|
|
1 | #include <stddef.h> |
|
|
2 | |
1 | #include "EXTERN.h" |
3 | #include "EXTERN.h" |
2 | #include "perl.h" |
4 | #include "perl.h" |
3 | #include "XSUB.h" |
5 | #include "XSUB.h" |
|
|
6 | |
|
|
7 | #define ECB_NO_THREADS 1 |
|
|
8 | #define ECB_NO_LIBM 1 |
|
|
9 | #include "ecb.h" |
4 | |
10 | |
5 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
11 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
6 | #include "xthread.h" |
12 | #include "xthread.h" |
7 | #include "schmorp.h" |
13 | #include "schmorp.h" |
8 | |
14 | |
… | |
… | |
26 | #if PREFER_1_1 |
32 | #if PREFER_1_1 |
27 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
33 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
28 | #endif |
34 | #endif |
29 | |
35 | |
30 | #ifdef __APPLE__ |
36 | #ifdef __APPLE__ |
31 | #include <OpenCL/opencl.h> |
37 | #define CLHDR(name) <OpenCL/name> |
32 | #else |
38 | #else |
|
|
39 | #define CLHDR(name) <CL/name> |
|
|
40 | #endif |
|
|
41 | |
33 | #include <CL/opencl.h> |
42 | #include CLHDR(opencl.h) |
|
|
43 | |
|
|
44 | #if 0 |
|
|
45 | #ifndef CL_VERSION_1_2 |
|
|
46 | #include CLHDR(cl_d3d9.h) |
|
|
47 | #endif |
|
|
48 | #endif |
|
|
49 | |
|
|
50 | #if _WIN32 |
|
|
51 | #include CLHDR(cl_d3d10.h) |
|
|
52 | #if CL_VERSION_1_2 |
|
|
53 | #include CLHDR(cl_d3d11.h) |
|
|
54 | #endif |
|
|
55 | #include CLHDR(cl_dx9_media_sharing.h) |
34 | #endif |
56 | #endif |
35 | |
57 | |
36 | #ifndef CL_VERSION_1_2 |
58 | #ifndef CL_VERSION_1_2 |
37 | #undef PREFER_1_1 |
59 | #undef PREFER_1_1 |
38 | #define PREFER_1_1 1 |
60 | #define PREFER_1_1 1 |
39 | #endif |
61 | #endif |
40 | |
62 | |
|
|
63 | // make sure all constants we might use are actually defined |
|
|
64 | #include "default.h" |
|
|
65 | |
41 | typedef cl_platform_id OpenCL__Platform; |
66 | typedef cl_platform_id OpenCL__Platform; |
42 | typedef cl_device_id OpenCL__Device; |
67 | typedef cl_device_id OpenCL__Device; |
|
|
68 | typedef cl_device_id OpenCL__SubDevice; |
43 | typedef cl_context OpenCL__Context; |
69 | typedef cl_context OpenCL__Context; |
44 | typedef cl_command_queue OpenCL__Queue; |
70 | typedef cl_command_queue OpenCL__Queue; |
45 | typedef cl_mem OpenCL__Memory; |
71 | typedef cl_mem OpenCL__Memory; |
46 | typedef cl_mem OpenCL__Buffer; |
72 | typedef cl_mem OpenCL__Buffer; |
47 | typedef cl_mem OpenCL__BufferObj; |
73 | typedef cl_mem OpenCL__BufferObj; |
48 | typedef cl_mem OpenCL__Image; |
74 | typedef cl_mem OpenCL__Image; |
49 | typedef cl_mem OpenCL__Image2D; |
|
|
50 | typedef cl_mem OpenCL__Image3D; |
|
|
51 | typedef cl_mem OpenCL__Memory_ornull; |
75 | typedef cl_mem OpenCL__Memory_ornull; |
52 | typedef cl_mem OpenCL__Buffer_ornull; |
76 | typedef cl_mem OpenCL__Buffer_ornull; |
53 | typedef cl_mem OpenCL__Image_ornull; |
77 | 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; |
78 | typedef cl_sampler OpenCL__Sampler; |
57 | typedef cl_program OpenCL__Program; |
79 | typedef cl_program OpenCL__Program; |
58 | typedef cl_kernel OpenCL__Kernel; |
80 | typedef cl_kernel OpenCL__Kernel; |
59 | typedef cl_event OpenCL__Event; |
81 | typedef cl_event OpenCL__Event; |
60 | typedef cl_event OpenCL__UserEvent; |
82 | typedef cl_event OpenCL__UserEvent; |
61 | |
83 | |
62 | typedef struct mapped * OpenCL__Mapped; |
84 | typedef struct mapped * OpenCL__Mapped; |
63 | |
85 | |
64 | typedef SV *FUTURE; |
|
|
65 | |
|
|
66 | static HV |
86 | static HV |
67 | *stash_platform, |
87 | *stash_platform, |
68 | *stash_device, |
88 | *stash_device, |
|
|
89 | *stash_subdevice, |
69 | *stash_context, |
90 | *stash_context, |
70 | *stash_queue, |
91 | *stash_queue, |
71 | *stash_program, |
92 | *stash_program, |
72 | *stash_kernel, |
93 | *stash_kernel, |
73 | *stash_sampler, |
94 | *stash_sampler, |
… | |
… | |
133 | } |
154 | } |
134 | |
155 | |
135 | /*****************************************************************************/ |
156 | /*****************************************************************************/ |
136 | |
157 | |
137 | /* up to two temporary buffers */ |
158 | /* up to two temporary buffers */ |
138 | static void * |
159 | static void * ecb_noinline |
139 | tmpbuf (size_t size) |
160 | tmpbuf (size_t size) |
140 | { |
161 | { |
141 | enum { buffers = 3 }; |
162 | enum { buffers = 4 }; |
142 | static int idx; |
163 | static int idx; |
143 | static void *buf [buffers]; |
164 | static void *buf [buffers]; |
144 | static size_t len [buffers]; |
165 | static size_t len [buffers]; |
145 | |
166 | |
146 | idx = (idx + 1) % buffers; |
167 | idx = (idx + 1) % buffers; |
… | |
… | |
153 | } |
174 | } |
154 | |
175 | |
155 | return buf [idx]; |
176 | return buf [idx]; |
156 | } |
177 | } |
157 | |
178 | |
|
|
179 | static const char * ecb_noinline |
|
|
180 | cv_get_name (CV *cv) |
|
|
181 | { |
|
|
182 | static char fullname [256]; |
|
|
183 | |
|
|
184 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
185 | |
|
|
186 | HV *stash = GvSTASH (gv); |
|
|
187 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
188 | const char *gvname = GvNAME (gv); |
|
|
189 | |
|
|
190 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
191 | return fullname; |
|
|
192 | } |
|
|
193 | |
158 | /*****************************************************************************/ |
194 | /*****************************************************************************/ |
159 | |
195 | |
160 | typedef struct |
196 | typedef struct |
161 | { |
197 | { |
162 | IV iv; |
198 | IV iv; |
163 | const char *name; |
199 | const char *name; |
164 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
200 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
165 | } ivstr; |
201 | } ivstr; |
|
|
202 | |
|
|
203 | typedef struct |
|
|
204 | { |
|
|
205 | NV nv; |
|
|
206 | const char *name; |
|
|
207 | #define const_nv(name) { (NV)CL_ ## name, # name }, |
|
|
208 | } nvstr; |
166 | |
209 | |
167 | static const char * |
210 | static const char * |
168 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
211 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
169 | { |
212 | { |
170 | int i; |
213 | int i; |
… | |
… | |
219 | if (res) \ |
262 | if (res) \ |
220 | FAIL (name); |
263 | FAIL (name); |
221 | |
264 | |
222 | /*****************************************************************************/ |
265 | /*****************************************************************************/ |
223 | |
266 | |
|
|
267 | static SV * |
|
|
268 | new_clobj (HV *stash, IV id) |
|
|
269 | { |
|
|
270 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
271 | } |
|
|
272 | |
|
|
273 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
274 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
275 | |
|
|
276 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
277 | /* they can be upgraded at runtime to the array form */ |
|
|
278 | static void * ecb_noinline |
|
|
279 | SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg) |
|
|
280 | { |
|
|
281 | // sv_derived_from is quite slow :( |
|
|
282 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
283 | return (void *)SvIV (SvRV (sv)); |
|
|
284 | |
|
|
285 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg); |
|
|
286 | } |
|
|
287 | |
|
|
288 | // the "no-inherit" version of the above |
|
|
289 | static void * ecb_noinline |
|
|
290 | SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash) |
|
|
291 | { |
|
|
292 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
293 | return (void *)SvIV (SvRV (sv)); |
|
|
294 | |
|
|
295 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash)); |
|
|
296 | } |
|
|
297 | |
|
|
298 | /*****************************************************************************/ |
|
|
299 | |
224 | static cl_context_properties * |
300 | static cl_context_properties * ecb_noinline |
225 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
301 | SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
226 | { |
302 | { |
227 | if (!sv || !SvOK (sv)) |
303 | if (!sv || !SvOK (sv)) |
228 | if (extra) |
304 | if (extra) |
229 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
305 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
230 | else |
306 | else |
… | |
… | |
236 | int i, len = av_len (av) + 1; |
312 | int i, len = av_len (av) + 1; |
237 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
313 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
238 | cl_context_properties *l = p; |
314 | cl_context_properties *l = p; |
239 | |
315 | |
240 | if (len & 1) |
316 | if (len & 1) |
241 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
317 | croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname); |
242 | |
318 | |
243 | while (extracount--) |
319 | while (extracount--) |
244 | *l++ = *extra++; |
320 | *l++ = *extra++; |
245 | |
321 | |
246 | for (i = 0; i < len; i += 2) |
322 | for (i = 0; i < len; i += 2) |
… | |
… | |
249 | SV *p_sv = *av_fetch (av, i + 1, 0); |
325 | SV *p_sv = *av_fetch (av, i + 1, 0); |
250 | cl_context_properties v = SvIV (p_sv); // code below can override |
326 | cl_context_properties v = SvIV (p_sv); // code below can override |
251 | |
327 | |
252 | switch (t) |
328 | switch (t) |
253 | { |
329 | { |
|
|
330 | case CL_CONTEXT_PLATFORM: |
|
|
331 | if (SvROK (p_sv)) |
|
|
332 | v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform"); |
|
|
333 | break; |
|
|
334 | |
254 | case CL_GLX_DISPLAY_KHR: |
335 | case CL_GLX_DISPLAY_KHR: |
255 | if (!SvOK (p_sv)) |
336 | if (!SvOK (p_sv)) |
256 | { |
337 | { |
257 | void *func = glsym ("_glXGetCurrentDisplay"); |
338 | void *func = glsym ("_glXGetCurrentDisplay"); |
258 | if (func) |
339 | if (func) |
… | |
… | |
281 | *l = 0; |
362 | *l = 0; |
282 | |
363 | |
283 | return p; |
364 | return p; |
284 | } |
365 | } |
285 | |
366 | |
286 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
367 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname); |
287 | } |
368 | } |
288 | |
369 | |
289 | /*****************************************************************************/ |
370 | // parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation |
290 | |
371 | // is a pointer (and only on well-behaved systems). |
291 | static SV * |
372 | static void * ecb_noinline |
292 | new_clobj (HV *stash, IV id) |
373 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
293 | { |
374 | { |
294 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
375 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
295 | } |
376 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
377 | cv_get_name (cv), argname, or_undef ? "undef or " : "", klass); |
296 | |
378 | |
297 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
379 | AV *av = (AV *)SvRV (arg); |
298 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
380 | void **list = 0; |
|
|
381 | cl_uint count = av_len (av) + 1; |
299 | |
382 | |
300 | /* cl objects are either \$iv, or [$iv, ...] */ |
383 | if (count) |
301 | /* they can be upgraded at runtime to the array form */ |
384 | { |
302 | static void * |
385 | list = tmpbuf (sizeof (*list) * count); |
303 | SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
386 | int i; |
304 | { |
387 | for (i = 0; i < count; ++i) |
305 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
388 | list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass); |
306 | return (void *)SvIV (SvRV (sv)); |
389 | } |
307 | |
390 | |
308 | croak ("%s: %s is not of type %s", func, svname, pkg); |
391 | if (!count && !or_undef) |
|
|
392 | croak ("%s: '%s' must contain at least one %s object", |
|
|
393 | cv_get_name (cv), argname, klass); |
|
|
394 | |
|
|
395 | *rcount = count; |
|
|
396 | return (void *)list; |
309 | } |
397 | } |
310 | |
398 | |
311 | /*****************************************************************************/ |
399 | /*****************************************************************************/ |
312 | /* callback stuff */ |
400 | /* callback stuff */ |
313 | |
401 | |
… | |
… | |
335 | static void (*eq_signal_func)(void *signal_arg, int value); |
423 | static void (*eq_signal_func)(void *signal_arg, int value); |
336 | static void *eq_signal_arg; |
424 | static void *eq_signal_arg; |
337 | static xmutex_t eq_lock = X_MUTEX_INIT; |
425 | static xmutex_t eq_lock = X_MUTEX_INIT; |
338 | static eq_item *eq_head, *eq_tail; |
426 | static eq_item *eq_head, *eq_tail; |
339 | |
427 | |
340 | static void |
428 | static void ecb_noinline |
341 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
429 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
342 | { |
430 | { |
343 | eq_item *item = malloc (sizeof (eq_item)); |
431 | eq_item *item = malloc (sizeof (eq_item)); |
344 | |
432 | |
345 | item->next = 0; |
433 | item->next = 0; |
… | |
… | |
412 | } |
500 | } |
413 | |
501 | |
414 | /*****************************************************************************/ |
502 | /*****************************************************************************/ |
415 | /* context notify */ |
503 | /* context notify */ |
416 | |
504 | |
417 | static void |
505 | static void ecb_noinline |
418 | eq_context_push (void *data1, void *data2, void *data3) |
506 | eq_context_push (void *data1, void *data2, void *data3) |
419 | { |
507 | { |
420 | dSP; |
508 | dSP; |
421 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
509 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
422 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
510 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
… | |
… | |
444 | { \ |
532 | { \ |
445 | pfn_notify = eq_context_notify; \ |
533 | pfn_notify = eq_context_notify; \ |
446 | user_data = s_get_cv (notify); \ |
534 | user_data = s_get_cv (notify); \ |
447 | } |
535 | } |
448 | |
536 | |
449 | static SV * |
537 | static SV * ecb_noinline |
450 | new_clobj_context (cl_context ctx, void *user_data) |
538 | new_clobj_context (cl_context ctx, void *user_data) |
451 | { |
539 | { |
452 | SV *sv = new_clobj (stash_context, (IV)ctx); |
540 | SV *sv = new_clobj (stash_context, (IV)ctx); |
453 | |
541 | |
454 | if (user_data) |
542 | if (user_data) |
… | |
… | |
473 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
561 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
474 | |
562 | |
475 | static void CL_CALLBACK |
563 | static void CL_CALLBACK |
476 | eq_program_notify (cl_program program, void *user_data) |
564 | eq_program_notify (cl_program program, void *user_data) |
477 | { |
565 | { |
|
|
566 | clRetainProgram (program); |
|
|
567 | |
478 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
568 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
569 | } |
|
|
570 | |
|
|
571 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
572 | |
|
|
573 | static program_callback ecb_noinline |
|
|
574 | make_program_callback (SV *notify, void **ruser_data) |
|
|
575 | { |
|
|
576 | if (SvOK (notify)) |
|
|
577 | { |
|
|
578 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
579 | return eq_program_notify; |
|
|
580 | } |
|
|
581 | else |
|
|
582 | { |
|
|
583 | *ruser_data = 0; |
|
|
584 | return 0; |
|
|
585 | } |
479 | } |
586 | } |
480 | |
587 | |
481 | struct build_args |
588 | struct build_args |
482 | { |
589 | { |
483 | cl_program program; |
590 | cl_program program; |
… | |
… | |
497 | else |
604 | else |
498 | clReleaseProgram (arg->program); |
605 | clReleaseProgram (arg->program); |
499 | |
606 | |
500 | free (arg->options); |
607 | free (arg->options); |
501 | free (arg); |
608 | free (arg); |
|
|
609 | |
|
|
610 | return 0; |
502 | } |
611 | } |
503 | |
612 | |
504 | static void |
613 | 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) |
614 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
506 | { |
615 | { |
… | |
… | |
511 | arg->user_data = user_data; |
620 | arg->user_data = user_data; |
512 | arg->num_devices = num_devices; |
621 | arg->num_devices = num_devices; |
513 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
622 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
514 | |
623 | |
515 | xthread_t id; |
624 | xthread_t id; |
516 | thread_create (&id, build_program_thread, arg); |
625 | xthread_create (&id, build_program_thread, arg); |
|
|
626 | } |
|
|
627 | |
|
|
628 | /*****************************************************************************/ |
|
|
629 | /* mem object destructor notify */ |
|
|
630 | |
|
|
631 | static void ecb_noinline |
|
|
632 | eq_destructor_push (void *data1, void *data2, void *data3) |
|
|
633 | { |
|
|
634 | } |
|
|
635 | |
|
|
636 | static eq_vtbl eq_destructor_vtbl = { 0, eq_destructor_push }; |
|
|
637 | |
|
|
638 | static void CL_CALLBACK |
|
|
639 | eq_destructor_notify (cl_mem memobj, void *user_data) |
|
|
640 | { |
|
|
641 | eq_enq (&eq_destructor_vtbl, (SV *)user_data, (void *)memobj, 0, 0); |
517 | } |
642 | } |
518 | |
643 | |
519 | /*****************************************************************************/ |
644 | /*****************************************************************************/ |
520 | /* event objects */ |
645 | /* event objects */ |
521 | |
646 | |
… | |
… | |
536 | clRetainEvent (event); |
661 | clRetainEvent (event); |
537 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
662 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
538 | } |
663 | } |
539 | |
664 | |
540 | /*****************************************************************************/ |
665 | /*****************************************************************************/ |
541 | /* mapped_xxx */ |
666 | /* utilities for XS code */ |
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 | |
667 | |
616 | static size_t |
668 | static size_t |
617 | img_row_pitch (cl_mem img) |
669 | img_row_pitch (cl_mem img) |
618 | { |
670 | { |
619 | size_t res; |
671 | size_t res; |
620 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
672 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
621 | return res; |
673 | return res; |
622 | } |
674 | } |
623 | |
675 | |
624 | static cl_event * |
676 | static cl_event * ecb_noinline |
625 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
677 | event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra) |
626 | { |
678 | { |
627 | cl_uint count = *rcount; |
679 | cl_uint count = *rcount; |
|
|
680 | |
|
|
681 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
682 | *rcount = count = 0; |
628 | |
683 | |
629 | if (!count && !extra) |
684 | if (!count && !extra) |
630 | return 0; |
685 | return 0; |
631 | |
686 | |
632 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
687 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
633 | int i = 0; |
688 | int i = 0; |
634 | |
689 | |
635 | while (count) |
690 | while (count--) |
636 | { |
|
|
637 | --count; |
|
|
638 | if (SvOK (items [count])) |
691 | if (SvOK (items [count])) |
639 | list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
692 | list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event"); |
640 | } |
|
|
641 | |
693 | |
642 | if (extra) |
694 | if (extra) |
643 | list [i++] = extra; |
695 | list [i++] = extra; |
644 | |
696 | |
645 | *rcount = i; |
697 | *rcount = i; |
646 | |
698 | |
647 | return i ? list : 0; |
699 | return i ? list : 0; |
648 | } |
700 | } |
649 | |
701 | |
650 | #define EVENT_LIST(skip) \ |
702 | #define EVENT_LIST(skip) \ |
651 | cl_uint event_list_count = items - skip; \ |
703 | cl_uint event_list_count = items - (skip); \ |
652 | cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) |
704 | cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0) |
653 | |
705 | |
654 | #define INFO(class) \ |
706 | #define INFO(class) \ |
655 | { \ |
707 | { \ |
656 | size_t size; \ |
708 | size_t size; \ |
657 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
709 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
661 | SvCUR_set (sv, size); \ |
713 | SvCUR_set (sv, size); \ |
662 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
714 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
663 | XPUSHs (sv); \ |
715 | XPUSHs (sv); \ |
664 | } |
716 | } |
665 | |
717 | |
|
|
718 | /*****************************************************************************/ |
|
|
719 | /* mapped_xxx */ |
|
|
720 | |
|
|
721 | static OpenCL__Mapped |
|
|
722 | SvMAPPED (SV *sv) |
|
|
723 | { |
|
|
724 | // no typechecking atm., keep your fingers crossed |
|
|
725 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
726 | } |
|
|
727 | |
|
|
728 | struct mapped |
|
|
729 | { |
|
|
730 | cl_command_queue queue; |
|
|
731 | cl_mem memobj; |
|
|
732 | void *ptr; |
|
|
733 | size_t cb; |
|
|
734 | cl_event event; |
|
|
735 | size_t row_pitch; |
|
|
736 | size_t slice_pitch; |
|
|
737 | |
|
|
738 | size_t element_size; |
|
|
739 | size_t width, height, depth; |
|
|
740 | }; |
|
|
741 | |
|
|
742 | static SV * |
|
|
743 | mapped_new ( |
|
|
744 | HV *stash, cl_command_queue queue, cl_mem memobj, cl_map_flags flags, |
|
|
745 | void *ptr, size_t cb, cl_event ev, |
|
|
746 | size_t row_pitch, size_t slice_pitch, size_t element_size, |
|
|
747 | size_t width, size_t height, size_t depth |
|
|
748 | ) |
|
|
749 | { |
|
|
750 | SV *data = newSV (0); |
|
|
751 | SvUPGRADE (data, SVt_PVMG); |
|
|
752 | |
|
|
753 | OpenCL__Mapped mapped; |
|
|
754 | New (0, mapped, 1, struct mapped); |
|
|
755 | |
|
|
756 | clRetainCommandQueue (queue); |
|
|
757 | |
|
|
758 | mapped->queue = queue; |
|
|
759 | mapped->memobj = memobj; |
|
|
760 | mapped->ptr = ptr; |
|
|
761 | mapped->cb = cb; |
|
|
762 | mapped->event = ev; |
|
|
763 | mapped->row_pitch = row_pitch; |
|
|
764 | mapped->slice_pitch = slice_pitch; |
|
|
765 | |
|
|
766 | mapped->element_size = element_size; |
|
|
767 | mapped->width = width; |
|
|
768 | mapped->height = height; |
|
|
769 | mapped->depth = depth; |
|
|
770 | |
|
|
771 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
772 | |
|
|
773 | if (SvLEN (data)) |
|
|
774 | Safefree (data); |
|
|
775 | |
|
|
776 | SvPVX (data) = (char *)ptr; |
|
|
777 | SvCUR_set (data, cb); |
|
|
778 | SvLEN_set (data, 0); |
|
|
779 | SvPOK_only (data); |
|
|
780 | |
|
|
781 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
782 | |
|
|
783 | if (!(flags & CL_MAP_WRITE)) |
|
|
784 | SvREADONLY_on (data); |
|
|
785 | |
|
|
786 | return obj; |
|
|
787 | } |
|
|
788 | |
|
|
789 | static void |
|
|
790 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
791 | { |
|
|
792 | SV *data = SvRV (sv); |
|
|
793 | |
|
|
794 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
795 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
796 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
797 | else |
|
|
798 | { |
|
|
799 | SvREADONLY_off (data); |
|
|
800 | SvCUR_set (data, 0); |
|
|
801 | SvPVX (data) = 0; |
|
|
802 | SvOK_off (data); |
|
|
803 | } |
|
|
804 | |
|
|
805 | mapped->ptr = 0; |
|
|
806 | } |
|
|
807 | |
|
|
808 | static void |
|
|
809 | mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
810 | { |
|
|
811 | cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event); |
|
|
812 | cl_event ev; |
|
|
813 | |
|
|
814 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
815 | |
|
|
816 | clReleaseEvent (mapped->event); |
|
|
817 | mapped->event = ev; |
|
|
818 | |
|
|
819 | mapped_detach (self, mapped); |
|
|
820 | } |
|
|
821 | |
|
|
822 | static size_t |
|
|
823 | mapped_element_size (OpenCL__Mapped self) |
|
|
824 | { |
|
|
825 | if (!self->element_size) |
|
|
826 | clGetImageInfo (self->memobj, CL_IMAGE_ELEMENT_SIZE, sizeof (self->element_size), &self->element_size, 0); |
|
|
827 | |
|
|
828 | return self->element_size; |
|
|
829 | } |
|
|
830 | |
|
|
831 | /*****************************************************************************/ |
|
|
832 | |
666 | MODULE = OpenCL PACKAGE = OpenCL |
833 | MODULE = OpenCL PACKAGE = OpenCL |
667 | |
834 | |
668 | PROTOTYPES: ENABLE |
835 | PROTOTYPES: ENABLE |
669 | |
836 | |
670 | void |
837 | void |
… | |
… | |
679 | eq_signal_arg = (void*)arg; |
846 | eq_signal_arg = (void*)arg; |
680 | |
847 | |
681 | BOOT: |
848 | BOOT: |
682 | { |
849 | { |
683 | HV *stash = gv_stashpv ("OpenCL", 1); |
850 | HV *stash = gv_stashpv ("OpenCL", 1); |
|
|
851 | |
684 | static const ivstr *civ, const_iv[] = { |
852 | static const ivstr *civ, const_iv[] = { |
685 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
853 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
686 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
854 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
687 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
855 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
688 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
856 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
… | |
… | |
691 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
859 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
692 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
860 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
693 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
861 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
694 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
862 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
695 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
863 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
|
|
864 | { PREFER_1_1 , "PREFER_1_1" }, |
696 | #include "constiv.h" |
865 | #include "constiv.h" |
697 | }; |
866 | }; |
698 | |
867 | |
699 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
868 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
700 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
869 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
701 | |
870 | |
|
|
871 | static const nvstr *cnv, const_nv[] = { |
|
|
872 | #include "constnv.h" |
|
|
873 | }; |
|
|
874 | |
|
|
875 | for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--) |
|
|
876 | newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv)); |
|
|
877 | |
|
|
878 | newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call |
|
|
879 | |
702 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
880 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
703 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
881 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
882 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
704 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
883 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
705 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
884 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
706 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
885 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
707 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
886 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
708 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
887 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
… | |
… | |
749 | list = tmpbuf (sizeof (*list) * count); |
928 | list = tmpbuf (sizeof (*list) * count); |
750 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
929 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
751 | |
930 | |
752 | EXTEND (SP, count); |
931 | EXTEND (SP, count); |
753 | for (i = 0; i < count; ++i) |
932 | for (i = 0; i < count; ++i) |
754 | PUSH_CLOBJ (stash_platform, (IV)list [i]); |
933 | PUSH_CLOBJ (stash_platform, list [i]); |
755 | |
934 | |
756 | void |
935 | void |
757 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
936 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
758 | PPCODE: |
937 | PPCODE: |
759 | CONTEXT_NOTIFY_CALLBACK; |
938 | CONTEXT_NOTIFY_CALLBACK; |
760 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
939 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
761 | XPUSH_CLOBJ_CONTEXT; |
940 | XPUSH_CLOBJ_CONTEXT; |
762 | |
941 | |
763 | void |
942 | void |
764 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
943 | context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
765 | PPCODE: |
944 | PPCODE: |
766 | /* der Gipfel der Kunst */ |
945 | cl_uint device_count; |
|
|
946 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
947 | |
|
|
948 | CONTEXT_NOTIFY_CALLBACK; |
|
|
949 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
950 | XPUSH_CLOBJ_CONTEXT; |
767 | |
951 | |
768 | void |
952 | void |
769 | wait_for_events (...) |
953 | wait_for_events (...) |
770 | CODE: |
954 | CODE: |
771 | EVENT_LIST (0); |
955 | EVENT_LIST (0); |
… | |
… | |
822 | EXTEND (SP, count); |
1006 | EXTEND (SP, count); |
823 | for (i = 0; i < count; ++i) |
1007 | for (i = 0; i < count; ++i) |
824 | PUSH_CLOBJ (stash_device, list [i]); |
1008 | PUSH_CLOBJ (stash_device, list [i]); |
825 | |
1009 | |
826 | void |
1010 | void |
827 | context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
1011 | context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef) |
828 | PPCODE: |
1012 | PPCODE: |
|
|
1013 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
1014 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
1015 | |
|
|
1016 | cl_uint device_count; |
|
|
1017 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1018 | |
|
|
1019 | CONTEXT_NOTIFY_CALLBACK; |
|
|
1020 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
1021 | XPUSH_CLOBJ_CONTEXT; |
|
|
1022 | |
|
|
1023 | void |
|
|
1024 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
1025 | PPCODE: |
|
|
1026 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
1027 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
1028 | |
|
|
1029 | CONTEXT_NOTIFY_CALLBACK; |
|
|
1030 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
1031 | XPUSH_CLOBJ_CONTEXT; |
|
|
1032 | |
|
|
1033 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
1034 | |
|
|
1035 | void |
|
|
1036 | info (OpenCL::Device self, cl_device_info name) |
|
|
1037 | PPCODE: |
|
|
1038 | INFO (Device) |
|
|
1039 | |
|
|
1040 | #if CL_VERSION_1_2 |
|
|
1041 | |
|
|
1042 | void |
|
|
1043 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
1044 | PPCODE: |
829 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
1045 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
830 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
1046 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
831 | |
1047 | |
832 | AV *av = (AV *)SvRV (devices); |
1048 | properties = SvRV (properties); |
833 | cl_uint num_devices = av_len (av) + 1; |
1049 | |
834 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
1050 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
1051 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
835 | |
1052 | |
836 | int i; |
1053 | int i; |
837 | for (i = num_devices; i--; ) |
1054 | for (i = 0; i < count; ++i) |
838 | device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
1055 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
839 | |
1056 | |
840 | CONTEXT_NOTIFY_CALLBACK; |
1057 | props [count] = 0; |
841 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res)); |
|
|
842 | XPUSH_CLOBJ_CONTEXT; |
|
|
843 | |
1058 | |
844 | void |
1059 | cl_uint num_devices; |
845 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
1060 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
846 | PPCODE: |
1061 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
847 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
1062 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
848 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
849 | |
1063 | |
850 | CONTEXT_NOTIFY_CALLBACK; |
1064 | EXTEND (SP, num_devices); |
851 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
1065 | for (i = 0; i < count; ++i) |
852 | XPUSH_CLOBJ_CONTEXT; |
1066 | PUSH_CLOBJ (stash_subdevice, list [i]); |
853 | |
1067 | |
854 | MODULE = OpenCL PACKAGE = OpenCL::Device |
1068 | #endif |
855 | |
|
|
856 | void |
|
|
857 | info (OpenCL::Device self, cl_device_info name) |
|
|
858 | PPCODE: |
|
|
859 | INFO (Device) |
|
|
860 | |
1069 | |
861 | #BEGIN:device |
1070 | #BEGIN:device |
862 | |
1071 | |
863 | void |
1072 | void |
864 | type (OpenCL::Device self) |
1073 | type (OpenCL::Device self) |
|
|
1074 | ALIAS: |
|
|
1075 | type = CL_DEVICE_TYPE |
|
|
1076 | address_bits = CL_DEVICE_ADDRESS_BITS |
|
|
1077 | max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE |
|
|
1078 | single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG |
|
|
1079 | global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE |
|
|
1080 | global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE |
|
|
1081 | max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE |
|
|
1082 | local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE |
|
|
1083 | execution_capabilities = CL_DEVICE_EXECUTION_CAPABILITIES |
|
|
1084 | properties = CL_DEVICE_QUEUE_PROPERTIES |
|
|
1085 | double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG |
|
|
1086 | half_fp_config = CL_DEVICE_HALF_FP_CONFIG |
865 | PPCODE: |
1087 | PPCODE: |
866 | cl_device_type value [1]; |
1088 | cl_ulong value [1]; |
867 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
1089 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
868 | EXTEND (SP, 1); |
1090 | EXTEND (SP, 1); |
869 | const int i = 0; |
1091 | const int i = 0; |
870 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1092 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
871 | |
1093 | |
872 | void |
1094 | void |
… | |
… | |
886 | max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS |
1108 | max_write_image_args = CL_DEVICE_MAX_WRITE_IMAGE_ARGS |
887 | image_support = CL_DEVICE_IMAGE_SUPPORT |
1109 | image_support = CL_DEVICE_IMAGE_SUPPORT |
888 | max_samplers = CL_DEVICE_MAX_SAMPLERS |
1110 | max_samplers = CL_DEVICE_MAX_SAMPLERS |
889 | mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN |
1111 | mem_base_addr_align = CL_DEVICE_MEM_BASE_ADDR_ALIGN |
890 | min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE |
1112 | min_data_type_align_size = CL_DEVICE_MIN_DATA_TYPE_ALIGN_SIZE |
|
|
1113 | global_mem_cache_type = CL_DEVICE_GLOBAL_MEM_CACHE_TYPE |
891 | global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
1114 | global_mem_cacheline_size = CL_DEVICE_GLOBAL_MEM_CACHELINE_SIZE |
892 | max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS |
1115 | max_constant_args = CL_DEVICE_MAX_CONSTANT_ARGS |
|
|
1116 | local_mem_type = CL_DEVICE_LOCAL_MEM_TYPE |
893 | preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF |
1117 | preferred_vector_width_half = CL_DEVICE_PREFERRED_VECTOR_WIDTH_HALF |
894 | native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR |
1118 | native_vector_width_char = CL_DEVICE_NATIVE_VECTOR_WIDTH_CHAR |
895 | native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT |
1119 | native_vector_width_short = CL_DEVICE_NATIVE_VECTOR_WIDTH_SHORT |
896 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
1120 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
897 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
1121 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
… | |
… | |
935 | EXTEND (SP, n); |
1159 | EXTEND (SP, n); |
936 | for (i = 0; i < n; ++i) |
1160 | for (i = 0; i < n; ++i) |
937 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1161 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
938 | |
1162 | |
939 | void |
1163 | void |
940 | address_bits (OpenCL::Device self) |
|
|
941 | PPCODE: |
|
|
942 | cl_bitfield value [1]; |
|
|
943 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_ADDRESS_BITS, sizeof (value), value, 0)); |
|
|
944 | EXTEND (SP, 1); |
|
|
945 | const int i = 0; |
|
|
946 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
947 | |
|
|
948 | void |
|
|
949 | max_mem_alloc_size (OpenCL::Device self) |
|
|
950 | ALIAS: |
|
|
951 | max_mem_alloc_size = CL_DEVICE_MAX_MEM_ALLOC_SIZE |
|
|
952 | global_mem_cache_size = CL_DEVICE_GLOBAL_MEM_CACHE_SIZE |
|
|
953 | global_mem_size = CL_DEVICE_GLOBAL_MEM_SIZE |
|
|
954 | max_constant_buffer_size = CL_DEVICE_MAX_CONSTANT_BUFFER_SIZE |
|
|
955 | local_mem_size = CL_DEVICE_LOCAL_MEM_SIZE |
|
|
956 | PPCODE: |
|
|
957 | cl_ulong value [1]; |
|
|
958 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
959 | EXTEND (SP, 1); |
|
|
960 | const int i = 0; |
|
|
961 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
962 | |
|
|
963 | void |
|
|
964 | single_fp_config (OpenCL::Device self) |
|
|
965 | ALIAS: |
|
|
966 | single_fp_config = CL_DEVICE_SINGLE_FP_CONFIG |
|
|
967 | double_fp_config = CL_DEVICE_DOUBLE_FP_CONFIG |
|
|
968 | half_fp_config = CL_DEVICE_HALF_FP_CONFIG |
|
|
969 | PPCODE: |
|
|
970 | cl_device_fp_config value [1]; |
|
|
971 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
|
|
972 | EXTEND (SP, 1); |
|
|
973 | const int i = 0; |
|
|
974 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
975 | |
|
|
976 | void |
|
|
977 | global_mem_cache_type (OpenCL::Device self) |
|
|
978 | PPCODE: |
|
|
979 | cl_device_mem_cache_type value [1]; |
|
|
980 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_GLOBAL_MEM_CACHE_TYPE, sizeof (value), value, 0)); |
|
|
981 | EXTEND (SP, 1); |
|
|
982 | const int i = 0; |
|
|
983 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
984 | |
|
|
985 | void |
|
|
986 | local_mem_type (OpenCL::Device self) |
|
|
987 | PPCODE: |
|
|
988 | cl_device_local_mem_type value [1]; |
|
|
989 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_LOCAL_MEM_TYPE, sizeof (value), value, 0)); |
|
|
990 | EXTEND (SP, 1); |
|
|
991 | const int i = 0; |
|
|
992 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
993 | |
|
|
994 | void |
|
|
995 | error_correction_support (OpenCL::Device self) |
1164 | error_correction_support (OpenCL::Device self) |
996 | ALIAS: |
1165 | ALIAS: |
997 | error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT |
1166 | error_correction_support = CL_DEVICE_ERROR_CORRECTION_SUPPORT |
998 | endian_little = CL_DEVICE_ENDIAN_LITTLE |
1167 | endian_little = CL_DEVICE_ENDIAN_LITTLE |
999 | available = CL_DEVICE_AVAILABLE |
1168 | available = CL_DEVICE_AVAILABLE |
1000 | compiler_available = CL_DEVICE_COMPILER_AVAILABLE |
1169 | compiler_available = CL_DEVICE_COMPILER_AVAILABLE |
1001 | host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY |
1170 | host_unified_memory = CL_DEVICE_HOST_UNIFIED_MEMORY |
1002 | PPCODE: |
1171 | PPCODE: |
1003 | cl_bool value [1]; |
1172 | cl_uint value [1]; |
1004 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
1173 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
1005 | EXTEND (SP, 1); |
1174 | EXTEND (SP, 1); |
1006 | const int i = 0; |
1175 | const int i = 0; |
1007 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
1176 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
1008 | |
|
|
1009 | void |
|
|
1010 | execution_capabilities (OpenCL::Device self) |
|
|
1011 | PPCODE: |
|
|
1012 | cl_device_exec_capabilities value [1]; |
|
|
1013 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_EXECUTION_CAPABILITIES, sizeof (value), value, 0)); |
|
|
1014 | EXTEND (SP, 1); |
|
|
1015 | const int i = 0; |
|
|
1016 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1017 | |
|
|
1018 | void |
|
|
1019 | properties (OpenCL::Device self) |
|
|
1020 | PPCODE: |
|
|
1021 | cl_command_queue_properties value [1]; |
|
|
1022 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
|
|
1023 | EXTEND (SP, 1); |
|
|
1024 | const int i = 0; |
|
|
1025 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1026 | |
1177 | |
1027 | void |
1178 | void |
1028 | platform (OpenCL::Device self) |
1179 | platform (OpenCL::Device self) |
1029 | PPCODE: |
1180 | PPCODE: |
1030 | cl_platform_id value [1]; |
1181 | cl_platform_id value [1]; |
1031 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1182 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1032 | EXTEND (SP, 1); |
1183 | EXTEND (SP, 1); |
1033 | const int i = 0; |
1184 | const int i = 0; |
1034 | { |
|
|
1035 | PUSH_CLOBJ (stash_platform, value [i]); |
1185 | PUSH_CLOBJ (stash_platform, value [i]); |
1036 | } |
|
|
1037 | |
1186 | |
1038 | void |
1187 | void |
1039 | name (OpenCL::Device self) |
1188 | name (OpenCL::Device self) |
1040 | ALIAS: |
1189 | ALIAS: |
1041 | name = CL_DEVICE_NAME |
1190 | name = CL_DEVICE_NAME |
… | |
… | |
1058 | PPCODE: |
1207 | PPCODE: |
1059 | cl_device_id value [1]; |
1208 | cl_device_id value [1]; |
1060 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1209 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1061 | EXTEND (SP, 1); |
1210 | EXTEND (SP, 1); |
1062 | const int i = 0; |
1211 | const int i = 0; |
1063 | { |
|
|
1064 | PUSH_CLOBJ (stash_device, value [i]); |
1212 | PUSH_CLOBJ (stash_device, value [i]); |
1065 | } |
|
|
1066 | |
1213 | |
1067 | void |
1214 | void |
1068 | partition_types_ext (OpenCL::Device self) |
1215 | partition_types_ext (OpenCL::Device self) |
1069 | ALIAS: |
1216 | ALIAS: |
1070 | partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT |
1217 | partition_types_ext = CL_DEVICE_PARTITION_TYPES_EXT |
… | |
… | |
1080 | for (i = 0; i < n; ++i) |
1227 | for (i = 0; i < n; ++i) |
1081 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1228 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1082 | |
1229 | |
1083 | #END:device |
1230 | #END:device |
1084 | |
1231 | |
|
|
1232 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1233 | |
|
|
1234 | #if CL_VERSION_1_2 |
|
|
1235 | |
|
|
1236 | void |
|
|
1237 | DESTROY (OpenCL::SubDevice self) |
|
|
1238 | CODE: |
|
|
1239 | clReleaseDevice (self); |
|
|
1240 | |
|
|
1241 | #endif |
|
|
1242 | |
1085 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1243 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1086 | |
1244 | |
1087 | void |
1245 | void |
1088 | DESTROY (OpenCL::Context context) |
1246 | DESTROY (OpenCL::Context self) |
1089 | CODE: |
1247 | CODE: |
1090 | clReleaseContext (context); |
1248 | clReleaseContext (self); |
1091 | |
1249 | |
1092 | void |
1250 | void |
1093 | info (OpenCL::Context self, cl_context_info name) |
1251 | info (OpenCL::Context self, cl_context_info name) |
1094 | PPCODE: |
1252 | PPCODE: |
1095 | INFO (Context) |
1253 | INFO (Context) |
… | |
… | |
1136 | const cl_image_desc desc = { |
1294 | const cl_image_desc desc = { |
1137 | type, |
1295 | type, |
1138 | width, height, depth, |
1296 | width, height, depth, |
1139 | array_size, row_pitch, slice_pitch, |
1297 | array_size, row_pitch, slice_pitch, |
1140 | num_mip_level, num_samples, |
1298 | num_mip_level, num_samples, |
1141 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
1299 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0 |
1142 | }; |
1300 | }; |
1143 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
1301 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
1144 | HV *stash = stash_image; |
1302 | HV *stash = stash_image; |
1145 | switch (type) |
1303 | switch (type) |
1146 | { |
1304 | { |
… | |
… | |
1279 | |
1437 | |
1280 | len2 = len; |
1438 | len2 = len; |
1281 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1439 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1282 | XPUSH_CLOBJ (stash_program, prog); |
1440 | XPUSH_CLOBJ (stash_program, prog); |
1283 | |
1441 | |
|
|
1442 | void |
|
|
1443 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1444 | PPCODE: |
|
|
1445 | cl_uint device_count; |
|
|
1446 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1447 | |
|
|
1448 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1449 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1450 | |
|
|
1451 | binaries = SvRV (binaries); |
|
|
1452 | |
|
|
1453 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1454 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1455 | |
|
|
1456 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1457 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1458 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1459 | |
|
|
1460 | int i; |
|
|
1461 | for (i = 0; i < device_count; ++i) |
|
|
1462 | { |
|
|
1463 | STRLEN len; |
|
|
1464 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1465 | length_list [i] = len; |
|
|
1466 | } |
|
|
1467 | |
|
|
1468 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1469 | length_list, binary_list, |
|
|
1470 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1471 | |
|
|
1472 | EXTEND (SP, 2); |
|
|
1473 | PUSH_CLOBJ (stash_program, prog); |
|
|
1474 | |
|
|
1475 | if (GIMME_V == G_ARRAY) |
|
|
1476 | { |
|
|
1477 | AV *av = newAV (); |
|
|
1478 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1479 | |
|
|
1480 | for (i = device_count; i--; ) |
|
|
1481 | av_store (av, i, newSViv (status_list [i])); |
|
|
1482 | } |
|
|
1483 | |
|
|
1484 | #if CL_VERSION_1_2 |
|
|
1485 | |
|
|
1486 | void |
|
|
1487 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1488 | PPCODE: |
|
|
1489 | cl_uint device_count; |
|
|
1490 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1491 | |
|
|
1492 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1493 | |
|
|
1494 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1495 | |
|
|
1496 | void |
|
|
1497 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1498 | CODE: |
|
|
1499 | cl_uint device_count = 0; |
|
|
1500 | cl_device_id *device_list = 0; |
|
|
1501 | |
|
|
1502 | if (SvOK (devices)) |
|
|
1503 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1504 | |
|
|
1505 | cl_uint program_count; |
|
|
1506 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1507 | |
|
|
1508 | void *user_data; |
|
|
1509 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1510 | |
|
|
1511 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1512 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1513 | |
|
|
1514 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1515 | |
|
|
1516 | #endif |
|
|
1517 | |
1284 | #BEGIN:context |
1518 | #BEGIN:context |
1285 | |
1519 | |
1286 | void |
1520 | void |
1287 | reference_count (OpenCL::Context self) |
1521 | reference_count (OpenCL::Context self) |
1288 | ALIAS: |
1522 | ALIAS: |
… | |
… | |
1303 | cl_device_id *value = tmpbuf (size); |
1537 | cl_device_id *value = tmpbuf (size); |
1304 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1538 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1305 | int i, n = size / sizeof (*value); |
1539 | int i, n = size / sizeof (*value); |
1306 | EXTEND (SP, n); |
1540 | EXTEND (SP, n); |
1307 | for (i = 0; i < n; ++i) |
1541 | for (i = 0; i < n; ++i) |
1308 | { |
|
|
1309 | PUSH_CLOBJ (stash_device, value [i]); |
1542 | PUSH_CLOBJ (stash_device, value [i]); |
1310 | } |
|
|
1311 | |
1543 | |
1312 | void |
1544 | void |
1313 | properties (OpenCL::Context self) |
1545 | properties (OpenCL::Context self) |
1314 | PPCODE: |
1546 | PPCODE: |
1315 | size_t size; |
1547 | size_t size; |
… | |
… | |
1333 | void |
1565 | void |
1334 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1566 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1335 | ALIAS: |
1567 | ALIAS: |
1336 | enqueue_read_buffer = 0 |
1568 | enqueue_read_buffer = 0 |
1337 | PPCODE: |
1569 | PPCODE: |
1338 | cl_event ev = 0; |
|
|
1339 | EVENT_LIST (6); |
1570 | EVENT_LIST (6); |
1340 | |
1571 | |
1341 | SvUPGRADE (data, SVt_PV); |
1572 | SvUPGRADE (data, SVt_PV); |
1342 | SvGROW (data, len); |
1573 | SvGROW (data, len); |
1343 | SvPOK_only (data); |
1574 | SvPOK_only (data); |
1344 | SvCUR_set (data, len); |
1575 | SvCUR_set (data, len); |
|
|
1576 | |
|
|
1577 | cl_event 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)); |
1578 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1346 | |
1579 | |
1347 | if (ev) |
1580 | if (ev) |
1348 | XPUSH_CLOBJ (stash_event, ev); |
1581 | XPUSH_CLOBJ (stash_event, ev); |
1349 | |
1582 | |
1350 | void |
1583 | void |
1351 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1584 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1352 | ALIAS: |
1585 | ALIAS: |
1353 | enqueue_write_buffer = 0 |
1586 | enqueue_write_buffer = 0 |
1354 | PPCODE: |
1587 | PPCODE: |
1355 | cl_event ev = 0; |
1588 | EVENT_LIST (5); |
|
|
1589 | |
1356 | STRLEN len; |
1590 | STRLEN len; |
1357 | char *ptr = SvPVbyte (data, len); |
1591 | char *ptr = SvPVbyte (data, len); |
1358 | EVENT_LIST (5); |
|
|
1359 | |
1592 | |
|
|
1593 | cl_event ev = 0; |
1360 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1594 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1361 | |
1595 | |
1362 | if (ev) |
1596 | if (ev) |
1363 | XPUSH_CLOBJ (stash_event, ev); |
1597 | XPUSH_CLOBJ (stash_event, ev); |
1364 | |
1598 | |
… | |
… | |
1367 | void |
1601 | void |
1368 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
1602 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
1369 | ALIAS: |
1603 | ALIAS: |
1370 | enqueue_fill_buffer = 0 |
1604 | enqueue_fill_buffer = 0 |
1371 | PPCODE: |
1605 | PPCODE: |
1372 | cl_event ev = 0; |
1606 | EVENT_LIST (5); |
|
|
1607 | |
1373 | STRLEN len; |
1608 | STRLEN len; |
1374 | char *ptr = SvPVbyte (data, len); |
1609 | char *ptr = SvPVbyte (data, len); |
1375 | EVENT_LIST (5); |
|
|
1376 | |
1610 | |
|
|
1611 | cl_event ev = 0; |
1377 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1612 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1378 | |
1613 | |
1379 | if (ev) |
1614 | if (ev) |
1380 | XPUSH_CLOBJ (stash_event, ev); |
1615 | XPUSH_CLOBJ (stash_event, ev); |
1381 | |
1616 | |
1382 | void |
1617 | 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, ...) |
1618 | 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: |
1619 | ALIAS: |
1385 | enqueue_fill_image = 0 |
1620 | enqueue_fill_image = 0 |
1386 | PPCODE: |
1621 | PPCODE: |
1387 | cl_event ev = 0; |
1622 | EVENT_LIST (12); |
1388 | STRLEN len; |
1623 | |
1389 | const size_t origin [3] = { x, y, z }; |
1624 | const size_t origin [3] = { x, y, z }; |
1390 | const size_t region [3] = { width, height, depth }; |
1625 | const size_t region [3] = { width, height, depth }; |
1391 | EVENT_LIST (12); |
|
|
1392 | |
1626 | |
1393 | const cl_float c_f [4] = { r, g, b, a }; |
1627 | const cl_float c_f [4] = { r, g, b, a }; |
1394 | const cl_uint c_u [4] = { r, g, b, a }; |
1628 | const cl_uint c_u [4] = { r, g, b, a }; |
1395 | const cl_int c_s [4] = { r, g, b, a }; |
1629 | const cl_int c_s [4] = { r, g, b, a }; |
1396 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
1630 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
… | |
… | |
1399 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
1633 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
1400 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
1634 | 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) |
1635 | 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."); |
1636 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
1403 | |
1637 | |
|
|
1638 | cl_event ev = 0; |
1404 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
1639 | 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)); |
1640 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1406 | |
1641 | |
1407 | if (ev) |
1642 | if (ev) |
1408 | XPUSH_CLOBJ (stash_event, ev); |
1643 | XPUSH_CLOBJ (stash_event, ev); |
… | |
… | |
1412 | void |
1647 | void |
1413 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1648 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1414 | ALIAS: |
1649 | ALIAS: |
1415 | enqueue_copy_buffer = 0 |
1650 | enqueue_copy_buffer = 0 |
1416 | PPCODE: |
1651 | PPCODE: |
|
|
1652 | EVENT_LIST (6); |
|
|
1653 | |
1417 | cl_event ev = 0; |
1654 | cl_event ev = 0; |
1418 | EVENT_LIST (6); |
|
|
1419 | |
|
|
1420 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1655 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1421 | |
1656 | |
1422 | if (ev) |
1657 | if (ev) |
1423 | XPUSH_CLOBJ (stash_event, ev); |
1658 | XPUSH_CLOBJ (stash_event, ev); |
1424 | |
1659 | |
1425 | void |
1660 | void |
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, ...) |
1661 | 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: |
1662 | ALIAS: |
1428 | enqueue_read_buffer_rect = 0 |
1663 | enqueue_read_buffer_rect = 0 |
1429 | PPCODE: |
1664 | PPCODE: |
1430 | cl_event ev = 0; |
1665 | EVENT_LIST (17); |
|
|
1666 | |
1431 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1667 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1432 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1668 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1433 | const size_t region[3] = { width, height, depth }; |
1669 | const size_t region[3] = { width, height, depth }; |
1434 | EVENT_LIST (17); |
|
|
1435 | |
1670 | |
1436 | if (!buf_row_pitch) |
1671 | if (!buf_row_pitch) |
1437 | buf_row_pitch = region [0]; |
1672 | buf_row_pitch = region [0]; |
1438 | |
1673 | |
1439 | if (!buf_slice_pitch) |
1674 | if (!buf_slice_pitch) |
… | |
… | |
1449 | |
1684 | |
1450 | SvUPGRADE (data, SVt_PV); |
1685 | SvUPGRADE (data, SVt_PV); |
1451 | SvGROW (data, len); |
1686 | SvGROW (data, len); |
1452 | SvPOK_only (data); |
1687 | SvPOK_only (data); |
1453 | SvCUR_set (data, len); |
1688 | SvCUR_set (data, len); |
|
|
1689 | |
|
|
1690 | cl_event 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)); |
1691 | 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)); |
1455 | |
1692 | |
1456 | if (ev) |
1693 | if (ev) |
1457 | XPUSH_CLOBJ (stash_event, ev); |
1694 | XPUSH_CLOBJ (stash_event, ev); |
1458 | |
1695 | |
1459 | void |
1696 | void |
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, ...) |
1697 | 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: |
1698 | ALIAS: |
1462 | enqueue_write_buffer_rect = 0 |
1699 | enqueue_write_buffer_rect = 0 |
1463 | PPCODE: |
1700 | PPCODE: |
1464 | cl_event ev = 0; |
1701 | EVENT_LIST (17); |
|
|
1702 | |
1465 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1703 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1466 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1704 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1467 | const size_t region[3] = { width, height, depth }; |
1705 | const size_t region[3] = { width, height, depth }; |
1468 | STRLEN len; |
1706 | STRLEN len; |
1469 | char *ptr = SvPVbyte (data, len); |
1707 | char *ptr = SvPVbyte (data, len); |
1470 | EVENT_LIST (17); |
|
|
1471 | |
1708 | |
1472 | if (!buf_row_pitch) |
1709 | if (!buf_row_pitch) |
1473 | buf_row_pitch = region [0]; |
1710 | buf_row_pitch = region [0]; |
1474 | |
1711 | |
1475 | if (!buf_slice_pitch) |
1712 | if (!buf_slice_pitch) |
… | |
… | |
1484 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1721 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1485 | |
1722 | |
1486 | if (len < min_len) |
1723 | if (len < min_len) |
1487 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1724 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1488 | |
1725 | |
|
|
1726 | cl_event 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)); |
1727 | 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)); |
1490 | |
1728 | |
1491 | if (ev) |
1729 | if (ev) |
1492 | XPUSH_CLOBJ (stash_event, ev); |
1730 | XPUSH_CLOBJ (stash_event, ev); |
1493 | |
1731 | |
1494 | void |
1732 | void |
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, ...) |
1733 | 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: |
1734 | ALIAS: |
1497 | enqueue_copy_buffer_rect = 0 |
1735 | enqueue_copy_buffer_rect = 0 |
1498 | PPCODE: |
1736 | PPCODE: |
1499 | cl_event ev = 0; |
1737 | EVENT_LIST (16); |
|
|
1738 | |
1500 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1739 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1501 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1740 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1502 | const size_t region[3] = { width, height, depth }; |
1741 | const size_t region[3] = { width, height, depth }; |
1503 | EVENT_LIST (16); |
|
|
1504 | |
1742 | |
|
|
1743 | cl_event 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)); |
1744 | 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)); |
1506 | |
1745 | |
1507 | if (ev) |
1746 | if (ev) |
1508 | XPUSH_CLOBJ (stash_event, ev); |
1747 | XPUSH_CLOBJ (stash_event, ev); |
1509 | |
1748 | |
1510 | void |
1749 | void |
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, ...) |
1750 | 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: |
1751 | ALIAS: |
1513 | enqueue_read_image = 0 |
1752 | enqueue_read_image = 0 |
1514 | PPCODE: |
1753 | PPCODE: |
1515 | cl_event ev = 0; |
1754 | EVENT_LIST (12); |
|
|
1755 | |
1516 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1756 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1517 | const size_t region[3] = { width, height, depth }; |
1757 | const size_t region[3] = { width, height, depth }; |
1518 | EVENT_LIST (12); |
|
|
1519 | |
1758 | |
1520 | if (!row_pitch) |
1759 | if (!row_pitch) |
1521 | row_pitch = img_row_pitch (src); |
1760 | row_pitch = img_row_pitch (src); |
1522 | |
1761 | |
1523 | if (depth > 1 && !slice_pitch) |
1762 | if (depth > 1 && !slice_pitch) |
… | |
… | |
1527 | |
1766 | |
1528 | SvUPGRADE (data, SVt_PV); |
1767 | SvUPGRADE (data, SVt_PV); |
1529 | SvGROW (data, len); |
1768 | SvGROW (data, len); |
1530 | SvPOK_only (data); |
1769 | SvPOK_only (data); |
1531 | SvCUR_set (data, len); |
1770 | SvCUR_set (data, len); |
|
|
1771 | |
|
|
1772 | cl_event 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)); |
1773 | 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)); |
1533 | |
1774 | |
1534 | if (ev) |
1775 | if (ev) |
1535 | XPUSH_CLOBJ (stash_event, ev); |
1776 | XPUSH_CLOBJ (stash_event, ev); |
1536 | |
1777 | |
1537 | void |
1778 | void |
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, ...) |
1779 | 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: |
1780 | ALIAS: |
1540 | enqueue_write_image = 0 |
1781 | enqueue_write_image = 0 |
1541 | PPCODE: |
1782 | PPCODE: |
1542 | cl_event ev = 0; |
1783 | EVENT_LIST (12); |
|
|
1784 | |
1543 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1785 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1544 | const size_t region[3] = { width, height, depth }; |
1786 | const size_t region[3] = { width, height, depth }; |
1545 | STRLEN len; |
1787 | STRLEN len; |
1546 | char *ptr = SvPVbyte (data, len); |
1788 | char *ptr = SvPVbyte (data, len); |
1547 | EVENT_LIST (12); |
|
|
1548 | |
1789 | |
1549 | if (!row_pitch) |
1790 | if (!row_pitch) |
1550 | row_pitch = img_row_pitch (dst); |
1791 | row_pitch = img_row_pitch (dst); |
1551 | |
1792 | |
1552 | if (depth > 1 && !slice_pitch) |
1793 | if (depth > 1 && !slice_pitch) |
… | |
… | |
1555 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1796 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1556 | |
1797 | |
1557 | if (len < min_len) |
1798 | if (len < min_len) |
1558 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1799 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1559 | |
1800 | |
|
|
1801 | cl_event 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)); |
1802 | 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)); |
1561 | |
1803 | |
1562 | if (ev) |
1804 | if (ev) |
1563 | XPUSH_CLOBJ (stash_event, ev); |
1805 | XPUSH_CLOBJ (stash_event, ev); |
1564 | |
1806 | |
1565 | void |
1807 | void |
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, ...) |
1808 | 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: |
1809 | ALIAS: |
1568 | enqueue_copy_image = 0 |
1810 | enqueue_copy_image = 0 |
1569 | PPCODE: |
1811 | PPCODE: |
1570 | cl_event ev = 0; |
1812 | EVENT_LIST (12); |
|
|
1813 | |
1571 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1814 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1572 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1815 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1573 | const size_t region[3] = { width, height, depth }; |
1816 | const size_t region[3] = { width, height, depth }; |
1574 | EVENT_LIST (12); |
|
|
1575 | |
1817 | |
|
|
1818 | cl_event 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)); |
1819 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1577 | |
1820 | |
1578 | if (ev) |
1821 | if (ev) |
1579 | XPUSH_CLOBJ (stash_event, ev); |
1822 | XPUSH_CLOBJ (stash_event, ev); |
1580 | |
1823 | |
1581 | void |
1824 | void |
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, ...) |
1825 | 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: |
1826 | ALIAS: |
1584 | enqueue_copy_image_to_buffer = 0 |
1827 | enqueue_copy_image_to_buffer = 0 |
1585 | PPCODE: |
1828 | PPCODE: |
1586 | cl_event ev = 0; |
1829 | EVENT_LIST (10); |
|
|
1830 | |
1587 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1831 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1588 | const size_t region [3] = { width, height, depth }; |
1832 | const size_t region [3] = { width, height, depth }; |
1589 | EVENT_LIST (10); |
|
|
1590 | |
1833 | |
|
|
1834 | cl_event ev = 0; |
1591 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1835 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1592 | |
1836 | |
1593 | if (ev) |
1837 | if (ev) |
1594 | XPUSH_CLOBJ (stash_event, ev); |
1838 | XPUSH_CLOBJ (stash_event, ev); |
1595 | |
1839 | |
1596 | void |
1840 | 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, ...) |
1841 | 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: |
1842 | ALIAS: |
1599 | enqueue_copy_buffer_to_image = 0 |
1843 | enqueue_copy_buffer_to_image = 0 |
1600 | PPCODE: |
1844 | PPCODE: |
1601 | cl_event ev = 0; |
1845 | EVENT_LIST (10); |
|
|
1846 | |
1602 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1847 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1603 | const size_t region [3] = { width, height, depth }; |
1848 | const size_t region [3] = { width, height, depth }; |
1604 | EVENT_LIST (10); |
|
|
1605 | |
1849 | |
|
|
1850 | cl_event ev = 0; |
1606 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1851 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1607 | |
1852 | |
1608 | if (ev) |
1853 | if (ev) |
1609 | XPUSH_CLOBJ (stash_event, ev); |
1854 | XPUSH_CLOBJ (stash_event, ev); |
1610 | |
1855 | |
1611 | void |
1856 | 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, ...) |
1857 | 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, ...) |
1613 | ALIAS: |
1858 | ALIAS: |
1614 | enqueue_map_buffer = 0 |
1859 | enqueue_map_buffer = 0 |
1615 | PPCODE: |
1860 | PPCODE: |
|
|
1861 | EVENT_LIST (6); |
|
|
1862 | |
|
|
1863 | size_t cb = SvIV (cb_); |
|
|
1864 | |
|
|
1865 | if (!SvOK (cb_)) |
|
|
1866 | { |
|
|
1867 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1868 | cb -= offset; |
|
|
1869 | } |
|
|
1870 | |
1616 | cl_event ev; |
1871 | 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)); |
1872 | 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)); |
1873 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0, 1, cb, 1, 1)); |
1621 | |
1874 | |
1622 | void |
1875 | 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, ...) |
1876 | 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, ...) |
1624 | ALIAS: |
1877 | ALIAS: |
1625 | enqueue_map_image = 0 |
1878 | enqueue_map_image = 0 |
1626 | PPCODE: |
1879 | PPCODE: |
1627 | cl_event ev; |
1880 | size_t width = SvIV (width_); |
|
|
1881 | if (!SvOK (width_)) |
|
|
1882 | { |
|
|
1883 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1884 | width -= x; |
|
|
1885 | } |
|
|
1886 | |
|
|
1887 | size_t height = SvIV (width_); |
|
|
1888 | if (!SvOK (height_)) |
|
|
1889 | { |
|
|
1890 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1891 | height -= y; |
|
|
1892 | |
|
|
1893 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1894 | if (!height) |
|
|
1895 | height = 1; |
|
|
1896 | } |
|
|
1897 | |
|
|
1898 | size_t depth = SvIV (width_); |
|
|
1899 | if (!SvOK (depth_)) |
|
|
1900 | { |
|
|
1901 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1902 | depth -= z; |
|
|
1903 | |
|
|
1904 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1905 | if (!depth) |
|
|
1906 | depth = 1; |
|
|
1907 | } |
|
|
1908 | |
1628 | const size_t origin[3] = { x, y, z }; |
1909 | const size_t origin[3] = { x, y, z }; |
1629 | const size_t region[3] = { width, height, depth }; |
1910 | const size_t region[3] = { width, height, depth }; |
1630 | size_t row_pitch, slice_pitch; |
1911 | size_t row_pitch, slice_pitch; |
1631 | EVENT_LIST (11); |
1912 | EVENT_LIST (10); |
1632 | |
1913 | |
|
|
1914 | cl_event ev; |
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)); |
1915 | 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 | |
1916 | |
1635 | size_t cb = slice_pitch ? slice_pitch * region [2] |
1917 | size_t cb = slice_pitch ? slice_pitch * region [2] |
1636 | : row_pitch ? row_pitch * region [1] |
1918 | : row_pitch ? row_pitch * region [1] |
1637 | : region [0]; |
1919 | : region [0]; |
1638 | |
1920 | |
1639 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
1921 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch, 0, width, height, depth)); |
1640 | |
1922 | |
1641 | void |
1923 | void |
1642 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
1924 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
1643 | PPCODE: |
1925 | PPCODE: |
1644 | cl_uint event_list_count = items - 2; |
1926 | mapped_unmap (cv, ST (1), mapped, self, &ST (2), 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) |
1927 | if (GIMME_V != G_VOID) |
|
|
1928 | { |
|
|
1929 | clRetainEvent (mapped->event); |
1651 | XPUSH_CLOBJ (stash_event, mapped->event); |
1930 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1931 | } |
1652 | |
1932 | |
1653 | void |
1933 | void |
1654 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
1934 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
1655 | ALIAS: |
1935 | ALIAS: |
1656 | enqueue_task = 0 |
1936 | enqueue_task = 0 |
1657 | PPCODE: |
1937 | PPCODE: |
|
|
1938 | EVENT_LIST (2); |
|
|
1939 | |
1658 | cl_event ev = 0; |
1940 | cl_event ev = 0; |
1659 | EVENT_LIST (2); |
|
|
1660 | |
|
|
1661 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1941 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1662 | |
1942 | |
1663 | if (ev) |
1943 | if (ev) |
1664 | XPUSH_CLOBJ (stash_event, ev); |
1944 | XPUSH_CLOBJ (stash_event, ev); |
1665 | |
1945 | |
1666 | void |
1946 | void |
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, ...) |
1947 | 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: |
1948 | ALIAS: |
1669 | enqueue_nd_range_kernel = 0 |
1949 | enqueue_nd_range_kernel = 0 |
1670 | PPCODE: |
1950 | PPCODE: |
1671 | cl_event ev = 0; |
1951 | EVENT_LIST (5); |
|
|
1952 | |
1672 | size_t *gwo = 0, *gws, *lws = 0; |
1953 | size_t *gwo = 0, *gws, *lws = 0; |
1673 | int gws_len; |
1954 | int gws_len; |
1674 | size_t *lists; |
1955 | size_t *lists; |
1675 | int i; |
1956 | int i; |
1676 | EVENT_LIST (5); |
|
|
1677 | |
1957 | |
1678 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1958 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1679 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1959 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1680 | |
1960 | |
1681 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1961 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
… | |
… | |
1720 | if (!lws [i]) |
2000 | if (!lws [i]) |
1721 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
2001 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
1722 | } |
2002 | } |
1723 | } |
2003 | } |
1724 | |
2004 | |
|
|
2005 | cl_event 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)); |
2006 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1726 | |
2007 | |
1727 | if (ev) |
2008 | if (ev) |
1728 | XPUSH_CLOBJ (stash_event, ev); |
2009 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2010 | |
|
|
2011 | #if CL_VERSION_1_2 |
|
|
2012 | |
|
|
2013 | void |
|
|
2014 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
2015 | ALIAS: |
|
|
2016 | enqueue_migrate_mem_objects = 0 |
|
|
2017 | PPCODE: |
|
|
2018 | EVENT_LIST (3); |
|
|
2019 | |
|
|
2020 | cl_uint obj_count; |
|
|
2021 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2022 | |
|
|
2023 | cl_event ev = 0; |
|
|
2024 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2025 | |
|
|
2026 | if (ev) |
|
|
2027 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2028 | |
|
|
2029 | #endif |
1729 | |
2030 | |
1730 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2031 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1731 | |
2032 | |
1732 | void |
2033 | void |
1733 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2034 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1734 | ALIAS: |
2035 | ALIAS: |
1735 | release_gl_objects = 1 |
2036 | release_gl_objects = 1 |
1736 | enqueue_acquire_gl_objects = 0 |
2037 | enqueue_acquire_gl_objects = 0 |
1737 | enqueue_release_gl_objects = 1 |
2038 | enqueue_release_gl_objects = 1 |
1738 | PPCODE: |
2039 | PPCODE: |
1739 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
2040 | EVENT_LIST (2); |
1740 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
2041 | |
|
|
2042 | cl_uint obj_count; |
|
|
2043 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
1741 | |
2044 | |
1742 | cl_event ev = 0; |
2045 | cl_event ev = 0; |
1743 | EVENT_LIST (2); |
|
|
1744 | AV *av = (AV *)SvRV (objects); |
|
|
1745 | cl_uint num_objects = av_len (av) + 1; |
|
|
1746 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1747 | int i; |
|
|
1748 | |
|
|
1749 | for (i = num_objects; i--; ) |
|
|
1750 | object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1751 | |
2046 | |
1752 | if (ix) |
2047 | if (ix) |
1753 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2048 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1754 | else |
2049 | else |
1755 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2050 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1756 | |
2051 | |
1757 | if (ev) |
2052 | if (ev) |
1758 | XPUSH_CLOBJ (stash_event, ev); |
2053 | XPUSH_CLOBJ (stash_event, ev); |
1759 | |
2054 | |
1760 | #endif |
2055 | #endif |
… | |
… | |
1774 | void |
2069 | void |
1775 | marker (OpenCL::Queue self, ...) |
2070 | marker (OpenCL::Queue self, ...) |
1776 | ALIAS: |
2071 | ALIAS: |
1777 | enqueue_marker = 0 |
2072 | enqueue_marker = 0 |
1778 | PPCODE: |
2073 | PPCODE: |
|
|
2074 | EVENT_LIST (1); |
1779 | cl_event ev = 0; |
2075 | cl_event ev = 0; |
1780 | EVENT_LIST (1); |
|
|
1781 | #if PREFER_1_1 |
2076 | #if PREFER_1_1 |
1782 | if (!event_list_count) |
2077 | if (!event_list_count) |
1783 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
2078 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
1784 | else |
2079 | else |
1785 | #if CL_VERSION_1_2 |
2080 | #if CL_VERSION_1_2 |
… | |
… | |
1799 | void |
2094 | void |
1800 | barrier (OpenCL::Queue self, ...) |
2095 | barrier (OpenCL::Queue self, ...) |
1801 | ALIAS: |
2096 | ALIAS: |
1802 | enqueue_barrier = 0 |
2097 | enqueue_barrier = 0 |
1803 | PPCODE: |
2098 | PPCODE: |
|
|
2099 | EVENT_LIST (1); |
1804 | cl_event ev = 0; |
2100 | cl_event ev = 0; |
1805 | EVENT_LIST (1); |
|
|
1806 | #if PREFER_1_1 |
2101 | #if PREFER_1_1 |
1807 | if (!event_list_count && GIMME_V == G_VOID) |
2102 | if (!event_list_count && GIMME_V == G_VOID) |
1808 | NEED_SUCCESS (EnqueueBarrier, (self)); |
2103 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1809 | else |
2104 | else |
1810 | #if CL_VERSION_1_2 |
2105 | #if CL_VERSION_1_2 |
… | |
… | |
1846 | PPCODE: |
2141 | PPCODE: |
1847 | cl_context value [1]; |
2142 | cl_context value [1]; |
1848 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
2143 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1849 | EXTEND (SP, 1); |
2144 | EXTEND (SP, 1); |
1850 | const int i = 0; |
2145 | const int i = 0; |
1851 | { |
|
|
1852 | NEED_SUCCESS (RetainContext, (value [i])); |
2146 | NEED_SUCCESS (RetainContext, (value [i])); |
1853 | PUSH_CLOBJ (stash_context, value [i]); |
2147 | PUSH_CLOBJ (stash_context, value [i]); |
1854 | } |
|
|
1855 | |
2148 | |
1856 | void |
2149 | void |
1857 | device (OpenCL::Queue self) |
2150 | device (OpenCL::Queue self) |
1858 | PPCODE: |
2151 | PPCODE: |
1859 | cl_device_id value [1]; |
2152 | cl_device_id value [1]; |
1860 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
2153 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1861 | EXTEND (SP, 1); |
2154 | EXTEND (SP, 1); |
1862 | const int i = 0; |
2155 | const int i = 0; |
1863 | { |
|
|
1864 | PUSH_CLOBJ (stash_device, value [i]); |
2156 | PUSH_CLOBJ (stash_device, value [i]); |
1865 | } |
|
|
1866 | |
2157 | |
1867 | void |
2158 | void |
1868 | reference_count (OpenCL::Queue self) |
2159 | reference_count (OpenCL::Queue self) |
1869 | PPCODE: |
2160 | PPCODE: |
1870 | cl_uint value [1]; |
2161 | cl_uint value [1]; |
… | |
… | |
1874 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2165 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1875 | |
2166 | |
1876 | void |
2167 | void |
1877 | properties (OpenCL::Queue self) |
2168 | properties (OpenCL::Queue self) |
1878 | PPCODE: |
2169 | PPCODE: |
1879 | cl_command_queue_properties value [1]; |
2170 | cl_ulong value [1]; |
1880 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
2171 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1881 | EXTEND (SP, 1); |
2172 | EXTEND (SP, 1); |
1882 | const int i = 0; |
2173 | const int i = 0; |
1883 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2174 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1884 | |
2175 | |
… | |
… | |
1894 | void |
2185 | void |
1895 | info (OpenCL::Memory self, cl_mem_info name) |
2186 | info (OpenCL::Memory self, cl_mem_info name) |
1896 | PPCODE: |
2187 | PPCODE: |
1897 | INFO (MemObject) |
2188 | INFO (MemObject) |
1898 | |
2189 | |
|
|
2190 | void |
|
|
2191 | destructor_callback (OpenCL::Memory self, SV *notify) |
|
|
2192 | PPCODE: |
|
|
2193 | clSetMemObjectDestructorCallback (self, eq_destructor_notify, SvREFCNT_inc (s_get_cv (notify))); |
|
|
2194 | |
1899 | #BEGIN:mem |
2195 | #BEGIN:mem |
1900 | |
2196 | |
1901 | void |
2197 | void |
1902 | type (OpenCL::Memory self) |
2198 | type (OpenCL::Memory self) |
|
|
2199 | ALIAS: |
|
|
2200 | type = CL_MEM_TYPE |
|
|
2201 | map_count = CL_MEM_MAP_COUNT |
|
|
2202 | reference_count = CL_MEM_REFERENCE_COUNT |
1903 | PPCODE: |
2203 | PPCODE: |
1904 | cl_mem_object_type value [1]; |
2204 | cl_uint value [1]; |
1905 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
2205 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
1906 | EXTEND (SP, 1); |
2206 | EXTEND (SP, 1); |
1907 | const int i = 0; |
2207 | const int i = 0; |
1908 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2208 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1909 | |
2209 | |
1910 | void |
2210 | void |
1911 | flags (OpenCL::Memory self) |
2211 | flags (OpenCL::Memory self) |
1912 | PPCODE: |
2212 | PPCODE: |
1913 | cl_mem_flags value [1]; |
2213 | cl_ulong value [1]; |
1914 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
2214 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1915 | EXTEND (SP, 1); |
2215 | EXTEND (SP, 1); |
1916 | const int i = 0; |
2216 | const int i = 0; |
1917 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2217 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1918 | |
2218 | |
… | |
… | |
1936 | EXTEND (SP, 1); |
2236 | EXTEND (SP, 1); |
1937 | const int i = 0; |
2237 | const int i = 0; |
1938 | PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i]))); |
2238 | PUSHs (sv_2mortal (newSVuv ((IV)(intptr_t)value [i]))); |
1939 | |
2239 | |
1940 | void |
2240 | void |
1941 | map_count (OpenCL::Memory self) |
|
|
1942 | ALIAS: |
|
|
1943 | map_count = CL_MEM_MAP_COUNT |
|
|
1944 | reference_count = CL_MEM_REFERENCE_COUNT |
|
|
1945 | PPCODE: |
|
|
1946 | cl_uint value [1]; |
|
|
1947 | NEED_SUCCESS (GetMemObjectInfo, (self, ix, sizeof (value), value, 0)); |
|
|
1948 | EXTEND (SP, 1); |
|
|
1949 | const int i = 0; |
|
|
1950 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
1951 | |
|
|
1952 | void |
|
|
1953 | context (OpenCL::Memory self) |
2241 | context (OpenCL::Memory self) |
1954 | PPCODE: |
2242 | PPCODE: |
1955 | cl_context value [1]; |
2243 | cl_context value [1]; |
1956 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
2244 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1957 | EXTEND (SP, 1); |
2245 | EXTEND (SP, 1); |
1958 | const int i = 0; |
2246 | const int i = 0; |
1959 | { |
|
|
1960 | NEED_SUCCESS (RetainContext, (value [i])); |
2247 | NEED_SUCCESS (RetainContext, (value [i])); |
1961 | PUSH_CLOBJ (stash_context, value [i]); |
2248 | PUSH_CLOBJ (stash_context, value [i]); |
1962 | } |
|
|
1963 | |
2249 | |
1964 | void |
2250 | void |
1965 | associated_memobject (OpenCL::Memory self) |
2251 | associated_memobject (OpenCL::Memory self) |
1966 | PPCODE: |
2252 | PPCODE: |
1967 | cl_mem value [1]; |
2253 | cl_mem value [1]; |
1968 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
2254 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1969 | EXTEND (SP, 1); |
2255 | EXTEND (SP, 1); |
1970 | const int i = 0; |
2256 | const int i = 0; |
1971 | { |
|
|
1972 | NEED_SUCCESS (RetainMemObject, (value [i])); |
2257 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1973 | PUSH_CLOBJ (stash_memory, value [i]); |
2258 | PUSH_CLOBJ (stash_memory, value [i]); |
1974 | } |
|
|
1975 | |
2259 | |
1976 | #END:mem |
2260 | #END:mem |
1977 | |
2261 | |
1978 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2262 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1979 | |
2263 | |
… | |
… | |
2078 | |
2362 | |
2079 | #BEGIN:sampler |
2363 | #BEGIN:sampler |
2080 | |
2364 | |
2081 | void |
2365 | void |
2082 | reference_count (OpenCL::Sampler self) |
2366 | reference_count (OpenCL::Sampler self) |
|
|
2367 | ALIAS: |
|
|
2368 | reference_count = CL_SAMPLER_REFERENCE_COUNT |
|
|
2369 | normalized_coords = CL_SAMPLER_NORMALIZED_COORDS |
|
|
2370 | addressing_mode = CL_SAMPLER_ADDRESSING_MODE |
2083 | PPCODE: |
2371 | PPCODE: |
2084 | cl_uint value [1]; |
2372 | cl_uint value [1]; |
2085 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_REFERENCE_COUNT, sizeof (value), value, 0)); |
2373 | NEED_SUCCESS (GetSamplerInfo, (self, ix, sizeof (value), value, 0)); |
2086 | EXTEND (SP, 1); |
2374 | EXTEND (SP, 1); |
2087 | const int i = 0; |
2375 | const int i = 0; |
2088 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2376 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2089 | |
2377 | |
2090 | void |
2378 | void |
… | |
… | |
2092 | PPCODE: |
2380 | PPCODE: |
2093 | cl_context value [1]; |
2381 | cl_context value [1]; |
2094 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2382 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2095 | EXTEND (SP, 1); |
2383 | EXTEND (SP, 1); |
2096 | const int i = 0; |
2384 | const int i = 0; |
2097 | { |
|
|
2098 | NEED_SUCCESS (RetainContext, (value [i])); |
2385 | NEED_SUCCESS (RetainContext, (value [i])); |
2099 | PUSH_CLOBJ (stash_context, value [i]); |
2386 | PUSH_CLOBJ (stash_context, value [i]); |
2100 | } |
|
|
2101 | |
|
|
2102 | void |
|
|
2103 | normalized_coords (OpenCL::Sampler self) |
|
|
2104 | PPCODE: |
|
|
2105 | cl_addressing_mode value [1]; |
|
|
2106 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
|
|
2107 | EXTEND (SP, 1); |
|
|
2108 | const int i = 0; |
|
|
2109 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2110 | |
|
|
2111 | void |
|
|
2112 | addressing_mode (OpenCL::Sampler self) |
|
|
2113 | PPCODE: |
|
|
2114 | cl_filter_mode value [1]; |
|
|
2115 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
|
|
2116 | EXTEND (SP, 1); |
|
|
2117 | const int i = 0; |
|
|
2118 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2119 | |
2387 | |
2120 | void |
2388 | void |
2121 | filter_mode (OpenCL::Sampler self) |
2389 | filter_mode (OpenCL::Sampler self) |
2122 | PPCODE: |
2390 | PPCODE: |
2123 | cl_bool value [1]; |
2391 | cl_uint value [1]; |
2124 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0)); |
2392 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_FILTER_MODE, sizeof (value), value, 0)); |
2125 | EXTEND (SP, 1); |
2393 | EXTEND (SP, 1); |
2126 | const int i = 0; |
2394 | const int i = 0; |
2127 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
2395 | PUSHs (sv_2mortal (value [i] ? &PL_sv_yes : &PL_sv_no)); |
2128 | |
2396 | |
… | |
… | |
2138 | void |
2406 | void |
2139 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
2407 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
2140 | ALIAS: |
2408 | ALIAS: |
2141 | build_async = 1 |
2409 | build_async = 1 |
2142 | CODE: |
2410 | 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; |
2411 | cl_uint device_count = 0; |
2146 | cl_device_id *device_list = 0; |
2412 | cl_device_id *device_list = 0; |
2147 | |
2413 | |
2148 | if (SvOK (devices)) |
2414 | if (SvOK (devices)) |
2149 | { |
2415 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2416 | |
|
|
2417 | void *user_data; |
|
|
2418 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2419 | |
|
|
2420 | if (ix) |
|
|
2421 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2422 | else |
|
|
2423 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2424 | |
|
|
2425 | #if CL_VERSION_1_2 |
|
|
2426 | |
|
|
2427 | void |
|
|
2428 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2429 | CODE: |
|
|
2430 | cl_uint device_count = 0; |
|
|
2431 | cl_device_id *device_list = 0; |
|
|
2432 | |
|
|
2433 | if (SvOK (devices)) |
|
|
2434 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2435 | |
|
|
2436 | cl_uint header_count = 0; |
|
|
2437 | cl_program *header_list = 0; |
|
|
2438 | const char **header_name = 0; |
|
|
2439 | |
|
|
2440 | if (SvOK (headers)) |
|
|
2441 | { |
2150 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
2442 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
2151 | croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); |
2443 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
2152 | |
2444 | |
2153 | AV *av = (AV *)SvRV (devices); |
2445 | HV *hv = (HV *)SvRV (devices); |
2154 | num_devices = av_len (av) + 1; |
|
|
2155 | |
2446 | |
2156 | if (num_devices) |
2447 | header_count = hv_iterinit (hv); |
|
|
2448 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2449 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2450 | |
|
|
2451 | HE *he; |
|
|
2452 | int i = 0; |
|
|
2453 | while (he = hv_iternext (hv)) |
2157 | { |
2454 | { |
2158 | device_list = tmpbuf (sizeof (*device_list) * num_devices); |
2455 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2456 | header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program"); |
2159 | int count; |
2457 | ++i; |
2160 | for (count = 0; count < num_devices; ++count) |
|
|
2161 | device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device"); |
|
|
2162 | } |
2458 | } |
2163 | } |
2459 | } |
2164 | |
2460 | |
2165 | if (SvOK (notify)) |
2461 | void *user_data; |
2166 | { |
2462 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
2167 | NEED_SUCCESS (RetainProgram, (self)); |
|
|
2168 | pfn_notify = eq_program_notify; |
|
|
2169 | user_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
2170 | } |
|
|
2171 | |
2463 | |
2172 | if (ix) |
|
|
2173 | build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2174 | else |
|
|
2175 | NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
2464 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2465 | header_count, header_list, header_name, pfn_notify, user_data)); |
|
|
2466 | |
|
|
2467 | #endif |
2176 | |
2468 | |
2177 | void |
2469 | void |
2178 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2470 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2179 | PPCODE: |
2471 | PPCODE: |
2180 | size_t size; |
2472 | size_t size; |
… | |
… | |
2189 | #BEGIN:program_build |
2481 | #BEGIN:program_build |
2190 | |
2482 | |
2191 | void |
2483 | void |
2192 | build_status (OpenCL::Program self, OpenCL::Device device) |
2484 | build_status (OpenCL::Program self, OpenCL::Device device) |
2193 | PPCODE: |
2485 | PPCODE: |
2194 | cl_build_status value [1]; |
2486 | cl_int value [1]; |
2195 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0)); |
2487 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BUILD_STATUS, sizeof (value), value, 0)); |
2196 | EXTEND (SP, 1); |
2488 | EXTEND (SP, 1); |
2197 | const int i = 0; |
2489 | const int i = 0; |
2198 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2490 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2199 | |
2491 | |
… | |
… | |
2208 | char *value = tmpbuf (size); |
2500 | char *value = tmpbuf (size); |
2209 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
2501 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
2210 | EXTEND (SP, 1); |
2502 | EXTEND (SP, 1); |
2211 | const int i = 0; |
2503 | const int i = 0; |
2212 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
2504 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2505 | |
|
|
2506 | void |
|
|
2507 | binary_type (OpenCL::Program self, OpenCL::Device device) |
|
|
2508 | PPCODE: |
|
|
2509 | cl_uint value [1]; |
|
|
2510 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0)); |
|
|
2511 | EXTEND (SP, 1); |
|
|
2512 | const int i = 0; |
|
|
2513 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
2213 | |
2514 | |
2214 | #END:program_build |
2515 | #END:program_build |
2215 | |
2516 | |
2216 | void |
2517 | void |
2217 | kernel (OpenCL::Program program, SV *function) |
2518 | kernel (OpenCL::Program program, SV *function) |
… | |
… | |
2284 | PPCODE: |
2585 | PPCODE: |
2285 | cl_context value [1]; |
2586 | cl_context value [1]; |
2286 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2587 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2287 | EXTEND (SP, 1); |
2588 | EXTEND (SP, 1); |
2288 | const int i = 0; |
2589 | const int i = 0; |
2289 | { |
|
|
2290 | NEED_SUCCESS (RetainContext, (value [i])); |
2590 | NEED_SUCCESS (RetainContext, (value [i])); |
2291 | PUSH_CLOBJ (stash_context, value [i]); |
2591 | PUSH_CLOBJ (stash_context, value [i]); |
2292 | } |
|
|
2293 | |
2592 | |
2294 | void |
2593 | void |
2295 | devices (OpenCL::Program self) |
2594 | devices (OpenCL::Program self) |
2296 | PPCODE: |
2595 | PPCODE: |
2297 | size_t size; |
2596 | size_t size; |
… | |
… | |
2299 | cl_device_id *value = tmpbuf (size); |
2598 | cl_device_id *value = tmpbuf (size); |
2300 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2599 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2301 | int i, n = size / sizeof (*value); |
2600 | int i, n = size / sizeof (*value); |
2302 | EXTEND (SP, n); |
2601 | EXTEND (SP, n); |
2303 | for (i = 0; i < n; ++i) |
2602 | for (i = 0; i < n; ++i) |
2304 | { |
|
|
2305 | PUSH_CLOBJ (stash_device, value [i]); |
2603 | PUSH_CLOBJ (stash_device, value [i]); |
2306 | } |
|
|
2307 | |
2604 | |
2308 | void |
2605 | void |
2309 | source (OpenCL::Program self) |
2606 | source (OpenCL::Program self) |
2310 | PPCODE: |
2607 | PPCODE: |
2311 | size_t size; |
2608 | size_t size; |
… | |
… | |
2384 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
2681 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
2385 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
2682 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
2386 | |
2683 | |
2387 | case 'z': nullarg = 1; size = SvIV (sv); break; |
2684 | case 'z': nullarg = 1; size = SvIV (sv); break; |
2388 | |
2685 | |
2389 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
2686 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "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; |
2687 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "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; |
2688 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
2392 | |
2689 | |
2393 | default: |
2690 | default: |
2394 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
2691 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
2395 | } |
2692 | } |
2396 | |
2693 | |
… | |
… | |
2467 | CODE: |
2764 | CODE: |
2468 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
2765 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
2469 | |
2766 | |
2470 | void |
2767 | void |
2471 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
2768 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
2472 | ALIAS: |
|
|
2473 | set_image2d = 0 |
|
|
2474 | set_image3d = 0 |
|
|
2475 | CODE: |
2769 | CODE: |
2476 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
2770 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
2477 | |
2771 | |
2478 | void |
2772 | void |
2479 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
2773 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
… | |
… | |
2525 | PPCODE: |
2819 | PPCODE: |
2526 | cl_context value [1]; |
2820 | cl_context value [1]; |
2527 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2821 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2528 | EXTEND (SP, 1); |
2822 | EXTEND (SP, 1); |
2529 | const int i = 0; |
2823 | const int i = 0; |
2530 | { |
|
|
2531 | NEED_SUCCESS (RetainContext, (value [i])); |
2824 | NEED_SUCCESS (RetainContext, (value [i])); |
2532 | PUSH_CLOBJ (stash_context, value [i]); |
2825 | PUSH_CLOBJ (stash_context, value [i]); |
2533 | } |
|
|
2534 | |
2826 | |
2535 | void |
2827 | void |
2536 | program (OpenCL::Kernel self) |
2828 | program (OpenCL::Kernel self) |
2537 | PPCODE: |
2829 | PPCODE: |
2538 | cl_program value [1]; |
2830 | cl_program value [1]; |
2539 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2831 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2540 | EXTEND (SP, 1); |
2832 | EXTEND (SP, 1); |
2541 | const int i = 0; |
2833 | const int i = 0; |
2542 | { |
|
|
2543 | NEED_SUCCESS (RetainProgram, (value [i])); |
2834 | NEED_SUCCESS (RetainProgram, (value [i])); |
2544 | PUSH_CLOBJ (stash_program, value [i]); |
2835 | PUSH_CLOBJ (stash_program, value [i]); |
2545 | } |
|
|
2546 | |
2836 | |
2547 | #END:kernel |
2837 | #END:kernel |
2548 | |
2838 | |
2549 | void |
2839 | void |
2550 | work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name) |
2840 | work_group_info (OpenCL::Kernel self, OpenCL::Device device, cl_kernel_work_group_info name) |
… | |
… | |
2596 | const int i = 0; |
2886 | const int i = 0; |
2597 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2887 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2598 | |
2888 | |
2599 | #END:kernel_work_group |
2889 | #END:kernel_work_group |
2600 | |
2890 | |
|
|
2891 | #if CL_VERSION_1_2 |
|
|
2892 | |
|
|
2893 | void |
|
|
2894 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2895 | PPCODE: |
|
|
2896 | size_t size; |
|
|
2897 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2898 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2899 | SvUPGRADE (sv, SVt_PV); |
|
|
2900 | SvPOK_only (sv); |
|
|
2901 | SvCUR_set (sv, size); |
|
|
2902 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2903 | XPUSHs (sv); |
|
|
2904 | |
|
|
2905 | #BEGIN:kernel_arg |
|
|
2906 | |
|
|
2907 | void |
|
|
2908 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2909 | ALIAS: |
|
|
2910 | arg_address_qualifier = CL_KERNEL_ARG_ADDRESS_QUALIFIER |
|
|
2911 | arg_access_qualifier = CL_KERNEL_ARG_ACCESS_QUALIFIER |
|
|
2912 | PPCODE: |
|
|
2913 | cl_uint value [1]; |
|
|
2914 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, sizeof (value), value, 0)); |
|
|
2915 | EXTEND (SP, 1); |
|
|
2916 | const int i = 0; |
|
|
2917 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2918 | |
|
|
2919 | void |
|
|
2920 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2921 | ALIAS: |
|
|
2922 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2923 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2924 | PPCODE: |
|
|
2925 | size_t size; |
|
|
2926 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2927 | char *value = tmpbuf (size); |
|
|
2928 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2929 | EXTEND (SP, 1); |
|
|
2930 | const int i = 0; |
|
|
2931 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2932 | |
|
|
2933 | void |
|
|
2934 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2935 | PPCODE: |
|
|
2936 | cl_ulong value [1]; |
|
|
2937 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2938 | EXTEND (SP, 1); |
|
|
2939 | const int i = 0; |
|
|
2940 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2941 | |
|
|
2942 | #END:kernel_arg |
|
|
2943 | |
|
|
2944 | #endif |
|
|
2945 | |
2601 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2946 | MODULE = OpenCL PACKAGE = OpenCL::Event |
2602 | |
2947 | |
2603 | void |
2948 | void |
2604 | DESTROY (OpenCL::Event self) |
2949 | DESTROY (OpenCL::Event self) |
2605 | CODE: |
2950 | CODE: |
… | |
… | |
2627 | PPCODE: |
2972 | PPCODE: |
2628 | cl_command_queue value [1]; |
2973 | cl_command_queue value [1]; |
2629 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
2974 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
2630 | EXTEND (SP, 1); |
2975 | EXTEND (SP, 1); |
2631 | const int i = 0; |
2976 | const int i = 0; |
2632 | { |
|
|
2633 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
2977 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
2634 | PUSH_CLOBJ (stash_queue, value [i]); |
2978 | PUSH_CLOBJ (stash_queue, value [i]); |
2635 | } |
|
|
2636 | |
2979 | |
2637 | void |
2980 | void |
2638 | command_type (OpenCL::Event self) |
2981 | command_type (OpenCL::Event self) |
2639 | PPCODE: |
|
|
2640 | cl_command_type value [1]; |
|
|
2641 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_TYPE, sizeof (value), value, 0)); |
|
|
2642 | EXTEND (SP, 1); |
|
|
2643 | const int i = 0; |
|
|
2644 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2645 | |
|
|
2646 | void |
|
|
2647 | reference_count (OpenCL::Event self) |
|
|
2648 | ALIAS: |
2982 | ALIAS: |
|
|
2983 | command_type = CL_EVENT_COMMAND_TYPE |
2649 | reference_count = CL_EVENT_REFERENCE_COUNT |
2984 | reference_count = CL_EVENT_REFERENCE_COUNT |
2650 | command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS |
2985 | command_execution_status = CL_EVENT_COMMAND_EXECUTION_STATUS |
2651 | PPCODE: |
2986 | PPCODE: |
2652 | cl_uint value [1]; |
2987 | cl_uint value [1]; |
2653 | NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0)); |
2988 | NEED_SUCCESS (GetEventInfo, (self, ix, sizeof (value), value, 0)); |
… | |
… | |
2660 | PPCODE: |
2995 | PPCODE: |
2661 | cl_context value [1]; |
2996 | cl_context value [1]; |
2662 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
2997 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
2663 | EXTEND (SP, 1); |
2998 | EXTEND (SP, 1); |
2664 | const int i = 0; |
2999 | const int i = 0; |
2665 | { |
|
|
2666 | NEED_SUCCESS (RetainContext, (value [i])); |
3000 | NEED_SUCCESS (RetainContext, (value [i])); |
2667 | PUSH_CLOBJ (stash_context, value [i]); |
3001 | PUSH_CLOBJ (stash_context, value [i]); |
2668 | } |
|
|
2669 | |
3002 | |
2670 | #END:event |
3003 | #END:event |
2671 | |
3004 | |
2672 | void |
3005 | void |
2673 | profiling_info (OpenCL::Event self, cl_profiling_info name) |
3006 | profiling_info (OpenCL::Event self, cl_profiling_info name) |
… | |
… | |
2708 | |
3041 | |
2709 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
3042 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
2710 | mapped_detach (self, mapped); |
3043 | mapped_detach (self, mapped); |
2711 | |
3044 | |
2712 | clReleaseCommandQueue (mapped->queue); |
3045 | clReleaseCommandQueue (mapped->queue); |
2713 | |
|
|
2714 | if (mapped->event) |
|
|
2715 | clReleaseEvent (mapped->event); |
3046 | clReleaseEvent (mapped->event); |
2716 | |
|
|
2717 | Safefree (mapped); |
3047 | Safefree (mapped); |
|
|
3048 | |
|
|
3049 | void |
|
|
3050 | unmap (OpenCL::Mapped self, ...) |
|
|
3051 | CODE: |
|
|
3052 | mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1); |
2718 | |
3053 | |
2719 | bool |
3054 | bool |
2720 | mapped (OpenCL::Mapped self) |
3055 | mapped (OpenCL::Mapped self) |
2721 | CODE: |
3056 | CODE: |
2722 | RETVAL = !!self->ptr; |
3057 | RETVAL = !!self->ptr; |
… | |
… | |
2736 | XSRETURN_UNDEF; |
3071 | XSRETURN_UNDEF; |
2737 | |
3072 | |
2738 | clRetainEvent (self->event); |
3073 | clRetainEvent (self->event); |
2739 | XPUSH_CLOBJ (stash_event, self->event); |
3074 | XPUSH_CLOBJ (stash_event, self->event); |
2740 | |
3075 | |
2741 | size_t |
3076 | #define MAPPED_OFFSET_CB offsetof (struct mapped, cb) |
|
|
3077 | #define MAPPED_OFFSET_ROW_PITCH offsetof (struct mapped, row_pitch) |
|
|
3078 | #define MAPPED_OFFSET_SLICE_PITCH offsetof (struct mapped, slice_pitch) |
|
|
3079 | #define MAPPED_OFFSET_WIDTH offsetof (struct mapped, width) |
|
|
3080 | #define MAPPED_OFFSET_HEIGHT offsetof (struct mapped, height) |
|
|
3081 | #define MAPPED_OFFSET_DEPTH offsetof (struct mapped, depth) |
|
|
3082 | |
|
|
3083 | IV |
2742 | size (OpenCL::Mapped self) |
3084 | size (OpenCL::Mapped self) |
|
|
3085 | ALIAS: |
|
|
3086 | size = MAPPED_OFFSET_CB |
|
|
3087 | row_pitch = MAPPED_OFFSET_ROW_PITCH |
|
|
3088 | slice_pitch = MAPPED_OFFSET_SLICE_PITCH |
|
|
3089 | width = MAPPED_OFFSET_WIDTH |
|
|
3090 | height = MAPPED_OFFSET_HEIGHT |
|
|
3091 | depth = MAPPED_OFFSET_DEPTH |
2743 | CODE: |
3092 | CODE: |
2744 | RETVAL = self->cb; |
3093 | RETVAL = *(size_t *)((char *)self + ix); |
2745 | OUTPUT: |
3094 | OUTPUT: |
2746 | RETVAL |
3095 | RETVAL |
2747 | |
3096 | |
2748 | IV |
3097 | IV |
2749 | ptr (OpenCL::Mapped self) |
3098 | ptr (OpenCL::Mapped self) |
2750 | CODE: |
3099 | CODE: |
2751 | RETVAL = PTR2IV (self->ptr); |
3100 | RETVAL = PTR2IV (self->ptr); |
2752 | OUTPUT: |
3101 | OUTPUT: |
2753 | RETVAL |
3102 | RETVAL |
2754 | |
3103 | |
|
|
3104 | void |
|
|
3105 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3106 | CODE: |
|
|
3107 | STRLEN len; |
|
|
3108 | const char *ptr = SvPVbyte (data, len); |
|
|
3109 | |
|
|
3110 | if (offset + len > self->cb) |
|
|
3111 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3112 | |
|
|
3113 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3114 | |
|
|
3115 | void |
|
|
3116 | get_row (OpenCL::Mapped self, size_t count, size_t x = 0, size_t y = 0, size_t z = 0) |
|
|
3117 | PPCODE: |
|
|
3118 | if (!SvOK (ST (1))) |
|
|
3119 | count = self->width - x; |
|
|
3120 | |
|
|
3121 | if (x + count > self->width) |
|
|
3122 | croak ("OpenCL::Mapped::get: x + count crosses a row boundary"); |
|
|
3123 | |
|
|
3124 | if (y >= self->height) |
|
|
3125 | croak ("OpenCL::Mapped::get: y coordinate out of bounds"); |
|
|
3126 | |
|
|
3127 | if (z >= self->depth) |
|
|
3128 | croak ("OpenCL::Mapped::get: z coordinate out of bounds"); |
|
|
3129 | |
|
|
3130 | size_t element = mapped_element_size (self); |
|
|
3131 | |
|
|
3132 | count *= element; |
|
|
3133 | x *= element; |
|
|
3134 | |
|
|
3135 | char *ptr = (char *)self->ptr + x + y * self->row_pitch + z * self->slice_pitch; |
|
|
3136 | XPUSHs (sv_2mortal (newSVpvn (ptr, count))); |
|
|
3137 | |
|
|
3138 | void |
|
|
3139 | set_row (OpenCL::Mapped self, SV *data, size_t x = 0, size_t y = 0, size_t z = 0) |
|
|
3140 | PPCODE: |
|
|
3141 | STRLEN count; |
|
|
3142 | char *dataptr = SvPVbyte (data, count); |
|
|
3143 | size_t element = mapped_element_size (self); |
|
|
3144 | |
|
|
3145 | x *= element; |
|
|
3146 | |
|
|
3147 | if (x + count > self->width * element) |
|
|
3148 | croak ("OpenCL::Mapped::set: x + data size crosses a row boundary"); |
|
|
3149 | |
|
|
3150 | if (y >= self->height) |
|
|
3151 | croak ("OpenCL::Mapped::set: y coordinate out of bounds"); |
|
|
3152 | |
|
|
3153 | if (z >= self->depth) |
|
|
3154 | croak ("OpenCL::Mapped::set: z coordinate out of bounds"); |
|
|
3155 | |
|
|
3156 | char *ptr = (char *)self->ptr + x + y * self->row_pitch + z * self->slice_pitch; |
|
|
3157 | memcpy (ptr, dataptr, count); |
|
|
3158 | |
2755 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
3159 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
2756 | |
3160 | |
2757 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
3161 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
2758 | |
3162 | |
2759 | IV |
3163 | IV |
2760 | row_pitch (OpenCL::Mapped self) |
3164 | element_size (OpenCL::Mapped self) |
2761 | ALIAS: |
|
|
2762 | slice_pitch = 1 |
|
|
2763 | CODE: |
3165 | CODE: |
2764 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
3166 | RETVAL = mapped_element_size (self); |
2765 | OUTPUT: |
3167 | OUTPUT: |
2766 | RETVAL |
3168 | RETVAL |
2767 | |
3169 | |
2768 | |
|
|
2769 | |
|
|