1); |
2095 | EXTEND (SP, 1); |
1417 | const int i = 0; |
2096 | const int i = 0; |
1418 | { |
2097 | { |
1419 | NEED_SUCCESS (RetainContext, (value [i])); |
2098 | NEED_SUCCESS (RetainContext, (value [i])); |
1420 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2099 | PUSH_CLOBJ (stash_context, value [i]); |
1421 | } |
2100 | } |
1422 | |
2101 | |
1423 | void |
2102 | void |
1424 | normalized_coords (OpenCL::Sampler self) |
2103 | normalized_coords (OpenCL::Sampler self) |
1425 | PPCODE: |
2104 | PPCODE: |
1426 | cl_addressing_mode value [1]; |
2105 | cl_addressing_mode value [1]; |
1427 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
2106 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
1428 | EXTEND (SP, 1); |
2107 | EXTEND (SP, 1); |
1429 | const int i = 0; |
2108 | const int i = 0; |
1430 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2109 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1431 | |
2110 | |
1432 | void |
2111 | void |
1433 | addressing_mode (OpenCL::Sampler self) |
2112 | addressing_mode (OpenCL::Sampler self) |
1434 | PPCODE: |
2113 | PPCODE: |
1435 | cl_filter_mode value [1]; |
2114 | cl_filter_mode value [1]; |
1436 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
2115 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
1437 | EXTEND (SP, 1); |
2116 | EXTEND (SP, 1); |
1438 | const int i = 0; |
2117 | const int i = 0; |
1439 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2118 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1440 | |
2119 | |
1441 | void |
2120 | void |
1442 | filter_mode (OpenCL::Sampler self) |
2121 | filter_mode (OpenCL::Sampler self) |
1443 | PPCODE: |
2122 | PPCODE: |
1444 | cl_bool value [1]; |
2123 | cl_bool value [1]; |
… | |
… | |
1455 | DESTROY (OpenCL::Program self) |
2134 | DESTROY (OpenCL::Program self) |
1456 | CODE: |
2135 | CODE: |
1457 | clReleaseProgram (self); |
2136 | clReleaseProgram (self); |
1458 | |
2137 | |
1459 | void |
2138 | void |
1460 | build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) |
2139 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2140 | ALIAS: |
|
|
2141 | build_async = 1 |
1461 | CODE: |
2142 | CODE: |
|
|
2143 | void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0; |
|
|
2144 | void *user_data = 0; |
|
|
2145 | cl_uint num_devices = 0; |
|
|
2146 | cl_device_id *device_list = 0; |
|
|
2147 | |
|
|
2148 | if (SvOK (devices)) |
|
|
2149 | { |
|
|
2150 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
2151 | croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); |
|
|
2152 | |
|
|
2153 | AV *av = (AV *)SvRV (devices); |
|
|
2154 | num_devices = av_len (av) + 1; |
|
|
2155 | |
|
|
2156 | if (num_devices) |
|
|
2157 | { |
|
|
2158 | device_list = tmpbuf (sizeof (*device_list) * num_devices); |
|
|
2159 | int count; |
|
|
2160 | for (count = 0; count < num_devices; ++count) |
|
|
2161 | device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device"); |
|
|
2162 | } |
|
|
2163 | } |
|
|
2164 | |
|
|
2165 | if (SvOK (notify)) |
|
|
2166 | { |
|
|
2167 | NEED_SUCCESS (RetainProgram, (self)); |
|
|
2168 | pfn_notify = eq_program_notify; |
|
|
2169 | user_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
2170 | } |
|
|
2171 | |
|
|
2172 | if (ix) |
|
|
2173 | build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2174 | else |
1462 | NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2175 | NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
1463 | |
2176 | |
1464 | void |
2177 | void |
1465 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2178 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
1466 | PPCODE: |
2179 | PPCODE: |
1467 | size_t size; |
2180 | size_t size; |
… | |
… | |
1502 | |
2215 | |
1503 | void |
2216 | void |
1504 | kernel (OpenCL::Program program, SV *function) |
2217 | kernel (OpenCL::Program program, SV *function) |
1505 | PPCODE: |
2218 | PPCODE: |
1506 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
2219 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1507 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2220 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2221 | |
|
|
2222 | void |
|
|
2223 | kernels_in_program (OpenCL::Program program) |
|
|
2224 | PPCODE: |
|
|
2225 | cl_uint num_kernels; |
|
|
2226 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2227 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2228 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2229 | |
|
|
2230 | int i; |
|
|
2231 | EXTEND (SP, num_kernels); |
|
|
2232 | for (i = 0; i < num_kernels; ++i) |
|
|
2233 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
1508 | |
2234 | |
1509 | void |
2235 | void |
1510 | info (OpenCL::Program self, cl_program_info name) |
2236 | info (OpenCL::Program self, cl_program_info name) |
1511 | PPCODE: |
2237 | PPCODE: |
1512 | INFO (Program) |
2238 | INFO (Program) |
… | |
… | |
1530 | { |
2256 | { |
1531 | SV *sv = sv_2mortal (newSV (sizes [i])); |
2257 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1532 | SvUPGRADE (sv, SVt_PV); |
2258 | SvUPGRADE (sv, SVt_PV); |
1533 | SvPOK_only (sv); |
2259 | SvPOK_only (sv); |
1534 | SvCUR_set (sv, sizes [i]); |
2260 | SvCUR_set (sv, sizes [i]); |
1535 | ptrs [i] = SvPVX (sv); |
2261 | ptrs [i] = (void *)SvPVX (sv); |
1536 | PUSHs (sv); |
2262 | PUSHs (sv); |
1537 | } |
2263 | } |
1538 | |
2264 | |
1539 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
2265 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1540 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
2266 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
… | |
… | |
1560 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2286 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
1561 | EXTEND (SP, 1); |
2287 | EXTEND (SP, 1); |
1562 | const int i = 0; |
2288 | const int i = 0; |
1563 | { |
2289 | { |
1564 | NEED_SUCCESS (RetainContext, (value [i])); |
2290 | NEED_SUCCESS (RetainContext, (value [i])); |
1565 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2291 | PUSH_CLOBJ (stash_context, value [i]); |
1566 | } |
2292 | } |
1567 | |
2293 | |
1568 | void |
2294 | void |
1569 | devices (OpenCL::Program self) |
2295 | devices (OpenCL::Program self) |
1570 | PPCODE: |
2296 | PPCODE: |
… | |
… | |
1574 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2300 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
1575 | int i, n = size / sizeof (*value); |
2301 | int i, n = size / sizeof (*value); |
1576 | EXTEND (SP, n); |
2302 | EXTEND (SP, n); |
1577 | for (i = 0; i < n; ++i) |
2303 | for (i = 0; i < n; ++i) |
1578 | { |
2304 | { |
1579 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2305 | PUSH_CLOBJ (stash_device, value [i]); |
1580 | } |
2306 | } |
1581 | |
2307 | |
1582 | void |
2308 | void |
1583 | source (OpenCL::Program self) |
2309 | source (OpenCL::Program self) |
1584 | PPCODE: |
2310 | PPCODE: |
… | |
… | |
1610 | DESTROY (OpenCL::Kernel self) |
2336 | DESTROY (OpenCL::Kernel self) |
1611 | CODE: |
2337 | CODE: |
1612 | clReleaseKernel (self); |
2338 | clReleaseKernel (self); |
1613 | |
2339 | |
1614 | void |
2340 | void |
|
|
2341 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2342 | CODE: |
|
|
2343 | int i; |
|
|
2344 | for (i = 2; ; ++i) |
|
|
2345 | { |
|
|
2346 | while (*format == ' ') |
|
|
2347 | ++format; |
|
|
2348 | |
|
|
2349 | char type = *format++; |
|
|
2350 | |
|
|
2351 | if (!type) |
|
|
2352 | break; |
|
|
2353 | |
|
|
2354 | if (i >= items) |
|
|
2355 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2356 | |
|
|
2357 | SV *sv = ST (i); |
|
|
2358 | |
|
|
2359 | union |
|
|
2360 | { |
|
|
2361 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2362 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2363 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2364 | cl_mem cm; |
|
|
2365 | cl_sampler ca; |
|
|
2366 | size_t cz; |
|
|
2367 | cl_event ce; |
|
|
2368 | } arg; |
|
|
2369 | size_t size; |
|
|
2370 | int nullarg = 0; |
|
|
2371 | |
|
|
2372 | switch (type) |
|
|
2373 | { |
|
|
2374 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2375 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2376 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2377 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2378 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2379 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2380 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2381 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2382 | |
|
|
2383 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2384 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2385 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2386 | |
|
|
2387 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2388 | |
|
|
2389 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2390 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2391 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2392 | |
|
|
2393 | default: |
|
|
2394 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2395 | } |
|
|
2396 | |
|
|
2397 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2398 | if (res) |
|
|
2399 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2400 | } |
|
|
2401 | |
|
|
2402 | if (i != items) |
|
|
2403 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2404 | |
|
|
2405 | void |
1615 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
2406 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
1616 | CODE: |
2407 | CODE: |
1617 | clSetKernelArg (self, idx, sizeof (value), &value); |
2408 | clSetKernelArg (self, idx, sizeof (value), &value); |
1618 | |
2409 | |
1619 | void |
2410 | void |
… | |
… | |
1667 | clSetKernelArg (self, idx, sizeof (value), &value); |
2458 | clSetKernelArg (self, idx, sizeof (value), &value); |
1668 | |
2459 | |
1669 | void |
2460 | void |
1670 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
2461 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
1671 | CODE: |
2462 | CODE: |
1672 | clSetKernelArg (self, idx, sizeof (value), &value); |
2463 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
1673 | |
2464 | |
1674 | void |
2465 | void |
1675 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
2466 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
1676 | CODE: |
2467 | CODE: |
1677 | clSetKernelArg (self, idx, sizeof (value), &value); |
2468 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
1678 | |
2469 | |
1679 | void |
2470 | void |
1680 | set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value) |
2471 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2472 | ALIAS: |
|
|
2473 | set_image2d = 0 |
|
|
2474 | set_image3d = 0 |
1681 | CODE: |
2475 | CODE: |
1682 | clSetKernelArg (self, idx, sizeof (value), &value); |
2476 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
1683 | |
|
|
1684 | void |
|
|
1685 | set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
1686 | CODE: |
|
|
1687 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1688 | |
2477 | |
1689 | void |
2478 | void |
1690 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
2479 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
1691 | CODE: |
2480 | CODE: |
1692 | clSetKernelArg (self, idx, sizeof (value), &value); |
2481 | clSetKernelArg (self, idx, sizeof (value), &value); |
… | |
… | |
1738 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2527 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
1739 | EXTEND (SP, 1); |
2528 | EXTEND (SP, 1); |
1740 | const int i = 0; |
2529 | const int i = 0; |
1741 | { |
2530 | { |
1742 | NEED_SUCCESS (RetainContext, (value [i])); |
2531 | NEED_SUCCESS (RetainContext, (value [i])); |
1743 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2532 | PUSH_CLOBJ (stash_context, value [i]); |
1744 | } |
2533 | } |
1745 | |
2534 | |
1746 | void |
2535 | void |
1747 | program (OpenCL::Kernel self) |
2536 | program (OpenCL::Kernel self) |
1748 | PPCODE: |
2537 | PPCODE: |
… | |
… | |
1750 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2539 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
1751 | EXTEND (SP, 1); |
2540 | EXTEND (SP, 1); |
1752 | const int i = 0; |
2541 | const int i = 0; |
1753 | { |
2542 | { |
1754 | NEED_SUCCESS (RetainProgram, (value [i])); |
2543 | NEED_SUCCESS (RetainProgram, (value [i])); |
1755 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); |
2544 | PUSH_CLOBJ (stash_program, value [i]); |
1756 | } |
2545 | } |
1757 | |
2546 | |
1758 | #END:kernel |
2547 | #END:kernel |
1759 | |
2548 | |
1760 | void |
2549 | void |
… | |
… | |
1820 | wait (OpenCL::Event self) |
2609 | wait (OpenCL::Event self) |
1821 | CODE: |
2610 | CODE: |
1822 | clWaitForEvents (1, &self); |
2611 | clWaitForEvents (1, &self); |
1823 | |
2612 | |
1824 | void |
2613 | void |
|
|
2614 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
2615 | CODE: |
|
|
2616 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
|
|
2617 | |
|
|
2618 | void |
1825 | info (OpenCL::Event self, cl_event_info name) |
2619 | info (OpenCL::Event self, cl_event_info name) |
1826 | PPCODE: |
2620 | PPCODE: |
1827 | INFO (Event) |
2621 | INFO (Event) |
1828 | |
2622 | |
1829 | #BEGIN:event |
2623 | #BEGIN:event |
… | |
… | |
1835 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
2629 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
1836 | EXTEND (SP, 1); |
2630 | EXTEND (SP, 1); |
1837 | const int i = 0; |
2631 | const int i = 0; |
1838 | { |
2632 | { |
1839 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
2633 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
1840 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); |
2634 | PUSH_CLOBJ (stash_queue, value [i]); |
1841 | } |
2635 | } |
1842 | |
2636 | |
1843 | void |
2637 | void |
1844 | command_type (OpenCL::Event self) |
2638 | command_type (OpenCL::Event self) |
1845 | PPCODE: |
2639 | PPCODE: |
… | |
… | |
1868 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
2662 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
1869 | EXTEND (SP, 1); |
2663 | EXTEND (SP, 1); |
1870 | const int i = 0; |
2664 | const int i = 0; |
1871 | { |
2665 | { |
1872 | NEED_SUCCESS (RetainContext, (value [i])); |
2666 | NEED_SUCCESS (RetainContext, (value [i])); |
1873 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2667 | PUSH_CLOBJ (stash_context, value [i]); |
1874 | } |
2668 | } |
1875 | |
2669 | |
1876 | #END:event |
2670 | #END:event |
1877 | |
2671 | |
1878 | void |
2672 | void |
… | |
… | |
1903 | void |
2697 | void |
1904 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
2698 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
1905 | CODE: |
2699 | CODE: |
1906 | clSetUserEventStatus (self, execution_status); |
2700 | clSetUserEventStatus (self, execution_status); |
1907 | |
2701 | |
|
|
2702 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
2703 | |
|
|
2704 | void |
|
|
2705 | DESTROY (SV *self) |
|
|
2706 | CODE: |
|
|
2707 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
2708 | |
|
|
2709 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
2710 | mapped_detach (self, mapped); |
|
|
2711 | |
|
|
2712 | clReleaseCommandQueue (mapped->queue); |
|
|
2713 | |
|
|
2714 | if (mapped->event) |
|
|
2715 | clReleaseEvent (mapped->event); |
|
|
2716 | |
|
|
2717 | Safefree (mapped); |
|
|
2718 | |
|
|
2719 | bool |
|
|
2720 | mapped (OpenCL::Mapped self) |
|
|
2721 | CODE: |
|
|
2722 | RETVAL = !!self->ptr; |
|
|
2723 | OUTPUT: |
|
|
2724 | RETVAL |
|
|
2725 | |
|
|
2726 | void |
|
|
2727 | wait (OpenCL::Mapped self) |
|
|
2728 | PPCODE: |
|
|
2729 | if (self->event) |
|
|
2730 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
2731 | |
|
|
2732 | void |
|
|
2733 | event (OpenCL::Mapped self) |
|
|
2734 | PPCODE: |
|
|
2735 | if (!self->event) |
|
|
2736 | XSRETURN_UNDEF; |
|
|
2737 | |
|
|
2738 | clRetainEvent (self->event); |
|
|
2739 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
2740 | |
|
|
2741 | size_t |
|
|
2742 | size (OpenCL::Mapped self) |
|
|
2743 | CODE: |
|
|
2744 | RETVAL = self->cb; |
|
|
2745 | OUTPUT: |
|
|
2746 | RETVAL |
|
|
2747 | |
|
|
2748 | IV |
|
|
2749 | ptr (OpenCL::Mapped self) |
|
|
2750 | CODE: |
|
|
2751 | RETVAL = PTR2IV (self->ptr); |
|
|
2752 | OUTPUT: |
|
|
2753 | RETVAL |
|
|
2754 | |
|
|
2755 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
2756 | |
|
|
2757 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
2758 | |
|
|
2759 | IV |
|
|
2760 | row_pitch (OpenCL::Mapped self) |
|
|
2761 | ALIAS: |
|
|
2762 | slice_pitch = 1 |
|
|
2763 | CODE: |
|
|
2764 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
2765 | OUTPUT: |
|
|
2766 | RETVAL |
|
|
2767 | |
|
|
2768 | |
|
|
2769 | |