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 | |
|
|
5 | #include "ecb.h"//D |
4 | |
6 | |
5 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
7 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
6 | #include "xthread.h" |
8 | #include "xthread.h" |
7 | #include "schmorp.h" |
9 | #include "schmorp.h" |
8 | |
10 | |
… | |
… | |
38 | #define PREFER_1_1 1 |
40 | #define PREFER_1_1 1 |
39 | #endif |
41 | #endif |
40 | |
42 | |
41 | typedef cl_platform_id OpenCL__Platform; |
43 | typedef cl_platform_id OpenCL__Platform; |
42 | typedef cl_device_id OpenCL__Device; |
44 | typedef cl_device_id OpenCL__Device; |
|
|
45 | typedef cl_device_id OpenCL__SubDevice; |
43 | typedef cl_context OpenCL__Context; |
46 | typedef cl_context OpenCL__Context; |
44 | typedef cl_command_queue OpenCL__Queue; |
47 | typedef cl_command_queue OpenCL__Queue; |
45 | typedef cl_mem OpenCL__Memory; |
48 | typedef cl_mem OpenCL__Memory; |
46 | typedef cl_mem OpenCL__Buffer; |
49 | typedef cl_mem OpenCL__Buffer; |
47 | typedef cl_mem OpenCL__BufferObj; |
50 | typedef cl_mem OpenCL__BufferObj; |
48 | typedef cl_mem OpenCL__Image; |
51 | typedef cl_mem OpenCL__Image; |
49 | typedef cl_mem OpenCL__Image2D; |
|
|
50 | typedef cl_mem OpenCL__Image3D; |
|
|
51 | typedef cl_mem OpenCL__Memory_ornull; |
52 | typedef cl_mem OpenCL__Memory_ornull; |
52 | typedef cl_mem OpenCL__Buffer_ornull; |
53 | typedef cl_mem OpenCL__Buffer_ornull; |
53 | typedef cl_mem OpenCL__Image_ornull; |
54 | typedef cl_mem OpenCL__Image_ornull; |
54 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
55 | typedef cl_mem OpenCL__Image3D_ornull; |
|
|
56 | typedef cl_sampler OpenCL__Sampler; |
55 | typedef cl_sampler OpenCL__Sampler; |
57 | typedef cl_program OpenCL__Program; |
56 | typedef cl_program OpenCL__Program; |
58 | typedef cl_kernel OpenCL__Kernel; |
57 | typedef cl_kernel OpenCL__Kernel; |
59 | typedef cl_event OpenCL__Event; |
58 | typedef cl_event OpenCL__Event; |
60 | typedef cl_event OpenCL__UserEvent; |
59 | typedef cl_event OpenCL__UserEvent; |
… | |
… | |
64 | typedef SV *FUTURE; |
63 | typedef SV *FUTURE; |
65 | |
64 | |
66 | static HV |
65 | static HV |
67 | *stash_platform, |
66 | *stash_platform, |
68 | *stash_device, |
67 | *stash_device, |
|
|
68 | *stash_subdevice, |
69 | *stash_context, |
69 | *stash_context, |
70 | *stash_queue, |
70 | *stash_queue, |
71 | *stash_program, |
71 | *stash_program, |
72 | *stash_kernel, |
72 | *stash_kernel, |
73 | *stash_sampler, |
73 | *stash_sampler, |
… | |
… | |
133 | } |
133 | } |
134 | |
134 | |
135 | /*****************************************************************************/ |
135 | /*****************************************************************************/ |
136 | |
136 | |
137 | /* up to two temporary buffers */ |
137 | /* up to two temporary buffers */ |
138 | static void * |
138 | static void * ecb_noinline |
139 | tmpbuf (size_t size) |
139 | tmpbuf (size_t size) |
140 | { |
140 | { |
141 | enum { buffers = 3 }; |
141 | enum { buffers = 4 }; |
142 | static int idx; |
142 | static int idx; |
143 | static void *buf [buffers]; |
143 | static void *buf [buffers]; |
144 | static size_t len [buffers]; |
144 | static size_t len [buffers]; |
145 | |
145 | |
146 | idx = (idx + 1) % buffers; |
146 | idx = (idx + 1) % buffers; |
… | |
… | |
151 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
151 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
152 | buf [idx] = malloc (len [idx]); |
152 | buf [idx] = malloc (len [idx]); |
153 | } |
153 | } |
154 | |
154 | |
155 | return buf [idx]; |
155 | return buf [idx]; |
|
|
156 | } |
|
|
157 | |
|
|
158 | static const char * ecb_noinline |
|
|
159 | cv_get_name (CV *cv) |
|
|
160 | { |
|
|
161 | static char fullname [256]; |
|
|
162 | |
|
|
163 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
164 | |
|
|
165 | HV *stash = GvSTASH (gv); |
|
|
166 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
167 | const char *gvname = GvNAME (gv); |
|
|
168 | |
|
|
169 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
170 | return fullname; |
156 | } |
171 | } |
157 | |
172 | |
158 | /*****************************************************************************/ |
173 | /*****************************************************************************/ |
159 | |
174 | |
160 | typedef struct |
175 | typedef struct |
… | |
… | |
219 | if (res) \ |
234 | if (res) \ |
220 | FAIL (name); |
235 | FAIL (name); |
221 | |
236 | |
222 | /*****************************************************************************/ |
237 | /*****************************************************************************/ |
223 | |
238 | |
|
|
239 | static SV * |
|
|
240 | new_clobj (HV *stash, IV id) |
|
|
241 | { |
|
|
242 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
243 | } |
|
|
244 | |
|
|
245 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
246 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
247 | |
|
|
248 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
249 | /* they can be upgraded at runtime to the array form */ |
|
|
250 | static void * ecb_noinline |
|
|
251 | SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
252 | { |
|
|
253 | // sv_derived_from is quite slow :( |
|
|
254 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
255 | return (void *)SvIV (SvRV (sv)); |
|
|
256 | |
|
|
257 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
258 | } |
|
|
259 | |
|
|
260 | // the "no-inherit" version of the above |
|
|
261 | static void * ecb_noinline |
|
|
262 | SvCLOBJ_ni (const char *func, const char *svname, SV *sv, HV *stash) |
|
|
263 | { |
|
|
264 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
265 | return (void *)SvIV (SvRV (sv)); |
|
|
266 | |
|
|
267 | croak ("%s: %s is not of type %s", func, svname, HvNAME (stash)); |
|
|
268 | } |
|
|
269 | |
|
|
270 | /*****************************************************************************/ |
|
|
271 | |
224 | static cl_context_properties * |
272 | static cl_context_properties * ecb_noinline |
225 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
273 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
226 | { |
274 | { |
227 | if (!sv || !SvOK (sv)) |
275 | if (!sv || !SvOK (sv)) |
228 | if (extra) |
276 | if (extra) |
229 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
277 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
… | |
… | |
236 | int i, len = av_len (av) + 1; |
284 | int i, len = av_len (av) + 1; |
237 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
285 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
238 | cl_context_properties *l = p; |
286 | cl_context_properties *l = p; |
239 | |
287 | |
240 | if (len & 1) |
288 | if (len & 1) |
241 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
289 | croak ("%s: %s is not a property list (must contain an even number of elements)", func, svname); |
242 | |
290 | |
243 | while (extracount--) |
291 | while (extracount--) |
244 | *l++ = *extra++; |
292 | *l++ = *extra++; |
245 | |
293 | |
246 | for (i = 0; i < len; i += 2) |
294 | for (i = 0; i < len; i += 2) |
… | |
… | |
249 | SV *p_sv = *av_fetch (av, i + 1, 0); |
297 | SV *p_sv = *av_fetch (av, i + 1, 0); |
250 | cl_context_properties v = SvIV (p_sv); // code below can override |
298 | cl_context_properties v = SvIV (p_sv); // code below can override |
251 | |
299 | |
252 | switch (t) |
300 | switch (t) |
253 | { |
301 | { |
|
|
302 | case CL_CONTEXT_PLATFORM: |
|
|
303 | if (SvROK (p_sv)) |
|
|
304 | v = (cl_context_properties)SvCLOBJ (func, svname, p_sv, "OpenCL::Platform"); |
|
|
305 | break; |
|
|
306 | |
254 | case CL_GLX_DISPLAY_KHR: |
307 | case CL_GLX_DISPLAY_KHR: |
255 | if (!SvOK (p_sv)) |
308 | if (!SvOK (p_sv)) |
256 | { |
309 | { |
257 | void *func = glsym ("_glXGetCurrentDisplay"); |
310 | void *func = glsym ("_glXGetCurrentDisplay"); |
258 | if (func) |
311 | if (func) |
… | |
… | |
284 | } |
337 | } |
285 | |
338 | |
286 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
339 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
287 | } |
340 | } |
288 | |
341 | |
289 | /*****************************************************************************/ |
342 | static void * ecb_noinline |
290 | |
343 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
291 | static SV * |
|
|
292 | new_clobj (HV *stash, IV id) |
|
|
293 | { |
344 | { |
294 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
345 | const char *funcname = cv_get_name (cv); |
295 | } |
|
|
296 | |
346 | |
297 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
347 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
298 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
348 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
349 | funcname, argname, or_undef ? "undef or " : "", klass); |
299 | |
350 | |
300 | /* cl objects are either \$iv, or [$iv, ...] */ |
351 | AV *av = (AV *)SvRV (arg); |
301 | /* they can be upgraded at runtime to the array form */ |
352 | void **list = 0; |
302 | static void * |
353 | cl_uint count = av_len (av) + 1; |
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 | |
354 | |
308 | croak ("%s: %s is not of type %s", func, svname, pkg); |
355 | if (count) |
|
|
356 | { |
|
|
357 | list = tmpbuf (sizeof (*list) * count); |
|
|
358 | int i; |
|
|
359 | for (i = 0; i < count; ++i) |
|
|
360 | list [i] = SvCLOBJ (funcname, argname, *av_fetch (av, i, 1), klass); |
|
|
361 | } |
|
|
362 | |
|
|
363 | if (!count && !or_undef) |
|
|
364 | croak ("%s: '%s' must contain at least one %s object", |
|
|
365 | funcname, argname, klass); |
|
|
366 | |
|
|
367 | *rcount = count; |
|
|
368 | return (void *)list; |
309 | } |
369 | } |
310 | |
370 | |
311 | /*****************************************************************************/ |
371 | /*****************************************************************************/ |
312 | /* callback stuff */ |
372 | /* callback stuff */ |
313 | |
373 | |
… | |
… | |
335 | static void (*eq_signal_func)(void *signal_arg, int value); |
395 | static void (*eq_signal_func)(void *signal_arg, int value); |
336 | static void *eq_signal_arg; |
396 | static void *eq_signal_arg; |
337 | static xmutex_t eq_lock = X_MUTEX_INIT; |
397 | static xmutex_t eq_lock = X_MUTEX_INIT; |
338 | static eq_item *eq_head, *eq_tail; |
398 | static eq_item *eq_head, *eq_tail; |
339 | |
399 | |
340 | static void |
400 | static void ecb_noinline |
341 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
401 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
342 | { |
402 | { |
343 | eq_item *item = malloc (sizeof (eq_item)); |
403 | eq_item *item = malloc (sizeof (eq_item)); |
344 | |
404 | |
345 | item->next = 0; |
405 | item->next = 0; |
… | |
… | |
412 | } |
472 | } |
413 | |
473 | |
414 | /*****************************************************************************/ |
474 | /*****************************************************************************/ |
415 | /* context notify */ |
475 | /* context notify */ |
416 | |
476 | |
417 | static void |
477 | static void ecb_noinline |
418 | eq_context_push (void *data1, void *data2, void *data3) |
478 | eq_context_push (void *data1, void *data2, void *data3) |
419 | { |
479 | { |
420 | dSP; |
480 | dSP; |
421 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
481 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
422 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
482 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
… | |
… | |
444 | { \ |
504 | { \ |
445 | pfn_notify = eq_context_notify; \ |
505 | pfn_notify = eq_context_notify; \ |
446 | user_data = s_get_cv (notify); \ |
506 | user_data = s_get_cv (notify); \ |
447 | } |
507 | } |
448 | |
508 | |
449 | static SV * |
509 | static SV * ecb_noinline |
450 | new_clobj_context (cl_context ctx, void *user_data) |
510 | new_clobj_context (cl_context ctx, void *user_data) |
451 | { |
511 | { |
452 | SV *sv = new_clobj (stash_context, (IV)ctx); |
512 | SV *sv = new_clobj (stash_context, (IV)ctx); |
453 | |
513 | |
454 | if (user_data) |
514 | if (user_data) |
… | |
… | |
473 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
533 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
474 | |
534 | |
475 | static void CL_CALLBACK |
535 | static void CL_CALLBACK |
476 | eq_program_notify (cl_program program, void *user_data) |
536 | eq_program_notify (cl_program program, void *user_data) |
477 | { |
537 | { |
|
|
538 | clRetainProgram (program); |
|
|
539 | |
478 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
540 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
541 | } |
|
|
542 | |
|
|
543 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
544 | |
|
|
545 | static program_callback ecb_noinline |
|
|
546 | make_program_callback (SV *notify, void **ruser_data) |
|
|
547 | { |
|
|
548 | if (SvOK (notify)) |
|
|
549 | { |
|
|
550 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
551 | return eq_program_notify; |
|
|
552 | } |
|
|
553 | else |
|
|
554 | { |
|
|
555 | *ruser_data = 0; |
|
|
556 | return 0; |
|
|
557 | } |
479 | } |
558 | } |
480 | |
559 | |
481 | struct build_args |
560 | struct build_args |
482 | { |
561 | { |
483 | cl_program program; |
562 | cl_program program; |
… | |
… | |
497 | else |
576 | else |
498 | clReleaseProgram (arg->program); |
577 | clReleaseProgram (arg->program); |
499 | |
578 | |
500 | free (arg->options); |
579 | free (arg->options); |
501 | free (arg); |
580 | free (arg); |
|
|
581 | |
|
|
582 | return 0; |
502 | } |
583 | } |
503 | |
584 | |
504 | static void |
585 | 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) |
586 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
506 | { |
587 | { |
… | |
… | |
546 | size_t res; |
627 | size_t res; |
547 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
628 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
548 | return res; |
629 | return res; |
549 | } |
630 | } |
550 | |
631 | |
551 | static cl_event * |
632 | static cl_event * ecb_noinline |
552 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
633 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
553 | { |
634 | { |
554 | cl_uint count = *rcount; |
635 | cl_uint count = *rcount; |
555 | |
636 | |
556 | if (count > 0x7fffffff) // yeha, it's a hack - the caller might have underflowed |
637 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
557 | count = 0; |
638 | *rcount = count = 0; |
558 | |
639 | |
559 | if (!count && !extra) |
640 | if (!count && !extra) |
560 | return 0; |
641 | return 0; |
561 | |
642 | |
562 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
643 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
… | |
… | |
719 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
800 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
720 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
801 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
721 | |
802 | |
722 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
803 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
723 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
804 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
805 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
724 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
806 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
725 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
807 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
726 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
808 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
727 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
809 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
728 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
810 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
… | |
… | |
769 | list = tmpbuf (sizeof (*list) * count); |
851 | list = tmpbuf (sizeof (*list) * count); |
770 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
852 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
771 | |
853 | |
772 | EXTEND (SP, count); |
854 | EXTEND (SP, count); |
773 | for (i = 0; i < count; ++i) |
855 | for (i = 0; i < count; ++i) |
774 | PUSH_CLOBJ (stash_platform, (IV)list [i]); |
856 | PUSH_CLOBJ (stash_platform, list [i]); |
775 | |
857 | |
776 | void |
858 | void |
777 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
859 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
778 | PPCODE: |
860 | PPCODE: |
779 | CONTEXT_NOTIFY_CALLBACK; |
861 | CONTEXT_NOTIFY_CALLBACK; |
780 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
862 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
781 | XPUSH_CLOBJ_CONTEXT; |
863 | XPUSH_CLOBJ_CONTEXT; |
782 | |
864 | |
783 | void |
865 | void |
784 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
866 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
785 | PPCODE: |
867 | PPCODE: |
… | |
… | |
842 | EXTEND (SP, count); |
924 | EXTEND (SP, count); |
843 | for (i = 0; i < count; ++i) |
925 | for (i = 0; i < count; ++i) |
844 | PUSH_CLOBJ (stash_device, list [i]); |
926 | PUSH_CLOBJ (stash_device, list [i]); |
845 | |
927 | |
846 | void |
928 | void |
847 | context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
929 | context (OpenCL::Platform self, SV *properties = 0, SV *devices, SV *notify = &PL_sv_undef) |
848 | PPCODE: |
930 | PPCODE: |
849 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
931 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
850 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
932 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context", "properties", properties, extra, 2); |
851 | |
933 | |
852 | AV *av = (AV *)SvRV (devices); |
934 | cl_uint device_count = 0; |
853 | cl_uint num_devices = av_len (av) + 1; |
935 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
854 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
|
|
855 | |
|
|
856 | int i; |
|
|
857 | for (i = num_devices; i--; ) |
|
|
858 | device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
|
|
859 | |
936 | |
860 | CONTEXT_NOTIFY_CALLBACK; |
937 | CONTEXT_NOTIFY_CALLBACK; |
861 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res)); |
938 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
862 | XPUSH_CLOBJ_CONTEXT; |
939 | XPUSH_CLOBJ_CONTEXT; |
863 | |
940 | |
864 | void |
941 | void |
865 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
942 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
866 | PPCODE: |
943 | PPCODE: |
867 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
944 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
868 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
945 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
869 | |
946 | |
870 | CONTEXT_NOTIFY_CALLBACK; |
947 | CONTEXT_NOTIFY_CALLBACK; |
871 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
948 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
872 | XPUSH_CLOBJ_CONTEXT; |
949 | XPUSH_CLOBJ_CONTEXT; |
873 | |
950 | |
874 | MODULE = OpenCL PACKAGE = OpenCL::Device |
951 | MODULE = OpenCL PACKAGE = OpenCL::Device |
875 | |
952 | |
876 | void |
953 | void |
877 | info (OpenCL::Device self, cl_device_info name) |
954 | info (OpenCL::Device self, cl_device_info name) |
878 | PPCODE: |
955 | PPCODE: |
879 | INFO (Device) |
956 | INFO (Device) |
|
|
957 | |
|
|
958 | #if CL_VERSION_1_2 |
|
|
959 | |
|
|
960 | void |
|
|
961 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
962 | PPCODE: |
|
|
963 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
|
|
964 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
|
|
965 | |
|
|
966 | properties = SvRV (properties); |
|
|
967 | |
|
|
968 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
969 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
970 | |
|
|
971 | int i; |
|
|
972 | for (i = 0; i < count; ++i) |
|
|
973 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
|
|
974 | |
|
|
975 | props [count] = 0; |
|
|
976 | |
|
|
977 | cl_uint num_devices; |
|
|
978 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
979 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
980 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
|
|
981 | |
|
|
982 | EXTEND (SP, num_devices); |
|
|
983 | for (i = 0; i < count; ++i) |
|
|
984 | PUSH_CLOBJ (stash_subdevice, list [i]); |
|
|
985 | |
|
|
986 | #endif |
880 | |
987 | |
881 | #BEGIN:device |
988 | #BEGIN:device |
882 | |
989 | |
883 | void |
990 | void |
884 | type (OpenCL::Device self) |
991 | type (OpenCL::Device self) |
… | |
… | |
1100 | for (i = 0; i < n; ++i) |
1207 | for (i = 0; i < n; ++i) |
1101 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1208 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1102 | |
1209 | |
1103 | #END:device |
1210 | #END:device |
1104 | |
1211 | |
|
|
1212 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1213 | |
|
|
1214 | #if CL_VERSION_1_2 |
|
|
1215 | |
|
|
1216 | void |
|
|
1217 | DESTROY (OpenCL::SubDevice self) |
|
|
1218 | CODE: |
|
|
1219 | clReleaseDevice (self); |
|
|
1220 | |
|
|
1221 | #endif |
|
|
1222 | |
1105 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1223 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1106 | |
1224 | |
1107 | void |
1225 | void |
1108 | DESTROY (OpenCL::Context context) |
1226 | DESTROY (OpenCL::Context self) |
1109 | CODE: |
1227 | CODE: |
1110 | clReleaseContext (context); |
1228 | clReleaseContext (self); |
1111 | |
1229 | |
1112 | void |
1230 | void |
1113 | info (OpenCL::Context self, cl_context_info name) |
1231 | info (OpenCL::Context self, cl_context_info name) |
1114 | PPCODE: |
1232 | PPCODE: |
1115 | INFO (Context) |
1233 | INFO (Context) |
… | |
… | |
1299 | |
1417 | |
1300 | len2 = len; |
1418 | len2 = len; |
1301 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1419 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1302 | XPUSH_CLOBJ (stash_program, prog); |
1420 | XPUSH_CLOBJ (stash_program, prog); |
1303 | |
1421 | |
|
|
1422 | void |
|
|
1423 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1424 | PPCODE: |
|
|
1425 | cl_uint device_count; |
|
|
1426 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1427 | |
|
|
1428 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1429 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1430 | |
|
|
1431 | binaries = SvRV (binaries); |
|
|
1432 | |
|
|
1433 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1434 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1435 | |
|
|
1436 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1437 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1438 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1439 | |
|
|
1440 | int i; |
|
|
1441 | for (i = 0; i < device_count; ++i) |
|
|
1442 | { |
|
|
1443 | STRLEN len; |
|
|
1444 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1445 | length_list [i] = len; |
|
|
1446 | } |
|
|
1447 | |
|
|
1448 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1449 | length_list, binary_list, |
|
|
1450 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1451 | |
|
|
1452 | EXTEND (SP, 2); |
|
|
1453 | PUSH_CLOBJ (stash_program, prog); |
|
|
1454 | |
|
|
1455 | if (GIMME_V == G_ARRAY) |
|
|
1456 | { |
|
|
1457 | AV *av = newAV (); |
|
|
1458 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1459 | |
|
|
1460 | for (i = device_count; i--; ) |
|
|
1461 | av_store (av, i, newSViv (status_list [i])); |
|
|
1462 | } |
|
|
1463 | |
|
|
1464 | #if CL_VERSION_1_2 |
|
|
1465 | |
|
|
1466 | void |
|
|
1467 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1468 | PPCODE: |
|
|
1469 | cl_uint device_count = 0; |
|
|
1470 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1471 | |
|
|
1472 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1473 | |
|
|
1474 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1475 | |
|
|
1476 | void |
|
|
1477 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1478 | CODE: |
|
|
1479 | cl_uint device_count = 0; |
|
|
1480 | cl_device_id *device_list = 0; |
|
|
1481 | |
|
|
1482 | if (SvOK (devices)) |
|
|
1483 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1484 | |
|
|
1485 | cl_uint program_count; |
|
|
1486 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1487 | |
|
|
1488 | void *user_data; |
|
|
1489 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1490 | |
|
|
1491 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1492 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1493 | |
|
|
1494 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1495 | |
|
|
1496 | #endif |
|
|
1497 | |
1304 | #BEGIN:context |
1498 | #BEGIN:context |
1305 | |
1499 | |
1306 | void |
1500 | void |
1307 | reference_count (OpenCL::Context self) |
1501 | reference_count (OpenCL::Context self) |
1308 | ALIAS: |
1502 | ALIAS: |
… | |
… | |
1353 | void |
1547 | void |
1354 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1548 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1355 | ALIAS: |
1549 | ALIAS: |
1356 | enqueue_read_buffer = 0 |
1550 | enqueue_read_buffer = 0 |
1357 | PPCODE: |
1551 | PPCODE: |
1358 | cl_event ev = 0; |
|
|
1359 | EVENT_LIST (6); |
1552 | EVENT_LIST (6); |
1360 | |
1553 | |
1361 | SvUPGRADE (data, SVt_PV); |
1554 | SvUPGRADE (data, SVt_PV); |
1362 | SvGROW (data, len); |
1555 | SvGROW (data, len); |
1363 | SvPOK_only (data); |
1556 | SvPOK_only (data); |
1364 | SvCUR_set (data, len); |
1557 | SvCUR_set (data, len); |
|
|
1558 | |
|
|
1559 | cl_event ev = 0; |
1365 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1560 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1366 | |
1561 | |
1367 | if (ev) |
1562 | if (ev) |
1368 | XPUSH_CLOBJ (stash_event, ev); |
1563 | XPUSH_CLOBJ (stash_event, ev); |
1369 | |
1564 | |
1370 | void |
1565 | void |
1371 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1566 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1372 | ALIAS: |
1567 | ALIAS: |
1373 | enqueue_write_buffer = 0 |
1568 | enqueue_write_buffer = 0 |
1374 | PPCODE: |
1569 | PPCODE: |
1375 | cl_event ev = 0; |
1570 | EVENT_LIST (5); |
|
|
1571 | |
1376 | STRLEN len; |
1572 | STRLEN len; |
1377 | char *ptr = SvPVbyte (data, len); |
1573 | char *ptr = SvPVbyte (data, len); |
1378 | EVENT_LIST (5); |
|
|
1379 | |
1574 | |
|
|
1575 | cl_event ev = 0; |
1380 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1576 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1381 | |
1577 | |
1382 | if (ev) |
1578 | if (ev) |
1383 | XPUSH_CLOBJ (stash_event, ev); |
1579 | XPUSH_CLOBJ (stash_event, ev); |
1384 | |
1580 | |
… | |
… | |
1387 | void |
1583 | void |
1388 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
1584 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
1389 | ALIAS: |
1585 | ALIAS: |
1390 | enqueue_fill_buffer = 0 |
1586 | enqueue_fill_buffer = 0 |
1391 | PPCODE: |
1587 | PPCODE: |
1392 | cl_event ev = 0; |
1588 | EVENT_LIST (5); |
|
|
1589 | |
1393 | STRLEN len; |
1590 | STRLEN len; |
1394 | char *ptr = SvPVbyte (data, len); |
1591 | char *ptr = SvPVbyte (data, len); |
1395 | EVENT_LIST (5); |
|
|
1396 | |
1592 | |
|
|
1593 | cl_event ev = 0; |
1397 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1594 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1398 | |
1595 | |
1399 | if (ev) |
1596 | if (ev) |
1400 | XPUSH_CLOBJ (stash_event, ev); |
1597 | XPUSH_CLOBJ (stash_event, ev); |
1401 | |
1598 | |
1402 | void |
1599 | void |
1403 | 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, ...) |
1600 | 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, ...) |
1404 | ALIAS: |
1601 | ALIAS: |
1405 | enqueue_fill_image = 0 |
1602 | enqueue_fill_image = 0 |
1406 | PPCODE: |
1603 | PPCODE: |
1407 | cl_event ev = 0; |
1604 | EVENT_LIST (12); |
1408 | STRLEN len; |
1605 | |
1409 | const size_t origin [3] = { x, y, z }; |
1606 | const size_t origin [3] = { x, y, z }; |
1410 | const size_t region [3] = { width, height, depth }; |
1607 | const size_t region [3] = { width, height, depth }; |
1411 | EVENT_LIST (12); |
|
|
1412 | |
1608 | |
1413 | const cl_float c_f [4] = { r, g, b, a }; |
1609 | const cl_float c_f [4] = { r, g, b, a }; |
1414 | const cl_uint c_u [4] = { r, g, b, a }; |
1610 | const cl_uint c_u [4] = { r, g, b, a }; |
1415 | const cl_int c_s [4] = { r, g, b, a }; |
1611 | const cl_int c_s [4] = { r, g, b, a }; |
1416 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
1612 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
… | |
… | |
1419 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
1615 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
1420 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
1616 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
1421 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
1617 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
1422 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
1618 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
1423 | |
1619 | |
|
|
1620 | cl_event ev = 0; |
1424 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
1621 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
1425 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1622 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1426 | |
1623 | |
1427 | if (ev) |
1624 | if (ev) |
1428 | XPUSH_CLOBJ (stash_event, ev); |
1625 | XPUSH_CLOBJ (stash_event, ev); |
… | |
… | |
1432 | void |
1629 | void |
1433 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1630 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1434 | ALIAS: |
1631 | ALIAS: |
1435 | enqueue_copy_buffer = 0 |
1632 | enqueue_copy_buffer = 0 |
1436 | PPCODE: |
1633 | PPCODE: |
|
|
1634 | EVENT_LIST (6); |
|
|
1635 | |
1437 | cl_event ev = 0; |
1636 | cl_event ev = 0; |
1438 | EVENT_LIST (6); |
|
|
1439 | |
|
|
1440 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1637 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1441 | |
1638 | |
1442 | if (ev) |
1639 | if (ev) |
1443 | XPUSH_CLOBJ (stash_event, ev); |
1640 | XPUSH_CLOBJ (stash_event, ev); |
1444 | |
1641 | |
1445 | void |
1642 | void |
1446 | 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, ...) |
1643 | 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, ...) |
1447 | ALIAS: |
1644 | ALIAS: |
1448 | enqueue_read_buffer_rect = 0 |
1645 | enqueue_read_buffer_rect = 0 |
1449 | PPCODE: |
1646 | PPCODE: |
1450 | cl_event ev = 0; |
1647 | EVENT_LIST (17); |
|
|
1648 | |
1451 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1649 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1452 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1650 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1453 | const size_t region[3] = { width, height, depth }; |
1651 | const size_t region[3] = { width, height, depth }; |
1454 | EVENT_LIST (17); |
|
|
1455 | |
1652 | |
1456 | if (!buf_row_pitch) |
1653 | if (!buf_row_pitch) |
1457 | buf_row_pitch = region [0]; |
1654 | buf_row_pitch = region [0]; |
1458 | |
1655 | |
1459 | if (!buf_slice_pitch) |
1656 | if (!buf_slice_pitch) |
… | |
… | |
1469 | |
1666 | |
1470 | SvUPGRADE (data, SVt_PV); |
1667 | SvUPGRADE (data, SVt_PV); |
1471 | SvGROW (data, len); |
1668 | SvGROW (data, len); |
1472 | SvPOK_only (data); |
1669 | SvPOK_only (data); |
1473 | SvCUR_set (data, len); |
1670 | SvCUR_set (data, len); |
|
|
1671 | |
|
|
1672 | cl_event ev = 0; |
1474 | 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)); |
1673 | 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)); |
1475 | |
1674 | |
1476 | if (ev) |
1675 | if (ev) |
1477 | XPUSH_CLOBJ (stash_event, ev); |
1676 | XPUSH_CLOBJ (stash_event, ev); |
1478 | |
1677 | |
1479 | void |
1678 | void |
1480 | 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, ...) |
1679 | 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, ...) |
1481 | ALIAS: |
1680 | ALIAS: |
1482 | enqueue_write_buffer_rect = 0 |
1681 | enqueue_write_buffer_rect = 0 |
1483 | PPCODE: |
1682 | PPCODE: |
1484 | cl_event ev = 0; |
1683 | EVENT_LIST (17); |
|
|
1684 | |
1485 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1685 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1486 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1686 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1487 | const size_t region[3] = { width, height, depth }; |
1687 | const size_t region[3] = { width, height, depth }; |
1488 | STRLEN len; |
1688 | STRLEN len; |
1489 | char *ptr = SvPVbyte (data, len); |
1689 | char *ptr = SvPVbyte (data, len); |
1490 | EVENT_LIST (17); |
|
|
1491 | |
1690 | |
1492 | if (!buf_row_pitch) |
1691 | if (!buf_row_pitch) |
1493 | buf_row_pitch = region [0]; |
1692 | buf_row_pitch = region [0]; |
1494 | |
1693 | |
1495 | if (!buf_slice_pitch) |
1694 | if (!buf_slice_pitch) |
… | |
… | |
1504 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1703 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1505 | |
1704 | |
1506 | if (len < min_len) |
1705 | if (len < min_len) |
1507 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1706 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1508 | |
1707 | |
|
|
1708 | cl_event ev = 0; |
1509 | 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)); |
1709 | 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)); |
1510 | |
1710 | |
1511 | if (ev) |
1711 | if (ev) |
1512 | XPUSH_CLOBJ (stash_event, ev); |
1712 | XPUSH_CLOBJ (stash_event, ev); |
1513 | |
1713 | |
1514 | void |
1714 | void |
1515 | 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, ...) |
1715 | 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, ...) |
1516 | ALIAS: |
1716 | ALIAS: |
1517 | enqueue_copy_buffer_rect = 0 |
1717 | enqueue_copy_buffer_rect = 0 |
1518 | PPCODE: |
1718 | PPCODE: |
1519 | cl_event ev = 0; |
1719 | EVENT_LIST (16); |
|
|
1720 | |
1520 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1721 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1521 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1722 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1522 | const size_t region[3] = { width, height, depth }; |
1723 | const size_t region[3] = { width, height, depth }; |
1523 | EVENT_LIST (16); |
|
|
1524 | |
1724 | |
|
|
1725 | cl_event ev = 0; |
1525 | 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)); |
1726 | 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)); |
1526 | |
1727 | |
1527 | if (ev) |
1728 | if (ev) |
1528 | XPUSH_CLOBJ (stash_event, ev); |
1729 | XPUSH_CLOBJ (stash_event, ev); |
1529 | |
1730 | |
1530 | void |
1731 | void |
1531 | 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, ...) |
1732 | 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, ...) |
1532 | ALIAS: |
1733 | ALIAS: |
1533 | enqueue_read_image = 0 |
1734 | enqueue_read_image = 0 |
1534 | PPCODE: |
1735 | PPCODE: |
1535 | cl_event ev = 0; |
1736 | EVENT_LIST (12); |
|
|
1737 | |
1536 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1738 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1537 | const size_t region[3] = { width, height, depth }; |
1739 | const size_t region[3] = { width, height, depth }; |
1538 | EVENT_LIST (12); |
|
|
1539 | |
1740 | |
1540 | if (!row_pitch) |
1741 | if (!row_pitch) |
1541 | row_pitch = img_row_pitch (src); |
1742 | row_pitch = img_row_pitch (src); |
1542 | |
1743 | |
1543 | if (depth > 1 && !slice_pitch) |
1744 | if (depth > 1 && !slice_pitch) |
… | |
… | |
1547 | |
1748 | |
1548 | SvUPGRADE (data, SVt_PV); |
1749 | SvUPGRADE (data, SVt_PV); |
1549 | SvGROW (data, len); |
1750 | SvGROW (data, len); |
1550 | SvPOK_only (data); |
1751 | SvPOK_only (data); |
1551 | SvCUR_set (data, len); |
1752 | SvCUR_set (data, len); |
|
|
1753 | |
|
|
1754 | cl_event ev = 0; |
1552 | 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)); |
1755 | 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)); |
1553 | |
1756 | |
1554 | if (ev) |
1757 | if (ev) |
1555 | XPUSH_CLOBJ (stash_event, ev); |
1758 | XPUSH_CLOBJ (stash_event, ev); |
1556 | |
1759 | |
1557 | void |
1760 | void |
1558 | 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, ...) |
1761 | 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, ...) |
1559 | ALIAS: |
1762 | ALIAS: |
1560 | enqueue_write_image = 0 |
1763 | enqueue_write_image = 0 |
1561 | PPCODE: |
1764 | PPCODE: |
1562 | cl_event ev = 0; |
1765 | EVENT_LIST (12); |
|
|
1766 | |
1563 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1767 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1564 | const size_t region[3] = { width, height, depth }; |
1768 | const size_t region[3] = { width, height, depth }; |
1565 | STRLEN len; |
1769 | STRLEN len; |
1566 | char *ptr = SvPVbyte (data, len); |
1770 | char *ptr = SvPVbyte (data, len); |
1567 | EVENT_LIST (12); |
|
|
1568 | |
1771 | |
1569 | if (!row_pitch) |
1772 | if (!row_pitch) |
1570 | row_pitch = img_row_pitch (dst); |
1773 | row_pitch = img_row_pitch (dst); |
1571 | |
1774 | |
1572 | if (depth > 1 && !slice_pitch) |
1775 | if (depth > 1 && !slice_pitch) |
… | |
… | |
1575 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1778 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1576 | |
1779 | |
1577 | if (len < min_len) |
1780 | if (len < min_len) |
1578 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1781 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1579 | |
1782 | |
|
|
1783 | cl_event ev = 0; |
1580 | 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)); |
1784 | 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)); |
1581 | |
1785 | |
1582 | if (ev) |
1786 | if (ev) |
1583 | XPUSH_CLOBJ (stash_event, ev); |
1787 | XPUSH_CLOBJ (stash_event, ev); |
1584 | |
1788 | |
1585 | void |
1789 | void |
1586 | 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, ...) |
1790 | 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, ...) |
1587 | ALIAS: |
1791 | ALIAS: |
1588 | enqueue_copy_image = 0 |
1792 | enqueue_copy_image = 0 |
1589 | PPCODE: |
1793 | PPCODE: |
1590 | cl_event ev = 0; |
1794 | EVENT_LIST (12); |
|
|
1795 | |
1591 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1796 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1592 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1797 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1593 | const size_t region[3] = { width, height, depth }; |
1798 | const size_t region[3] = { width, height, depth }; |
1594 | EVENT_LIST (12); |
|
|
1595 | |
1799 | |
|
|
1800 | cl_event ev = 0; |
1596 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1801 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1597 | |
1802 | |
1598 | if (ev) |
1803 | if (ev) |
1599 | XPUSH_CLOBJ (stash_event, ev); |
1804 | XPUSH_CLOBJ (stash_event, ev); |
1600 | |
1805 | |
1601 | void |
1806 | void |
1602 | 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, ...) |
1807 | 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, ...) |
1603 | ALIAS: |
1808 | ALIAS: |
1604 | enqueue_copy_image_to_buffer = 0 |
1809 | enqueue_copy_image_to_buffer = 0 |
1605 | PPCODE: |
1810 | PPCODE: |
1606 | cl_event ev = 0; |
1811 | EVENT_LIST (10); |
|
|
1812 | |
1607 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1813 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1608 | const size_t region [3] = { width, height, depth }; |
1814 | const size_t region [3] = { width, height, depth }; |
1609 | EVENT_LIST (10); |
|
|
1610 | |
1815 | |
|
|
1816 | cl_event ev = 0; |
1611 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1817 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1612 | |
1818 | |
1613 | if (ev) |
1819 | if (ev) |
1614 | XPUSH_CLOBJ (stash_event, ev); |
1820 | XPUSH_CLOBJ (stash_event, ev); |
1615 | |
1821 | |
1616 | void |
1822 | void |
1617 | 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, ...) |
1823 | 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, ...) |
1618 | ALIAS: |
1824 | ALIAS: |
1619 | enqueue_copy_buffer_to_image = 0 |
1825 | enqueue_copy_buffer_to_image = 0 |
1620 | PPCODE: |
1826 | PPCODE: |
1621 | cl_event ev = 0; |
1827 | EVENT_LIST (10); |
|
|
1828 | |
1622 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1829 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1623 | const size_t region [3] = { width, height, depth }; |
1830 | const size_t region [3] = { width, height, depth }; |
1624 | EVENT_LIST (10); |
|
|
1625 | |
1831 | |
|
|
1832 | cl_event ev = 0; |
1626 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1833 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1627 | |
1834 | |
1628 | if (ev) |
1835 | if (ev) |
1629 | XPUSH_CLOBJ (stash_event, ev); |
1836 | XPUSH_CLOBJ (stash_event, ev); |
1630 | |
1837 | |
1631 | void |
1838 | void |
1632 | 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, ...) |
1839 | 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, SV *cb_ = &PL_sv_undef, ...) |
1633 | ALIAS: |
1840 | ALIAS: |
1634 | enqueue_map_buffer = 0 |
1841 | enqueue_map_buffer = 0 |
1635 | PPCODE: |
1842 | PPCODE: |
|
|
1843 | EVENT_LIST (6); |
|
|
1844 | |
|
|
1845 | size_t cb = SvIV (cb_); |
|
|
1846 | |
|
|
1847 | if (!SvOK (cb_)) |
|
|
1848 | { |
|
|
1849 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1850 | cb -= offset; |
|
|
1851 | } |
|
|
1852 | |
1636 | cl_event ev; |
1853 | cl_event ev; |
1637 | EVENT_LIST (6); |
|
|
1638 | |
|
|
1639 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
1854 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
1640 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
1855 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
1641 | |
1856 | |
1642 | void |
1857 | void |
1643 | 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, ...) |
1858 | 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, SV *width_ = &PL_sv_undef, SV *height_ = &PL_sv_undef, SV *depth_ = &PL_sv_undef, ...) |
1644 | ALIAS: |
1859 | ALIAS: |
1645 | enqueue_map_image = 0 |
1860 | enqueue_map_image = 0 |
1646 | PPCODE: |
1861 | PPCODE: |
1647 | cl_event ev; |
1862 | size_t width = SvIV (width_); |
|
|
1863 | if (!SvOK (width_)) |
|
|
1864 | { |
|
|
1865 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1866 | width -= x; |
|
|
1867 | } |
|
|
1868 | |
|
|
1869 | size_t height = SvIV (width_); |
|
|
1870 | if (!SvOK (height_)) |
|
|
1871 | { |
|
|
1872 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1873 | height -= y; |
|
|
1874 | } |
|
|
1875 | |
|
|
1876 | size_t depth = SvIV (width_); |
|
|
1877 | if (!SvOK (depth_)) |
|
|
1878 | { |
|
|
1879 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1880 | depth -= z; |
|
|
1881 | |
|
|
1882 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1883 | if (!depth) |
|
|
1884 | depth = 1; |
|
|
1885 | } |
|
|
1886 | |
1648 | const size_t origin[3] = { x, y, z }; |
1887 | const size_t origin[3] = { x, y, z }; |
1649 | const size_t region[3] = { width, height, depth }; |
1888 | const size_t region[3] = { width, height, depth }; |
1650 | size_t row_pitch, slice_pitch; |
1889 | size_t row_pitch, slice_pitch; |
1651 | EVENT_LIST (10); |
1890 | EVENT_LIST (10); |
1652 | |
1891 | |
|
|
1892 | cl_event ev; |
1653 | 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)); |
1893 | 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)); |
1654 | |
1894 | |
1655 | size_t cb = slice_pitch ? slice_pitch * region [2] |
1895 | size_t cb = slice_pitch ? slice_pitch * region [2] |
1656 | : row_pitch ? row_pitch * region [1] |
1896 | : row_pitch ? row_pitch * region [1] |
1657 | : region [0]; |
1897 | : region [0]; |
… | |
… | |
1671 | void |
1911 | void |
1672 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
1912 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
1673 | ALIAS: |
1913 | ALIAS: |
1674 | enqueue_task = 0 |
1914 | enqueue_task = 0 |
1675 | PPCODE: |
1915 | PPCODE: |
|
|
1916 | EVENT_LIST (2); |
|
|
1917 | |
1676 | cl_event ev = 0; |
1918 | cl_event ev = 0; |
1677 | EVENT_LIST (2); |
|
|
1678 | |
|
|
1679 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1919 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1680 | |
1920 | |
1681 | if (ev) |
1921 | if (ev) |
1682 | XPUSH_CLOBJ (stash_event, ev); |
1922 | XPUSH_CLOBJ (stash_event, ev); |
1683 | |
1923 | |
1684 | void |
1924 | void |
1685 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1925 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1686 | ALIAS: |
1926 | ALIAS: |
1687 | enqueue_nd_range_kernel = 0 |
1927 | enqueue_nd_range_kernel = 0 |
1688 | PPCODE: |
1928 | PPCODE: |
1689 | cl_event ev = 0; |
1929 | EVENT_LIST (5); |
|
|
1930 | |
1690 | size_t *gwo = 0, *gws, *lws = 0; |
1931 | size_t *gwo = 0, *gws, *lws = 0; |
1691 | int gws_len; |
1932 | int gws_len; |
1692 | size_t *lists; |
1933 | size_t *lists; |
1693 | int i; |
1934 | int i; |
1694 | EVENT_LIST (5); |
|
|
1695 | |
1935 | |
1696 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1936 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1697 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1937 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1698 | |
1938 | |
1699 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1939 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
… | |
… | |
1738 | if (!lws [i]) |
1978 | if (!lws [i]) |
1739 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
1979 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
1740 | } |
1980 | } |
1741 | } |
1981 | } |
1742 | |
1982 | |
|
|
1983 | cl_event ev = 0; |
1743 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1984 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1744 | |
1985 | |
1745 | if (ev) |
1986 | if (ev) |
1746 | XPUSH_CLOBJ (stash_event, ev); |
1987 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1988 | |
|
|
1989 | #if CL_VERSION_1_2 |
|
|
1990 | |
|
|
1991 | void |
|
|
1992 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
1993 | ALIAS: |
|
|
1994 | enqueue_migrate_mem_objects = 0 |
|
|
1995 | PPCODE: |
|
|
1996 | EVENT_LIST (3); |
|
|
1997 | |
|
|
1998 | cl_uint obj_count; |
|
|
1999 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2000 | |
|
|
2001 | cl_event ev = 0; |
|
|
2002 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2003 | |
|
|
2004 | if (ev) |
|
|
2005 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2006 | |
|
|
2007 | #endif |
1747 | |
2008 | |
1748 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2009 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1749 | |
2010 | |
1750 | void |
2011 | void |
1751 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2012 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1752 | ALIAS: |
2013 | ALIAS: |
1753 | release_gl_objects = 1 |
2014 | release_gl_objects = 1 |
1754 | enqueue_acquire_gl_objects = 0 |
2015 | enqueue_acquire_gl_objects = 0 |
1755 | enqueue_release_gl_objects = 1 |
2016 | enqueue_release_gl_objects = 1 |
1756 | PPCODE: |
2017 | PPCODE: |
1757 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
2018 | EVENT_LIST (2); |
1758 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
2019 | |
|
|
2020 | cl_uint obj_count; |
|
|
2021 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
1759 | |
2022 | |
1760 | cl_event ev = 0; |
2023 | cl_event ev = 0; |
1761 | EVENT_LIST (2); |
|
|
1762 | AV *av = (AV *)SvRV (objects); |
|
|
1763 | cl_uint num_objects = av_len (av) + 1; |
|
|
1764 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1765 | int i; |
|
|
1766 | |
|
|
1767 | for (i = num_objects; i--; ) |
|
|
1768 | object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1769 | |
2024 | |
1770 | if (ix) |
2025 | if (ix) |
1771 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2026 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1772 | else |
2027 | else |
1773 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2028 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1774 | |
2029 | |
1775 | if (ev) |
2030 | if (ev) |
1776 | XPUSH_CLOBJ (stash_event, ev); |
2031 | XPUSH_CLOBJ (stash_event, ev); |
1777 | |
2032 | |
1778 | #endif |
2033 | #endif |
… | |
… | |
1792 | void |
2047 | void |
1793 | marker (OpenCL::Queue self, ...) |
2048 | marker (OpenCL::Queue self, ...) |
1794 | ALIAS: |
2049 | ALIAS: |
1795 | enqueue_marker = 0 |
2050 | enqueue_marker = 0 |
1796 | PPCODE: |
2051 | PPCODE: |
|
|
2052 | EVENT_LIST (1); |
1797 | cl_event ev = 0; |
2053 | cl_event ev = 0; |
1798 | EVENT_LIST (1); |
|
|
1799 | #if PREFER_1_1 |
2054 | #if PREFER_1_1 |
1800 | if (!event_list_count) |
2055 | if (!event_list_count) |
1801 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
2056 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
1802 | else |
2057 | else |
1803 | #if CL_VERSION_1_2 |
2058 | #if CL_VERSION_1_2 |
… | |
… | |
1817 | void |
2072 | void |
1818 | barrier (OpenCL::Queue self, ...) |
2073 | barrier (OpenCL::Queue self, ...) |
1819 | ALIAS: |
2074 | ALIAS: |
1820 | enqueue_barrier = 0 |
2075 | enqueue_barrier = 0 |
1821 | PPCODE: |
2076 | PPCODE: |
|
|
2077 | EVENT_LIST (1); |
1822 | cl_event ev = 0; |
2078 | cl_event ev = 0; |
1823 | EVENT_LIST (1); |
|
|
1824 | #if PREFER_1_1 |
2079 | #if PREFER_1_1 |
1825 | if (!event_list_count && GIMME_V == G_VOID) |
2080 | if (!event_list_count && GIMME_V == G_VOID) |
1826 | NEED_SUCCESS (EnqueueBarrier, (self)); |
2081 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1827 | else |
2082 | else |
1828 | #if CL_VERSION_1_2 |
2083 | #if CL_VERSION_1_2 |
… | |
… | |
2156 | void |
2411 | void |
2157 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
2412 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
2158 | ALIAS: |
2413 | ALIAS: |
2159 | build_async = 1 |
2414 | build_async = 1 |
2160 | CODE: |
2415 | CODE: |
2161 | void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0; |
|
|
2162 | void *user_data = 0; |
|
|
2163 | cl_uint num_devices = 0; |
2416 | cl_uint device_count = 0; |
2164 | cl_device_id *device_list = 0; |
2417 | cl_device_id *device_list = 0; |
2165 | |
2418 | |
2166 | if (SvOK (devices)) |
2419 | if (SvOK (devices)) |
2167 | { |
2420 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2421 | |
|
|
2422 | void *user_data; |
|
|
2423 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2424 | |
|
|
2425 | if (ix) |
|
|
2426 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2427 | else |
|
|
2428 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2429 | |
|
|
2430 | #if CL_VERSION_1_2 |
|
|
2431 | |
|
|
2432 | void |
|
|
2433 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2434 | CODE: |
|
|
2435 | cl_uint device_count = 0; |
|
|
2436 | cl_device_id *device_list = 0; |
|
|
2437 | |
|
|
2438 | if (SvOK (devices)) |
|
|
2439 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2440 | |
|
|
2441 | cl_uint header_count = 0; |
|
|
2442 | cl_program *header_list = 0; |
|
|
2443 | const char **header_name = 0; |
|
|
2444 | |
|
|
2445 | if (SvOK (headers)) |
|
|
2446 | { |
2168 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
2447 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
2169 | croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); |
2448 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
2170 | |
2449 | |
2171 | AV *av = (AV *)SvRV (devices); |
2450 | HV *hv = (HV *)SvRV (devices); |
2172 | num_devices = av_len (av) + 1; |
|
|
2173 | |
2451 | |
2174 | if (num_devices) |
2452 | header_count = hv_iterinit (hv); |
|
|
2453 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2454 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2455 | |
|
|
2456 | HE *he; |
|
|
2457 | int i = 0; |
|
|
2458 | while (he = hv_iternext (hv)) |
2175 | { |
2459 | { |
2176 | device_list = tmpbuf (sizeof (*device_list) * num_devices); |
2460 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2461 | header_list [i] = SvCLOBJ ("clCompileProgram", "headers", HeVAL (he), "OpenCL::Program"); |
2177 | int count; |
2462 | ++i; |
2178 | for (count = 0; count < num_devices; ++count) |
|
|
2179 | device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device"); |
|
|
2180 | } |
2463 | } |
2181 | } |
2464 | } |
2182 | |
2465 | |
2183 | if (SvOK (notify)) |
2466 | void *user_data; |
2184 | { |
2467 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
2185 | NEED_SUCCESS (RetainProgram, (self)); |
|
|
2186 | pfn_notify = eq_program_notify; |
|
|
2187 | user_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
2188 | } |
|
|
2189 | |
2468 | |
2190 | if (ix) |
|
|
2191 | build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2192 | else |
|
|
2193 | NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
2469 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2470 | header_count, header_list, header_name, pfn_notify, user_data)); |
|
|
2471 | |
|
|
2472 | #endif |
2194 | |
2473 | |
2195 | void |
2474 | void |
2196 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2475 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2197 | PPCODE: |
2476 | PPCODE: |
2198 | size_t size; |
2477 | size_t size; |
… | |
… | |
2614 | const int i = 0; |
2893 | const int i = 0; |
2615 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2894 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2616 | |
2895 | |
2617 | #END:kernel_work_group |
2896 | #END:kernel_work_group |
2618 | |
2897 | |
|
|
2898 | #if CL_VERSION_1_2 |
|
|
2899 | |
|
|
2900 | void |
|
|
2901 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2902 | PPCODE: |
|
|
2903 | size_t size; |
|
|
2904 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2905 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2906 | SvUPGRADE (sv, SVt_PV); |
|
|
2907 | SvPOK_only (sv); |
|
|
2908 | SvCUR_set (sv, size); |
|
|
2909 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2910 | XPUSHs (sv); |
|
|
2911 | |
|
|
2912 | #BEGIN:kernel_arg |
|
|
2913 | |
|
|
2914 | void |
|
|
2915 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2916 | PPCODE: |
|
|
2917 | cl_kernel_arg_address_qualifier value [1]; |
|
|
2918 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2919 | EXTEND (SP, 1); |
|
|
2920 | const int i = 0; |
|
|
2921 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2922 | |
|
|
2923 | void |
|
|
2924 | arg_access_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2925 | PPCODE: |
|
|
2926 | cl_kernel_arg_access_qualifier value [1]; |
|
|
2927 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2928 | EXTEND (SP, 1); |
|
|
2929 | const int i = 0; |
|
|
2930 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2931 | |
|
|
2932 | void |
|
|
2933 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2934 | ALIAS: |
|
|
2935 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2936 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2937 | PPCODE: |
|
|
2938 | size_t size; |
|
|
2939 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2940 | char *value = tmpbuf (size); |
|
|
2941 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2942 | EXTEND (SP, 1); |
|
|
2943 | const int i = 0; |
|
|
2944 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2945 | |
|
|
2946 | void |
|
|
2947 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2948 | PPCODE: |
|
|
2949 | cl_kernel_arg_type_qualifier value [1]; |
|
|
2950 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2951 | EXTEND (SP, 1); |
|
|
2952 | const int i = 0; |
|
|
2953 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2954 | |
|
|
2955 | #END:kernel_arg |
|
|
2956 | |
|
|
2957 | #endif |
|
|
2958 | |
2619 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2959 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2620 | |
2960 | |
2621 | void |
2961 | void |
2622 | DESTROY (OpenCL::Event self) |
2962 | DESTROY (OpenCL::Event self) |
2623 | CODE: |
2963 | CODE: |
… | |
… | |
2770 | CODE: |
3110 | CODE: |
2771 | RETVAL = PTR2IV (self->ptr); |
3111 | RETVAL = PTR2IV (self->ptr); |
2772 | OUTPUT: |
3112 | OUTPUT: |
2773 | RETVAL |
3113 | RETVAL |
2774 | |
3114 | |
|
|
3115 | void |
|
|
3116 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3117 | CODE: |
|
|
3118 | STRLEN len; |
|
|
3119 | const char *ptr = SvPVbyte (data, len); |
|
|
3120 | |
|
|
3121 | if (offset + len > self->cb) |
|
|
3122 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3123 | |
|
|
3124 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3125 | |
2775 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
3126 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
2776 | |
3127 | |
2777 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
3128 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
2778 | |
3129 | |
2779 | IV |
3130 | IV |