1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
4 | |
4 | |
|
|
5 | #include "ecb.h"//D |
|
|
6 | |
|
|
7 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
8 | #include "xthread.h" |
|
|
9 | #include "schmorp.h" |
|
|
10 | |
5 | #ifdef I_DLFCN |
11 | #ifdef I_DLFCN |
6 | #include <dlfcn.h> |
12 | #include <dlfcn.h> |
7 | #endif |
13 | #endif |
8 | |
14 | |
|
|
15 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
16 | // but then fail to define the api functions unless you ALSO define |
|
|
17 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
18 | // after all, the functions are deprecated, not removed. |
|
|
19 | // in addition, you cannot test for this in any future-proof way. |
|
|
20 | // each time a new opencl version comes out, you need to make a new |
|
|
21 | // release. |
|
|
22 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
23 | |
|
|
24 | #ifndef PREFER_1_1 |
|
|
25 | #define PREFER_1_1 1 |
|
|
26 | #endif |
|
|
27 | |
|
|
28 | #if PREFER_1_1 |
|
|
29 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
30 | #endif |
|
|
31 | |
9 | #ifdef __APPLE__ |
32 | #ifdef __APPLE__ |
10 | #include <OpenCL/opencl.h> |
33 | #define CLHDR(name) <OpenCL/name> |
11 | #else |
34 | #else |
12 | #include <CL/opencl.h> |
35 | #define CLHDR(name) <CL/name> |
13 | #endif |
36 | #endif |
|
|
37 | |
|
|
38 | #include CLHDR(opencl.h) |
|
|
39 | |
|
|
40 | #ifndef CL_VERSION_1_2 |
|
|
41 | #include CLHDR(cl_d3d9.h) |
|
|
42 | #endif |
|
|
43 | |
|
|
44 | #if _WIN32 |
|
|
45 | #include CLHDR(cl_d3d10.h) |
|
|
46 | #if CL_VERSION_1_2 |
|
|
47 | #include CLHDR<cl_d3d11.h> |
|
|
48 | #endif |
|
|
49 | #include CLHDR<cl_dx9_media_sharing.h.h> |
|
|
50 | #endif |
|
|
51 | |
|
|
52 | #ifndef CL_VERSION_1_2 |
|
|
53 | #undef PREFER_1_1 |
|
|
54 | #define PREFER_1_1 1 |
|
|
55 | #endif |
|
|
56 | |
|
|
57 | // make sure all constants we might use are actually defined |
|
|
58 | #include "default.h" |
14 | |
59 | |
15 | typedef cl_platform_id OpenCL__Platform; |
60 | typedef cl_platform_id OpenCL__Platform; |
16 | typedef cl_device_id OpenCL__Device; |
61 | typedef cl_device_id OpenCL__Device; |
|
|
62 | typedef cl_device_id OpenCL__SubDevice; |
17 | typedef cl_context OpenCL__Context; |
63 | typedef cl_context OpenCL__Context; |
18 | typedef cl_command_queue OpenCL__Queue; |
64 | typedef cl_command_queue OpenCL__Queue; |
19 | typedef cl_mem OpenCL__Memory; |
65 | typedef cl_mem OpenCL__Memory; |
20 | typedef cl_mem OpenCL__Buffer; |
66 | typedef cl_mem OpenCL__Buffer; |
21 | typedef cl_mem OpenCL__BufferObj; |
67 | typedef cl_mem OpenCL__BufferObj; |
22 | typedef cl_mem OpenCL__Image; |
68 | typedef cl_mem OpenCL__Image; |
23 | typedef cl_mem OpenCL__Image2D; |
|
|
24 | typedef cl_mem OpenCL__Image3D; |
|
|
25 | typedef cl_mem OpenCL__Memory_ornull; |
69 | typedef cl_mem OpenCL__Memory_ornull; |
26 | typedef cl_mem OpenCL__Buffer_ornull; |
70 | typedef cl_mem OpenCL__Buffer_ornull; |
27 | typedef cl_mem OpenCL__Image_ornull; |
71 | typedef cl_mem OpenCL__Image_ornull; |
28 | typedef cl_mem OpenCL__Image2D_ornull; |
|
|
29 | typedef cl_mem OpenCL__Image3D_ornull; |
|
|
30 | typedef cl_sampler OpenCL__Sampler; |
72 | typedef cl_sampler OpenCL__Sampler; |
31 | typedef cl_program OpenCL__Program; |
73 | typedef cl_program OpenCL__Program; |
32 | typedef cl_kernel OpenCL__Kernel; |
74 | typedef cl_kernel OpenCL__Kernel; |
33 | typedef cl_event OpenCL__Event; |
75 | typedef cl_event OpenCL__Event; |
34 | typedef cl_event OpenCL__UserEvent; |
76 | typedef cl_event OpenCL__UserEvent; |
35 | |
77 | |
36 | typedef SV *FUTURE; |
78 | typedef struct mapped * OpenCL__Mapped; |
|
|
79 | |
|
|
80 | static HV |
|
|
81 | *stash_platform, |
|
|
82 | *stash_device, |
|
|
83 | *stash_subdevice, |
|
|
84 | *stash_context, |
|
|
85 | *stash_queue, |
|
|
86 | *stash_program, |
|
|
87 | *stash_kernel, |
|
|
88 | *stash_sampler, |
|
|
89 | *stash_event, |
|
|
90 | *stash_userevent, |
|
|
91 | *stash_memory, |
|
|
92 | *stash_buffer, |
|
|
93 | *stash_bufferobj, |
|
|
94 | *stash_image, |
|
|
95 | *stash_image1d, |
|
|
96 | *stash_image1darray, |
|
|
97 | *stash_image1dbuffer, |
|
|
98 | *stash_image2d, |
|
|
99 | *stash_image2darray, |
|
|
100 | *stash_image3d, |
|
|
101 | *stash_mapped, |
|
|
102 | *stash_mappedbuffer, |
|
|
103 | *stash_mappedimage; |
37 | |
104 | |
38 | /*****************************************************************************/ |
105 | /*****************************************************************************/ |
39 | |
106 | |
40 | // name must include a leading underscore |
107 | // name must include a leading underscore |
41 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
108 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
… | |
… | |
81 | } |
148 | } |
82 | |
149 | |
83 | /*****************************************************************************/ |
150 | /*****************************************************************************/ |
84 | |
151 | |
85 | /* up to two temporary buffers */ |
152 | /* up to two temporary buffers */ |
86 | static void * |
153 | static void * ecb_noinline |
87 | tmpbuf (size_t size) |
154 | tmpbuf (size_t size) |
88 | { |
155 | { |
89 | enum { buffers = 3 }; |
156 | enum { buffers = 4 }; |
90 | static int idx; |
157 | static int idx; |
91 | static void *buf [buffers]; |
158 | static void *buf [buffers]; |
92 | static size_t len [buffers]; |
159 | static size_t len [buffers]; |
93 | |
160 | |
94 | idx = ++idx % buffers; |
161 | idx = (idx + 1) % buffers; |
95 | |
162 | |
96 | if (len [idx] < size) |
163 | if (len [idx] < size) |
97 | { |
164 | { |
98 | free (buf [idx]); |
165 | free (buf [idx]); |
99 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
166 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
101 | } |
168 | } |
102 | |
169 | |
103 | return buf [idx]; |
170 | return buf [idx]; |
104 | } |
171 | } |
105 | |
172 | |
|
|
173 | static const char * ecb_noinline |
|
|
174 | cv_get_name (CV *cv) |
|
|
175 | { |
|
|
176 | static char fullname [256]; |
|
|
177 | |
|
|
178 | GV *gv = CvGV (cv); // gv better be nonzero |
|
|
179 | |
|
|
180 | HV *stash = GvSTASH (gv); |
|
|
181 | const char *hvname = HvNAME_get (stash); // stash also better be nonzero |
|
|
182 | const char *gvname = GvNAME (gv); |
|
|
183 | |
|
|
184 | snprintf (fullname, sizeof (fullname), "%s::%s", hvname, gvname); |
|
|
185 | return fullname; |
|
|
186 | } |
|
|
187 | |
106 | /*****************************************************************************/ |
188 | /*****************************************************************************/ |
107 | |
189 | |
108 | typedef struct |
190 | typedef struct |
109 | { |
191 | { |
110 | IV iv; |
192 | IV iv; |
111 | const char *name; |
193 | const char *name; |
112 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
194 | #define const_iv(name) { (IV)CL_ ## name, # name }, |
113 | } ivstr; |
195 | } ivstr; |
|
|
196 | |
|
|
197 | typedef struct |
|
|
198 | { |
|
|
199 | NV nv; |
|
|
200 | const char *name; |
|
|
201 | #define const_nv(name) { (NV)CL_ ## name, # name }, |
|
|
202 | } nvstr; |
114 | |
203 | |
115 | static const char * |
204 | static const char * |
116 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
205 | iv2str (IV value, const ivstr *base, int count, const char *fallback) |
117 | { |
206 | { |
118 | int i; |
207 | int i; |
… | |
… | |
167 | if (res) \ |
256 | if (res) \ |
168 | FAIL (name); |
257 | FAIL (name); |
169 | |
258 | |
170 | /*****************************************************************************/ |
259 | /*****************************************************************************/ |
171 | |
260 | |
172 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
173 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
174 | |
|
|
175 | static void * |
261 | static SV * |
|
|
262 | new_clobj (HV *stash, IV id) |
|
|
263 | { |
|
|
264 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
265 | } |
|
|
266 | |
|
|
267 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
268 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
269 | |
|
|
270 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
271 | /* they can be upgraded at runtime to the array form */ |
|
|
272 | static void * ecb_noinline |
176 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
273 | SvCLOBJ (CV *cv, const char *svname, SV *sv, const char *pkg) |
177 | { |
274 | { |
|
|
275 | // sv_derived_from is quite slow :( |
178 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
276 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
179 | return (void *)SvIV (SvRV (sv)); |
277 | return (void *)SvIV (SvRV (sv)); |
180 | |
278 | |
181 | croak ("%s: %s is not of type %s", func, svname, pkg); |
279 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, pkg); |
|
|
280 | } |
|
|
281 | |
|
|
282 | // the "no-inherit" version of the above |
|
|
283 | static void * ecb_noinline |
|
|
284 | SvCLOBJ_ni (CV *cv, const char *svname, SV *sv, HV *stash) |
|
|
285 | { |
|
|
286 | if (SvROK (sv) && SvSTASH (SvRV (sv)) == stash) |
|
|
287 | return (void *)SvIV (SvRV (sv)); |
|
|
288 | |
|
|
289 | croak ("%s: %s is not of type %s", cv_get_name (cv), svname, HvNAME (stash)); |
182 | } |
290 | } |
183 | |
291 | |
184 | /*****************************************************************************/ |
292 | /*****************************************************************************/ |
185 | |
293 | |
186 | static cl_context_properties * |
294 | static cl_context_properties * ecb_noinline |
187 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
295 | SvCONTEXTPROPERTIES (CV *cv, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
188 | { |
296 | { |
189 | if (!sv || !SvOK (sv)) |
297 | if (!sv || !SvOK (sv)) |
190 | if (extra) |
298 | if (extra) |
191 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
299 | sv = sv_2mortal (newRV_noinc ((SV *)newAV ())); // slow, but rarely used hopefully |
192 | else |
300 | else |
… | |
… | |
198 | int i, len = av_len (av) + 1; |
306 | int i, len = av_len (av) + 1; |
199 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
307 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
200 | cl_context_properties *l = p; |
308 | cl_context_properties *l = p; |
201 | |
309 | |
202 | if (len & 1) |
310 | if (len & 1) |
203 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
311 | croak ("%s: %s is not a property list (must contain an even number of elements)", cv_get_name (cv), svname); |
204 | |
312 | |
205 | while (extracount--) |
313 | while (extracount--) |
206 | *l++ = *extra++; |
314 | *l++ = *extra++; |
207 | |
315 | |
208 | for (i = 0; i < len; i += 2) |
316 | for (i = 0; i < len; i += 2) |
… | |
… | |
211 | SV *p_sv = *av_fetch (av, i + 1, 0); |
319 | SV *p_sv = *av_fetch (av, i + 1, 0); |
212 | cl_context_properties v = SvIV (p_sv); // code below can override |
320 | cl_context_properties v = SvIV (p_sv); // code below can override |
213 | |
321 | |
214 | switch (t) |
322 | switch (t) |
215 | { |
323 | { |
|
|
324 | case CL_CONTEXT_PLATFORM: |
|
|
325 | if (SvROK (p_sv)) |
|
|
326 | v = (cl_context_properties)SvCLOBJ (cv, svname, p_sv, "OpenCL::Platform"); |
|
|
327 | break; |
|
|
328 | |
216 | case CL_GLX_DISPLAY_KHR: |
329 | case CL_GLX_DISPLAY_KHR: |
217 | if (!SvOK (p_sv)) |
330 | if (!SvOK (p_sv)) |
218 | { |
331 | { |
219 | void *func = glsym ("_glXGetCurrentDisplay"); |
332 | void *func = glsym ("_glXGetCurrentDisplay"); |
220 | if (func) |
333 | if (func) |
… | |
… | |
243 | *l = 0; |
356 | *l = 0; |
244 | |
357 | |
245 | return p; |
358 | return p; |
246 | } |
359 | } |
247 | |
360 | |
248 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
361 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", cv_get_name (cv), svname); |
|
|
362 | } |
|
|
363 | |
|
|
364 | // parse an array of CLOBJ into a void ** array in C - works only for CLOBJs whose representation |
|
|
365 | // is a pointer (and only on well-behaved systems). |
|
|
366 | static void * ecb_noinline |
|
|
367 | object_list (CV *cv, int or_undef, const char *argname, SV *arg, const char *klass, cl_uint *rcount) |
|
|
368 | { |
|
|
369 | if (!SvROK (arg) || SvTYPE (SvRV (arg)) != SVt_PVAV) |
|
|
370 | croak ("%s: '%s' parameter must be %sa reference to an array of %s objects", |
|
|
371 | cv_get_name (cv), argname, or_undef ? "undef or " : "", klass); |
|
|
372 | |
|
|
373 | AV *av = (AV *)SvRV (arg); |
|
|
374 | void **list = 0; |
|
|
375 | cl_uint count = av_len (av) + 1; |
|
|
376 | |
|
|
377 | if (count) |
|
|
378 | { |
|
|
379 | list = tmpbuf (sizeof (*list) * count); |
|
|
380 | int i; |
|
|
381 | for (i = 0; i < count; ++i) |
|
|
382 | list [i] = SvCLOBJ (cv, argname, *av_fetch (av, i, 1), klass); |
|
|
383 | } |
|
|
384 | |
|
|
385 | if (!count && !or_undef) |
|
|
386 | croak ("%s: '%s' must contain at least one %s object", |
|
|
387 | cv_get_name (cv), argname, klass); |
|
|
388 | |
|
|
389 | *rcount = count; |
|
|
390 | return (void *)list; |
249 | } |
391 | } |
250 | |
392 | |
251 | /*****************************************************************************/ |
393 | /*****************************************************************************/ |
|
|
394 | /* callback stuff */ |
|
|
395 | |
|
|
396 | /* default context callback, log to stderr */ |
|
|
397 | static void CL_CALLBACK |
|
|
398 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
399 | { |
|
|
400 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
401 | } |
|
|
402 | |
|
|
403 | typedef struct |
|
|
404 | { |
|
|
405 | int free_cb; |
|
|
406 | void (*push)(void *data1, void *data2, void *data3); |
|
|
407 | } eq_vtbl; |
|
|
408 | |
|
|
409 | typedef struct eq_item |
|
|
410 | { |
|
|
411 | struct eq_item *next; |
|
|
412 | eq_vtbl *vtbl; |
|
|
413 | SV *cb; |
|
|
414 | void *data1, *data2, *data3; |
|
|
415 | } eq_item; |
|
|
416 | |
|
|
417 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
418 | static void *eq_signal_arg; |
|
|
419 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
420 | static eq_item *eq_head, *eq_tail; |
|
|
421 | |
|
|
422 | static void ecb_noinline |
|
|
423 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
424 | { |
|
|
425 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
426 | |
|
|
427 | item->next = 0; |
|
|
428 | item->vtbl = vtbl; |
|
|
429 | item->cb = cb; |
|
|
430 | item->data1 = data1; |
|
|
431 | item->data2 = data2; |
|
|
432 | item->data3 = data3; |
|
|
433 | |
|
|
434 | X_LOCK (eq_lock); |
|
|
435 | |
|
|
436 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
437 | eq_tail = item; |
|
|
438 | |
|
|
439 | X_UNLOCK (eq_lock); |
|
|
440 | |
|
|
441 | eq_signal_func (eq_signal_arg, 0); |
|
|
442 | } |
|
|
443 | |
|
|
444 | static eq_item * |
|
|
445 | eq_dec (void) |
|
|
446 | { |
|
|
447 | eq_item *res; |
|
|
448 | |
|
|
449 | X_LOCK (eq_lock); |
|
|
450 | |
|
|
451 | res = eq_head; |
|
|
452 | if (res) |
|
|
453 | eq_head = res->next; |
|
|
454 | |
|
|
455 | X_UNLOCK (eq_lock); |
|
|
456 | |
|
|
457 | return res; |
|
|
458 | } |
|
|
459 | |
|
|
460 | static void |
|
|
461 | eq_poll (void) |
|
|
462 | { |
|
|
463 | eq_item *item; |
|
|
464 | |
|
|
465 | while ((item = eq_dec ())) |
|
|
466 | { |
|
|
467 | ENTER; |
|
|
468 | SAVETMPS; |
|
|
469 | |
|
|
470 | dSP; |
|
|
471 | PUSHMARK (SP); |
|
|
472 | EXTEND (SP, 2); |
|
|
473 | |
|
|
474 | if (item->vtbl->free_cb) |
|
|
475 | sv_2mortal (item->cb); |
|
|
476 | |
|
|
477 | PUTBACK; |
|
|
478 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
479 | |
|
|
480 | SV *cb = item->cb; |
|
|
481 | free (item); |
|
|
482 | |
|
|
483 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
484 | |
|
|
485 | FREETMPS; |
|
|
486 | LEAVE; |
|
|
487 | } |
|
|
488 | } |
|
|
489 | |
|
|
490 | static void |
|
|
491 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
492 | { |
|
|
493 | eq_poll (); |
|
|
494 | } |
|
|
495 | |
|
|
496 | /*****************************************************************************/ |
|
|
497 | /* context notify */ |
|
|
498 | |
|
|
499 | static void ecb_noinline |
|
|
500 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
501 | { |
|
|
502 | dSP; |
|
|
503 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
504 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
505 | PUTBACK; |
|
|
506 | |
|
|
507 | free (data1); |
|
|
508 | free (data2); |
|
|
509 | } |
|
|
510 | |
|
|
511 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
512 | |
|
|
513 | static void CL_CALLBACK |
|
|
514 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
515 | { |
|
|
516 | void *pvt_copy = malloc (cb); |
|
|
517 | memcpy (pvt_copy, pvt, cb); |
|
|
518 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
519 | } |
|
|
520 | |
|
|
521 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
522 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
523 | void *user_data = 0; \ |
|
|
524 | \ |
|
|
525 | if (SvOK (notify)) \ |
|
|
526 | { \ |
|
|
527 | pfn_notify = eq_context_notify; \ |
|
|
528 | user_data = s_get_cv (notify); \ |
|
|
529 | } |
|
|
530 | |
|
|
531 | static SV * ecb_noinline |
|
|
532 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
533 | { |
|
|
534 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
535 | |
|
|
536 | if (user_data) |
|
|
537 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
538 | |
|
|
539 | return sv; |
|
|
540 | } |
|
|
541 | |
|
|
542 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
543 | |
|
|
544 | /*****************************************************************************/ |
|
|
545 | /* build/compile/link notify */ |
|
|
546 | |
|
|
547 | static void |
|
|
548 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
549 | { |
|
|
550 | dSP; |
|
|
551 | PUSH_CLOBJ (stash_program, data1); |
|
|
552 | PUTBACK; |
|
|
553 | } |
|
|
554 | |
|
|
555 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
556 | |
|
|
557 | static void CL_CALLBACK |
|
|
558 | eq_program_notify (cl_program program, void *user_data) |
|
|
559 | { |
|
|
560 | clRetainProgram (program); |
|
|
561 | |
|
|
562 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
563 | } |
|
|
564 | |
|
|
565 | typedef void (CL_CALLBACK *program_callback)(cl_program program, void *user_data); |
|
|
566 | |
|
|
567 | static program_callback ecb_noinline |
|
|
568 | make_program_callback (SV *notify, void **ruser_data) |
|
|
569 | { |
|
|
570 | if (SvOK (notify)) |
|
|
571 | { |
|
|
572 | *ruser_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
573 | return eq_program_notify; |
|
|
574 | } |
|
|
575 | else |
|
|
576 | { |
|
|
577 | *ruser_data = 0; |
|
|
578 | return 0; |
|
|
579 | } |
|
|
580 | } |
|
|
581 | |
|
|
582 | struct build_args |
|
|
583 | { |
|
|
584 | cl_program program; |
|
|
585 | char *options; |
|
|
586 | void *user_data; |
|
|
587 | cl_uint num_devices; |
|
|
588 | }; |
|
|
589 | |
|
|
590 | X_THREAD_PROC (build_program_thread) |
|
|
591 | { |
|
|
592 | struct build_args *arg = thr_arg; |
|
|
593 | |
|
|
594 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
595 | |
|
|
596 | if (arg->user_data) |
|
|
597 | eq_program_notify (arg->program, arg->user_data); |
|
|
598 | else |
|
|
599 | clReleaseProgram (arg->program); |
|
|
600 | |
|
|
601 | free (arg->options); |
|
|
602 | free (arg); |
|
|
603 | |
|
|
604 | return 0; |
|
|
605 | } |
|
|
606 | |
|
|
607 | static void |
|
|
608 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
609 | { |
|
|
610 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
611 | |
|
|
612 | arg->program = program; |
|
|
613 | arg->options = strdup (options); |
|
|
614 | arg->user_data = user_data; |
|
|
615 | arg->num_devices = num_devices; |
|
|
616 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
617 | |
|
|
618 | xthread_t id; |
|
|
619 | thread_create (&id, build_program_thread, arg); |
|
|
620 | } |
|
|
621 | |
|
|
622 | /*****************************************************************************/ |
|
|
623 | /* event objects */ |
|
|
624 | |
|
|
625 | static void |
|
|
626 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
627 | { |
|
|
628 | dSP; |
|
|
629 | PUSH_CLOBJ (stash_event, data1); |
|
|
630 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
631 | PUTBACK; |
|
|
632 | } |
|
|
633 | |
|
|
634 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
635 | |
|
|
636 | static void CL_CALLBACK |
|
|
637 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
638 | { |
|
|
639 | clRetainEvent (event); |
|
|
640 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
641 | } |
|
|
642 | |
|
|
643 | /*****************************************************************************/ |
|
|
644 | /* utilities for XS code */ |
252 | |
645 | |
253 | static size_t |
646 | static size_t |
254 | img_row_pitch (cl_mem img) |
647 | img_row_pitch (cl_mem img) |
255 | { |
648 | { |
256 | size_t res; |
649 | size_t res; |
257 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
650 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
258 | return res; |
651 | return res; |
259 | } |
652 | } |
260 | |
653 | |
261 | static cl_event * |
654 | static cl_event * ecb_noinline |
262 | event_list (SV **items, int count) |
655 | event_list (CV *cv, SV **items, cl_uint *rcount, cl_event extra) |
263 | { |
656 | { |
264 | if (!count) |
657 | cl_uint count = *rcount; |
|
|
658 | |
|
|
659 | if (count > 0x7fffffffU) // yeah, it's a hack - the caller might have underflowed |
|
|
660 | *rcount = count = 0; |
|
|
661 | |
|
|
662 | if (!count && !extra) |
265 | return 0; |
663 | return 0; |
266 | |
664 | |
267 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
665 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
666 | int i = 0; |
268 | |
667 | |
269 | while (count--) |
668 | while (count--) |
|
|
669 | if (SvOK (items [count])) |
270 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
670 | list [i++] = SvCLOBJ (cv, "wait_events", items [count], "OpenCL::Event"); |
271 | |
671 | |
|
|
672 | if (extra) |
|
|
673 | list [i++] = extra; |
|
|
674 | |
|
|
675 | *rcount = i; |
|
|
676 | |
272 | return list; |
677 | return i ? list : 0; |
273 | } |
678 | } |
274 | |
679 | |
275 | #define EVENT_LIST(items,count) \ |
680 | #define EVENT_LIST(skip) \ |
276 | cl_uint event_list_count = (count); \ |
681 | cl_uint event_list_count = items - (skip); \ |
277 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
682 | cl_event *event_list_ptr = event_list (cv, &ST (skip), &event_list_count, 0) |
278 | |
683 | |
279 | #define INFO(class) \ |
684 | #define INFO(class) \ |
280 | { \ |
685 | { \ |
281 | size_t size; \ |
686 | size_t size; \ |
282 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
687 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
286 | SvCUR_set (sv, size); \ |
691 | SvCUR_set (sv, size); \ |
287 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
692 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
288 | XPUSHs (sv); \ |
693 | XPUSHs (sv); \ |
289 | } |
694 | } |
290 | |
695 | |
|
|
696 | /*****************************************************************************/ |
|
|
697 | /* mapped_xxx */ |
|
|
698 | |
|
|
699 | static OpenCL__Mapped |
|
|
700 | SvMAPPED (SV *sv) |
|
|
701 | { |
|
|
702 | // no typechecking atm., keep your fingers crossed |
|
|
703 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
704 | } |
|
|
705 | |
|
|
706 | struct mapped |
|
|
707 | { |
|
|
708 | cl_command_queue queue; |
|
|
709 | cl_mem memobj; |
|
|
710 | void *ptr; |
|
|
711 | size_t cb; |
|
|
712 | cl_event event; |
|
|
713 | size_t row_pitch; |
|
|
714 | size_t slice_pitch; |
|
|
715 | }; |
|
|
716 | |
|
|
717 | static SV * |
|
|
718 | 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) |
|
|
719 | { |
|
|
720 | SV *data = newSV (0); |
|
|
721 | SvUPGRADE (data, SVt_PVMG); |
|
|
722 | |
|
|
723 | OpenCL__Mapped mapped; |
|
|
724 | New (0, mapped, 1, struct mapped); |
|
|
725 | |
|
|
726 | clRetainCommandQueue (queue); |
|
|
727 | |
|
|
728 | mapped->queue = queue; |
|
|
729 | mapped->memobj = memobj; |
|
|
730 | mapped->ptr = ptr; |
|
|
731 | mapped->cb = cb; |
|
|
732 | mapped->event = ev; |
|
|
733 | mapped->row_pitch = row_pitch; |
|
|
734 | mapped->slice_pitch = slice_pitch; |
|
|
735 | |
|
|
736 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
737 | |
|
|
738 | if (SvLEN (data)) |
|
|
739 | Safefree (data); |
|
|
740 | |
|
|
741 | SvPVX (data) = (char *)ptr; |
|
|
742 | SvCUR_set (data, cb); |
|
|
743 | SvLEN_set (data, 0); |
|
|
744 | SvPOK_only (data); |
|
|
745 | |
|
|
746 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
747 | |
|
|
748 | if (!(flags & CL_MAP_WRITE)) |
|
|
749 | SvREADONLY_on (data); |
|
|
750 | |
|
|
751 | return obj; |
|
|
752 | } |
|
|
753 | |
|
|
754 | static void |
|
|
755 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
756 | { |
|
|
757 | SV *data = SvRV (sv); |
|
|
758 | |
|
|
759 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
760 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
761 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
762 | else |
|
|
763 | { |
|
|
764 | SvREADONLY_off (data); |
|
|
765 | SvCUR_set (data, 0); |
|
|
766 | SvPVX (data) = 0; |
|
|
767 | SvOK_off (data); |
|
|
768 | } |
|
|
769 | |
|
|
770 | mapped->ptr = 0; |
|
|
771 | } |
|
|
772 | |
|
|
773 | static void |
|
|
774 | mapped_unmap (CV *cv, SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
775 | { |
|
|
776 | cl_event *event_list_ptr = event_list (cv, wait_list, &event_list_count, mapped->event); |
|
|
777 | cl_event ev; |
|
|
778 | |
|
|
779 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
780 | |
|
|
781 | clReleaseEvent (mapped->event); |
|
|
782 | mapped->event = ev; |
|
|
783 | |
|
|
784 | mapped_detach (self, mapped); |
|
|
785 | } |
|
|
786 | |
|
|
787 | /*****************************************************************************/ |
|
|
788 | |
291 | MODULE = OpenCL PACKAGE = OpenCL |
789 | MODULE = OpenCL PACKAGE = OpenCL |
292 | |
790 | |
293 | PROTOTYPES: ENABLE |
791 | PROTOTYPES: ENABLE |
294 | |
792 | |
|
|
793 | void |
|
|
794 | poll () |
|
|
795 | CODE: |
|
|
796 | eq_poll (); |
|
|
797 | |
|
|
798 | void |
|
|
799 | _eq_initialise (IV func, IV arg) |
|
|
800 | CODE: |
|
|
801 | eq_signal_func = (void (*)(void *, int))func; |
|
|
802 | eq_signal_arg = (void*)arg; |
|
|
803 | |
295 | BOOT: |
804 | BOOT: |
296 | { |
805 | { |
297 | HV *stash = gv_stashpv ("OpenCL", 1); |
806 | HV *stash = gv_stashpv ("OpenCL", 1); |
|
|
807 | |
298 | static const ivstr *civ, const_iv[] = { |
808 | static const ivstr *civ, const_iv[] = { |
299 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
809 | { sizeof (cl_char ), "SIZEOF_CHAR" }, |
300 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
810 | { sizeof (cl_uchar ), "SIZEOF_UCHAR" }, |
301 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
811 | { sizeof (cl_short ), "SIZEOF_SHORT" }, |
302 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
812 | { sizeof (cl_ushort), "SIZEOF_USHORT" }, |
… | |
… | |
305 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
815 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
306 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
816 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
307 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
817 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
308 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
818 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
309 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
819 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
|
|
820 | { PREFER_1_1 , "PREFER_1_1" }, |
310 | #include "constiv.h" |
821 | #include "constiv.h" |
311 | }; |
822 | }; |
|
|
823 | |
312 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
824 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
313 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
825 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
826 | |
|
|
827 | static const nvstr *cnv, const_nv[] = { |
|
|
828 | #include "constnv.h" |
|
|
829 | }; |
|
|
830 | |
|
|
831 | for (cnv = const_nv + sizeof (const_nv) / sizeof (const_nv [0]); cnv > const_nv; cnv--) |
|
|
832 | newCONSTSUB (stash, (char *)cnv[-1].name, newSVnv (cnv[-1].nv)); |
|
|
833 | |
|
|
834 | newCONSTSUB (stash, "NAN", newSVnv (CL_NAN)); // CL_NAN might be a function call |
|
|
835 | |
|
|
836 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
837 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
838 | stash_subdevice = gv_stashpv ("OpenCL::SubDevice", GV_ADD); |
|
|
839 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
840 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
841 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
842 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
843 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
844 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
845 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
846 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
847 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
848 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
849 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
850 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
851 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
852 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
853 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
854 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
855 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
856 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
857 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
858 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
859 | |
|
|
860 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
314 | } |
861 | } |
315 | |
862 | |
316 | cl_int |
863 | cl_int |
317 | errno () |
864 | errno () |
318 | CODE: |
865 | CODE: |
319 | errno = res; |
866 | RETVAL = res; |
|
|
867 | OUTPUT: |
|
|
868 | RETVAL |
320 | |
869 | |
321 | const char * |
870 | const char * |
322 | err2str (cl_int err) |
871 | err2str (cl_int err = res) |
323 | |
872 | |
324 | const char * |
873 | const char * |
325 | enum2str (cl_uint value) |
874 | enum2str (cl_uint value) |
326 | |
875 | |
327 | void |
876 | void |
… | |
… | |
335 | list = tmpbuf (sizeof (*list) * count); |
884 | list = tmpbuf (sizeof (*list) * count); |
336 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
885 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
337 | |
886 | |
338 | EXTEND (SP, count); |
887 | EXTEND (SP, count); |
339 | for (i = 0; i < count; ++i) |
888 | for (i = 0; i < count; ++i) |
340 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
889 | PUSH_CLOBJ (stash_platform, list [i]); |
341 | |
890 | |
342 | void |
891 | void |
343 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
892 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
344 | PPCODE: |
893 | PPCODE: |
|
|
894 | CONTEXT_NOTIFY_CALLBACK; |
345 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
895 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, pfn_notify, user_data, &res)); |
346 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
896 | XPUSH_CLOBJ_CONTEXT; |
347 | |
897 | |
348 | void |
898 | void |
349 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
899 | context (cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
350 | PPCODE: |
900 | PPCODE: |
351 | /* der Gipfel der Kunst */ |
901 | cl_uint device_count; |
|
|
902 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
903 | |
|
|
904 | CONTEXT_NOTIFY_CALLBACK; |
|
|
905 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
906 | XPUSH_CLOBJ_CONTEXT; |
352 | |
907 | |
353 | void |
908 | void |
354 | wait_for_events (...) |
909 | wait_for_events (...) |
355 | CODE: |
910 | CODE: |
356 | EVENT_LIST (0, items); |
911 | EVENT_LIST (0); |
357 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
912 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
358 | |
913 | |
359 | PROTOTYPES: DISABLE |
914 | PROTOTYPES: DISABLE |
360 | |
915 | |
361 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
916 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
362 | |
917 | |
363 | void |
918 | void |
364 | info (OpenCL::Platform self, cl_platform_info name) |
919 | info (OpenCL::Platform self, cl_platform_info name) |
365 | PPCODE: |
920 | PPCODE: |
366 | INFO (Platform) |
921 | INFO (Platform) |
|
|
922 | |
|
|
923 | void |
|
|
924 | unload_compiler (OpenCL::Platform self) |
|
|
925 | CODE: |
|
|
926 | #if CL_VERSION_1_2 |
|
|
927 | clUnloadPlatformCompiler (self); |
|
|
928 | #endif |
367 | |
929 | |
368 | #BEGIN:platform |
930 | #BEGIN:platform |
369 | |
931 | |
370 | void |
932 | void |
371 | profile (OpenCL::Platform self) |
933 | profile (OpenCL::Platform self) |
… | |
… | |
397 | list = tmpbuf (sizeof (*list) * count); |
959 | list = tmpbuf (sizeof (*list) * count); |
398 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
960 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
399 | |
961 | |
400 | EXTEND (SP, count); |
962 | EXTEND (SP, count); |
401 | for (i = 0; i < count; ++i) |
963 | for (i = 0; i < count; ++i) |
402 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
964 | PUSH_CLOBJ (stash_device, list [i]); |
403 | |
965 | |
404 | void |
966 | void |
405 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
967 | context (OpenCL::Platform self, SV *properties, SV *devices, SV *notify = &PL_sv_undef) |
406 | PPCODE: |
968 | PPCODE: |
|
|
969 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
970 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
971 | |
|
|
972 | cl_uint device_count; |
|
|
973 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
974 | |
|
|
975 | CONTEXT_NOTIFY_CALLBACK; |
|
|
976 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (props, device_count, device_list, pfn_notify, user_data, &res)); |
|
|
977 | XPUSH_CLOBJ_CONTEXT; |
|
|
978 | |
|
|
979 | void |
|
|
980 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
|
|
981 | PPCODE: |
|
|
982 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
983 | cl_context_properties *props = SvCONTEXTPROPERTIES (cv, "properties", properties, extra, 2); |
|
|
984 | |
|
|
985 | CONTEXT_NOTIFY_CALLBACK; |
|
|
986 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, pfn_notify, user_data, &res)); |
|
|
987 | XPUSH_CLOBJ_CONTEXT; |
|
|
988 | |
|
|
989 | MODULE = OpenCL PACKAGE = OpenCL::Device |
|
|
990 | |
|
|
991 | void |
|
|
992 | info (OpenCL::Device self, cl_device_info name) |
|
|
993 | PPCODE: |
|
|
994 | INFO (Device) |
|
|
995 | |
|
|
996 | #if CL_VERSION_1_2 |
|
|
997 | |
|
|
998 | void |
|
|
999 | sub_devices (OpenCL::Device self, SV *properties) |
|
|
1000 | PPCODE: |
407 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
1001 | if (!SvROK (properties) || SvTYPE (SvRV (properties)) != SVt_PVAV) |
408 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
1002 | croak ("OpenCL::Device::sub_devices: properties must be specified as reference to an array of property-value pairs"); |
409 | |
1003 | |
410 | AV *av = (AV *)SvRV (devices); |
1004 | properties = SvRV (properties); |
411 | cl_uint num_devices = av_len (av) + 1; |
1005 | |
412 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
1006 | cl_uint count = av_len ((AV *)properties) + 1; |
|
|
1007 | cl_device_partition_property *props = tmpbuf (sizeof (*props) * count + 1); |
|
|
1008 | |
413 | int i; |
1009 | int i; |
|
|
1010 | for (i = 0; i < count; ++i) |
|
|
1011 | props [i] = (cl_device_partition_property)SvIV (*av_fetch ((AV *)properties, i, 0)); |
414 | |
1012 | |
415 | for (i = num_devices; i--; ) |
1013 | props [count] = 0; |
416 | device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
|
|
417 | |
1014 | |
418 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res)); |
1015 | cl_uint num_devices; |
419 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
1016 | NEED_SUCCESS (CreateSubDevices, (self, props, 0, 0, &num_devices)); |
|
|
1017 | cl_device_id *list = tmpbuf (sizeof (*list) * num_devices); |
|
|
1018 | NEED_SUCCESS (CreateSubDevices, (self, props, num_devices, list, 0)); |
420 | |
1019 | |
421 | void |
1020 | EXTEND (SP, num_devices); |
422 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
1021 | for (i = 0; i < count; ++i) |
423 | PPCODE: |
1022 | PUSH_CLOBJ (stash_subdevice, list [i]); |
424 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
|
|
425 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
426 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
|
|
427 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
|
|
428 | |
1023 | |
429 | MODULE = OpenCL PACKAGE = OpenCL::Device |
1024 | #endif |
430 | |
|
|
431 | void |
|
|
432 | info (OpenCL::Device self, cl_device_info name) |
|
|
433 | PPCODE: |
|
|
434 | INFO (Device) |
|
|
435 | |
1025 | |
436 | #BEGIN:device |
1026 | #BEGIN:device |
437 | |
1027 | |
438 | void |
1028 | void |
439 | type (OpenCL::Device self) |
1029 | type (OpenCL::Device self) |
440 | PPCODE: |
1030 | PPCODE: |
441 | cl_device_type value [1]; |
1031 | cl_device_type value [1]; |
442 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
1032 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
443 | EXTEND (SP, 1); |
1033 | EXTEND (SP, 1); |
444 | const int i = 0; |
1034 | const int i = 0; |
445 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1035 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
446 | |
1036 | |
447 | void |
1037 | void |
448 | vendor_id (OpenCL::Device self) |
1038 | vendor_id (OpenCL::Device self) |
449 | ALIAS: |
1039 | ALIAS: |
450 | vendor_id = CL_DEVICE_VENDOR_ID |
1040 | vendor_id = CL_DEVICE_VENDOR_ID |
… | |
… | |
471 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
1061 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
472 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
1062 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
473 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
1063 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
474 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
1064 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
475 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
1065 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
476 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
1066 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
477 | PPCODE: |
1067 | PPCODE: |
478 | cl_uint value [1]; |
1068 | cl_uint value [1]; |
479 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
1069 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
480 | EXTEND (SP, 1); |
1070 | EXTEND (SP, 1); |
481 | const int i = 0; |
1071 | const int i = 0; |
… | |
… | |
595 | PPCODE: |
1185 | PPCODE: |
596 | cl_command_queue_properties value [1]; |
1186 | cl_command_queue_properties value [1]; |
597 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1187 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
598 | EXTEND (SP, 1); |
1188 | EXTEND (SP, 1); |
599 | const int i = 0; |
1189 | const int i = 0; |
600 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1190 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
601 | |
1191 | |
602 | void |
1192 | void |
603 | platform (OpenCL::Device self) |
1193 | platform (OpenCL::Device self) |
604 | PPCODE: |
1194 | PPCODE: |
605 | cl_platform_id value [1]; |
1195 | cl_platform_id value [1]; |
606 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1196 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
607 | EXTEND (SP, 1); |
1197 | EXTEND (SP, 1); |
608 | const int i = 0; |
1198 | const int i = 0; |
609 | { |
1199 | { |
610 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); |
1200 | PUSH_CLOBJ (stash_platform, value [i]); |
611 | } |
1201 | } |
612 | |
1202 | |
613 | void |
1203 | void |
614 | name (OpenCL::Device self) |
1204 | name (OpenCL::Device self) |
615 | ALIAS: |
1205 | ALIAS: |
… | |
… | |
634 | cl_device_id value [1]; |
1224 | cl_device_id value [1]; |
635 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1225 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
636 | EXTEND (SP, 1); |
1226 | EXTEND (SP, 1); |
637 | const int i = 0; |
1227 | const int i = 0; |
638 | { |
1228 | { |
639 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1229 | PUSH_CLOBJ (stash_device, value [i]); |
640 | } |
1230 | } |
641 | |
1231 | |
642 | void |
1232 | void |
643 | partition_types_ext (OpenCL::Device self) |
1233 | partition_types_ext (OpenCL::Device self) |
644 | ALIAS: |
1234 | ALIAS: |
… | |
… | |
655 | for (i = 0; i < n; ++i) |
1245 | for (i = 0; i < n; ++i) |
656 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1246 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
657 | |
1247 | |
658 | #END:device |
1248 | #END:device |
659 | |
1249 | |
|
|
1250 | MODULE = OpenCL PACKAGE = OpenCL::SubDevice |
|
|
1251 | |
|
|
1252 | #if CL_VERSION_1_2 |
|
|
1253 | |
|
|
1254 | void |
|
|
1255 | DESTROY (OpenCL::SubDevice self) |
|
|
1256 | CODE: |
|
|
1257 | clReleaseDevice (self); |
|
|
1258 | |
|
|
1259 | #endif |
|
|
1260 | |
660 | MODULE = OpenCL PACKAGE = OpenCL::Context |
1261 | MODULE = OpenCL PACKAGE = OpenCL::Context |
661 | |
1262 | |
662 | void |
1263 | void |
663 | DESTROY (OpenCL::Context context) |
1264 | DESTROY (OpenCL::Context self) |
664 | CODE: |
1265 | CODE: |
665 | clReleaseContext (context); |
1266 | clReleaseContext (self); |
666 | |
1267 | |
667 | void |
1268 | void |
668 | info (OpenCL::Context self, cl_context_info name) |
1269 | info (OpenCL::Context self, cl_context_info name) |
669 | PPCODE: |
1270 | PPCODE: |
670 | INFO (Context) |
1271 | INFO (Context) |
671 | |
1272 | |
672 | void |
1273 | void |
673 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
1274 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
674 | PPCODE: |
1275 | PPCODE: |
675 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
1276 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
676 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
1277 | XPUSH_CLOBJ (stash_queue, queue); |
677 | |
1278 | |
678 | void |
1279 | void |
679 | user_event (OpenCL::Context self) |
1280 | user_event (OpenCL::Context self) |
680 | PPCODE: |
1281 | PPCODE: |
681 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
1282 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
682 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
1283 | XPUSH_CLOBJ (stash_userevent, ev); |
683 | |
1284 | |
684 | void |
1285 | void |
685 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
1286 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
686 | PPCODE: |
1287 | PPCODE: |
687 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1288 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
688 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1289 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
689 | |
1290 | |
690 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
1291 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
691 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1292 | XPUSH_CLOBJ (stash_bufferobj, mem); |
692 | |
1293 | |
693 | void |
1294 | void |
694 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
1295 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
695 | PPCODE: |
1296 | PPCODE: |
696 | STRLEN len; |
1297 | STRLEN len; |
697 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1298 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
698 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1299 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
699 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1300 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
700 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
1301 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
701 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1302 | XPUSH_CLOBJ (stash_bufferobj, mem); |
702 | |
1303 | |
|
|
1304 | #if CL_VERSION_1_2 |
|
|
1305 | |
703 | void |
1306 | void |
704 | image2d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t row_pitch = 0, SV *data = &PL_sv_undef) |
1307 | image (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, cl_mem_object_type type, size_t width, size_t height, size_t depth = 0, size_t array_size = 0, size_t row_pitch = 0, size_t slice_pitch = 0, cl_uint num_mip_level = 0, cl_uint num_samples = 0, SV *data = &PL_sv_undef) |
705 | PPCODE: |
1308 | PPCODE: |
706 | STRLEN len; |
1309 | STRLEN len; |
707 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1310 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
708 | const cl_image_format format = { channel_order, channel_type }; |
1311 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1312 | const cl_image_desc desc = { |
|
|
1313 | type, |
|
|
1314 | width, height, depth, |
|
|
1315 | array_size, row_pitch, slice_pitch, |
|
|
1316 | num_mip_level, num_samples, |
|
|
1317 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ (cv, "data", data, "OpenCL::Buffer") : 0 |
|
|
1318 | }; |
709 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
1319 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
710 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1320 | HV *stash = stash_image; |
|
|
1321 | switch (type) |
|
|
1322 | { |
|
|
1323 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1324 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1325 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1326 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1327 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1328 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1329 | } |
|
|
1330 | XPUSH_CLOBJ (stash, mem); |
711 | |
1331 | |
|
|
1332 | #endif |
|
|
1333 | |
712 | void |
1334 | void |
713 | image3d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t depth, size_t row_pitch = 0, size_t slice_pitch = 0, SV *data = &PL_sv_undef) |
1335 | image2d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t row_pitch = 0, SV *data = &PL_sv_undef) |
714 | PPCODE: |
1336 | PPCODE: |
715 | STRLEN len; |
1337 | STRLEN len; |
716 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1338 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
717 | const cl_image_format format = { channel_order, channel_type }; |
1339 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1340 | #if PREFER_1_1 |
|
|
1341 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1342 | #else |
|
|
1343 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1344 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1345 | #endif |
|
|
1346 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1347 | |
|
|
1348 | void |
|
|
1349 | image3d (OpenCL::Context self, cl_mem_flags flags, cl_channel_order channel_order, cl_channel_type channel_type, size_t width, size_t height, size_t depth, size_t row_pitch = 0, size_t slice_pitch = 0, SV *data = &PL_sv_undef) |
|
|
1350 | PPCODE: |
|
|
1351 | STRLEN len; |
|
|
1352 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1353 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1354 | #if PREFER_1_1 |
718 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
1355 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
719 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1356 | #else |
|
|
1357 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1358 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1359 | #endif |
|
|
1360 | XPUSH_CLOBJ (stash_image3d, mem); |
720 | |
1361 | |
721 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1362 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
722 | |
1363 | |
723 | void |
1364 | void |
724 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
1365 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
725 | PPCODE: |
1366 | PPCODE: |
726 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
1367 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
727 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1368 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1369 | |
|
|
1370 | void |
|
|
1371 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1372 | PPCODE: |
|
|
1373 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1374 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1375 | |
|
|
1376 | #if CL_VERSION_1_2 |
|
|
1377 | |
|
|
1378 | void |
|
|
1379 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1380 | ALIAS: |
|
|
1381 | PPCODE: |
|
|
1382 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1383 | cl_gl_object_type type; |
|
|
1384 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1385 | HV *stash = stash_memory; |
|
|
1386 | switch (type) |
|
|
1387 | { |
|
|
1388 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1389 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1390 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1391 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1392 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1393 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1394 | } |
|
|
1395 | XPUSH_CLOBJ (stash, mem); |
|
|
1396 | |
|
|
1397 | #endif |
728 | |
1398 | |
729 | void |
1399 | void |
730 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1400 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
731 | PPCODE: |
1401 | PPCODE: |
|
|
1402 | #if PREFER_1_1 |
732 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
1403 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
733 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1404 | #else |
|
|
1405 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1406 | #endif |
|
|
1407 | XPUSH_CLOBJ (stash_image2d, mem); |
734 | |
1408 | |
735 | void |
1409 | void |
736 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
1410 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
737 | PPCODE: |
1411 | PPCODE: |
|
|
1412 | #if PREFER_1_1 |
738 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
1413 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
739 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1414 | #else |
740 | |
1415 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
741 | void |
1416 | #endif |
742 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
1417 | XPUSH_CLOBJ (stash_image3d, mem); |
743 | PPCODE: |
|
|
744 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
745 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
|
|
746 | |
1418 | |
747 | #endif |
1419 | #endif |
748 | |
1420 | |
749 | void |
1421 | void |
750 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
1422 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
… | |
… | |
770 | |
1442 | |
771 | void |
1443 | void |
772 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1444 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
773 | PPCODE: |
1445 | PPCODE: |
774 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
1446 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
775 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1447 | XPUSH_CLOBJ (stash_sampler, sampler); |
776 | |
1448 | |
777 | void |
1449 | void |
778 | program_with_source (OpenCL::Context self, SV *program) |
1450 | program_with_source (OpenCL::Context self, SV *program) |
779 | PPCODE: |
1451 | PPCODE: |
780 | STRLEN len; |
1452 | STRLEN len; |
781 | size_t len2; |
1453 | size_t len2; |
782 | const char *ptr = SvPVbyte (program, len); |
1454 | const char *ptr = SvPVbyte (program, len); |
783 | |
1455 | |
784 | len2 = len; |
1456 | len2 = len; |
785 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1457 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
786 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1458 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1459 | |
|
|
1460 | void |
|
|
1461 | program_with_binary (OpenCL::Context self, SV *devices, SV *binaries) |
|
|
1462 | PPCODE: |
|
|
1463 | cl_uint device_count; |
|
|
1464 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1465 | |
|
|
1466 | if (!SvROK (binaries) || SvTYPE (SvRV (binaries)) != SVt_PVAV) |
|
|
1467 | croak ("OpenCL::Context::program_with_binary: binaries must be specified as reference to an array of strings"); |
|
|
1468 | |
|
|
1469 | binaries = SvRV (binaries); |
|
|
1470 | |
|
|
1471 | if (device_count != av_len ((AV *)binaries) + 1) |
|
|
1472 | croak ("OpenCL::Context::program_with_binary: differing numbers of devices and binaries are not allowed"); |
|
|
1473 | |
|
|
1474 | size_t *length_list = tmpbuf (sizeof (*length_list) * device_count); |
|
|
1475 | const unsigned char **binary_list = tmpbuf (sizeof (*binary_list) * device_count); |
|
|
1476 | cl_int *status_list = tmpbuf (sizeof (*status_list) * device_count); |
|
|
1477 | |
|
|
1478 | int i; |
|
|
1479 | for (i = 0; i < device_count; ++i) |
|
|
1480 | { |
|
|
1481 | STRLEN len; |
|
|
1482 | binary_list [i] = (const unsigned char *)SvPVbyte (*av_fetch ((AV *)binaries, i, 0), len); |
|
|
1483 | length_list [i] = len; |
|
|
1484 | } |
|
|
1485 | |
|
|
1486 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBinary, (self, device_count, device_list, |
|
|
1487 | length_list, binary_list, |
|
|
1488 | GIMME_V == G_ARRAY ? status_list : 0, &res)); |
|
|
1489 | |
|
|
1490 | EXTEND (SP, 2); |
|
|
1491 | PUSH_CLOBJ (stash_program, prog); |
|
|
1492 | |
|
|
1493 | if (GIMME_V == G_ARRAY) |
|
|
1494 | { |
|
|
1495 | AV *av = newAV (); |
|
|
1496 | PUSHs (sv_2mortal (newRV_noinc ((SV *)av))); |
|
|
1497 | |
|
|
1498 | for (i = device_count; i--; ) |
|
|
1499 | av_store (av, i, newSViv (status_list [i])); |
|
|
1500 | } |
|
|
1501 | |
|
|
1502 | #if CL_VERSION_1_2 |
|
|
1503 | |
|
|
1504 | void |
|
|
1505 | program_with_built_in_kernels (OpenCL::Context self, SV *devices, SV *kernel_names) |
|
|
1506 | PPCODE: |
|
|
1507 | cl_uint device_count; |
|
|
1508 | cl_device_id *device_list = object_list (cv, 0, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1509 | |
|
|
1510 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithBuiltInKernels, (self, device_count, device_list, SvPVbyte_nolen (kernel_names), &res)); |
|
|
1511 | |
|
|
1512 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1513 | |
|
|
1514 | void |
|
|
1515 | link_program (OpenCL::Context self, SV *devices, SV *options, SV *programs, SV *notify = &PL_sv_undef) |
|
|
1516 | CODE: |
|
|
1517 | cl_uint device_count = 0; |
|
|
1518 | cl_device_id *device_list = 0; |
|
|
1519 | |
|
|
1520 | if (SvOK (devices)) |
|
|
1521 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
1522 | |
|
|
1523 | cl_uint program_count; |
|
|
1524 | cl_program *program_list = object_list (cv, 0, "programs", programs, "OpenCL::Program", &program_count); |
|
|
1525 | |
|
|
1526 | void *user_data; |
|
|
1527 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
1528 | |
|
|
1529 | NEED_SUCCESS_ARG (cl_program prog, LinkProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
1530 | program_count, program_list, pfn_notify, user_data, &res)); |
|
|
1531 | |
|
|
1532 | XPUSH_CLOBJ (stash_program, prog); |
|
|
1533 | |
|
|
1534 | #endif |
787 | |
1535 | |
788 | #BEGIN:context |
1536 | #BEGIN:context |
789 | |
1537 | |
790 | void |
1538 | void |
791 | reference_count (OpenCL::Context self) |
1539 | reference_count (OpenCL::Context self) |
… | |
… | |
808 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1556 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
809 | int i, n = size / sizeof (*value); |
1557 | int i, n = size / sizeof (*value); |
810 | EXTEND (SP, n); |
1558 | EXTEND (SP, n); |
811 | for (i = 0; i < n; ++i) |
1559 | for (i = 0; i < n; ++i) |
812 | { |
1560 | { |
813 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1561 | PUSH_CLOBJ (stash_device, value [i]); |
814 | } |
1562 | } |
815 | |
1563 | |
816 | void |
1564 | void |
817 | properties (OpenCL::Context self) |
1565 | properties (OpenCL::Context self) |
818 | PPCODE: |
1566 | PPCODE: |
… | |
… | |
833 | DESTROY (OpenCL::Queue self) |
1581 | DESTROY (OpenCL::Queue self) |
834 | CODE: |
1582 | CODE: |
835 | clReleaseCommandQueue (self); |
1583 | clReleaseCommandQueue (self); |
836 | |
1584 | |
837 | void |
1585 | void |
838 | enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1586 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1587 | ALIAS: |
|
|
1588 | enqueue_read_buffer = 0 |
839 | PPCODE: |
1589 | PPCODE: |
840 | cl_event ev = 0; |
|
|
841 | EVENT_LIST (6, items - 6); |
1590 | EVENT_LIST (6); |
842 | |
1591 | |
843 | SvUPGRADE (data, SVt_PV); |
1592 | SvUPGRADE (data, SVt_PV); |
844 | SvGROW (data, len); |
1593 | SvGROW (data, len); |
845 | SvPOK_only (data); |
1594 | SvPOK_only (data); |
846 | SvCUR_set (data, len); |
1595 | SvCUR_set (data, len); |
|
|
1596 | |
|
|
1597 | cl_event ev = 0; |
847 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1598 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
848 | |
1599 | |
849 | if (ev) |
1600 | if (ev) |
850 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1601 | XPUSH_CLOBJ (stash_event, ev); |
851 | |
1602 | |
852 | void |
1603 | void |
853 | enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1604 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1605 | ALIAS: |
|
|
1606 | enqueue_write_buffer = 0 |
854 | PPCODE: |
1607 | PPCODE: |
855 | cl_event ev = 0; |
1608 | EVENT_LIST (5); |
|
|
1609 | |
856 | STRLEN len; |
1610 | STRLEN len; |
857 | char *ptr = SvPVbyte (data, len); |
1611 | char *ptr = SvPVbyte (data, len); |
858 | EVENT_LIST (5, items - 5); |
|
|
859 | |
1612 | |
|
|
1613 | cl_event ev = 0; |
860 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1614 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
861 | |
1615 | |
862 | if (ev) |
1616 | if (ev) |
863 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1617 | XPUSH_CLOBJ (stash_event, ev); |
864 | |
1618 | |
|
|
1619 | #if CL_VERSION_1_2 |
|
|
1620 | |
865 | void |
1621 | void |
866 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1622 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1623 | ALIAS: |
|
|
1624 | enqueue_fill_buffer = 0 |
867 | PPCODE: |
1625 | PPCODE: |
|
|
1626 | EVENT_LIST (5); |
|
|
1627 | |
|
|
1628 | STRLEN len; |
|
|
1629 | char *ptr = SvPVbyte (data, len); |
|
|
1630 | |
868 | cl_event ev = 0; |
1631 | cl_event ev = 0; |
|
|
1632 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1633 | |
|
|
1634 | if (ev) |
|
|
1635 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1636 | |
|
|
1637 | void |
|
|
1638 | 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, ...) |
|
|
1639 | ALIAS: |
|
|
1640 | enqueue_fill_image = 0 |
|
|
1641 | PPCODE: |
|
|
1642 | EVENT_LIST (12); |
|
|
1643 | |
|
|
1644 | const size_t origin [3] = { x, y, z }; |
|
|
1645 | const size_t region [3] = { width, height, depth }; |
|
|
1646 | |
|
|
1647 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1648 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1649 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1650 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1651 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1652 | cl_image_format format; |
|
|
1653 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1654 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1655 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1656 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1657 | |
|
|
1658 | cl_event ev = 0; |
|
|
1659 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1660 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1661 | |
|
|
1662 | if (ev) |
|
|
1663 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1664 | |
|
|
1665 | #endif |
|
|
1666 | |
|
|
1667 | void |
|
|
1668 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1669 | ALIAS: |
|
|
1670 | enqueue_copy_buffer = 0 |
|
|
1671 | PPCODE: |
869 | EVENT_LIST (6, items - 6); |
1672 | EVENT_LIST (6); |
870 | |
1673 | |
|
|
1674 | cl_event ev = 0; |
871 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1675 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
872 | |
1676 | |
873 | if (ev) |
1677 | if (ev) |
874 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1678 | XPUSH_CLOBJ (stash_event, ev); |
875 | |
1679 | |
876 | void |
1680 | void |
877 | enqueue_read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1681 | 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, ...) |
|
|
1682 | ALIAS: |
|
|
1683 | enqueue_read_buffer_rect = 0 |
878 | PPCODE: |
1684 | PPCODE: |
879 | cl_event ev = 0; |
1685 | EVENT_LIST (17); |
|
|
1686 | |
880 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1687 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
881 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1688 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
882 | const size_t region[3] = { width, height, depth }; |
1689 | const size_t region[3] = { width, height, depth }; |
883 | EVENT_LIST (17, items - 17); |
|
|
884 | |
1690 | |
885 | if (!buf_row_pitch) |
1691 | if (!buf_row_pitch) |
886 | buf_row_pitch = region [0]; |
1692 | buf_row_pitch = region [0]; |
887 | |
1693 | |
888 | if (!buf_slice_pitch) |
1694 | if (!buf_slice_pitch) |
… | |
… | |
898 | |
1704 | |
899 | SvUPGRADE (data, SVt_PV); |
1705 | SvUPGRADE (data, SVt_PV); |
900 | SvGROW (data, len); |
1706 | SvGROW (data, len); |
901 | SvPOK_only (data); |
1707 | SvPOK_only (data); |
902 | SvCUR_set (data, len); |
1708 | SvCUR_set (data, len); |
|
|
1709 | |
|
|
1710 | cl_event ev = 0; |
903 | NEED_SUCCESS (EnqueueReadBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1711 | NEED_SUCCESS (EnqueueReadBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
904 | |
1712 | |
905 | if (ev) |
1713 | if (ev) |
906 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1714 | XPUSH_CLOBJ (stash_event, ev); |
907 | |
1715 | |
908 | void |
1716 | void |
909 | enqueue_write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
1717 | 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, ...) |
|
|
1718 | ALIAS: |
|
|
1719 | enqueue_write_buffer_rect = 0 |
910 | PPCODE: |
1720 | PPCODE: |
911 | cl_event ev = 0; |
1721 | EVENT_LIST (17); |
|
|
1722 | |
912 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1723 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
913 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1724 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
914 | const size_t region[3] = { width, height, depth }; |
1725 | const size_t region[3] = { width, height, depth }; |
915 | STRLEN len; |
1726 | STRLEN len; |
916 | char *ptr = SvPVbyte (data, len); |
1727 | char *ptr = SvPVbyte (data, len); |
917 | EVENT_LIST (17, items - 17); |
|
|
918 | |
1728 | |
919 | if (!buf_row_pitch) |
1729 | if (!buf_row_pitch) |
920 | buf_row_pitch = region [0]; |
1730 | buf_row_pitch = region [0]; |
921 | |
1731 | |
922 | if (!buf_slice_pitch) |
1732 | if (!buf_slice_pitch) |
… | |
… | |
931 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1741 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
932 | |
1742 | |
933 | if (len < min_len) |
1743 | if (len < min_len) |
934 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1744 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
935 | |
1745 | |
|
|
1746 | cl_event ev = 0; |
936 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1747 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
937 | |
1748 | |
938 | if (ev) |
1749 | if (ev) |
939 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1750 | XPUSH_CLOBJ (stash_event, ev); |
940 | |
1751 | |
941 | void |
1752 | void |
942 | enqueue_copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
1753 | 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, ...) |
|
|
1754 | ALIAS: |
|
|
1755 | enqueue_copy_buffer_rect = 0 |
943 | PPCODE: |
1756 | PPCODE: |
944 | cl_event ev = 0; |
1757 | EVENT_LIST (16); |
|
|
1758 | |
945 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1759 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
946 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1760 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
947 | const size_t region[3] = { width, height, depth }; |
1761 | const size_t region[3] = { width, height, depth }; |
948 | EVENT_LIST (16, items - 16); |
|
|
949 | |
1762 | |
|
|
1763 | cl_event ev = 0; |
950 | NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1764 | NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
951 | |
1765 | |
952 | if (ev) |
1766 | if (ev) |
953 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1767 | XPUSH_CLOBJ (stash_event, ev); |
954 | |
1768 | |
955 | void |
1769 | void |
956 | enqueue_read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1770 | 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, ...) |
|
|
1771 | ALIAS: |
|
|
1772 | enqueue_read_image = 0 |
957 | PPCODE: |
1773 | PPCODE: |
958 | cl_event ev = 0; |
1774 | EVENT_LIST (12); |
|
|
1775 | |
959 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1776 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
960 | const size_t region[3] = { width, height, depth }; |
1777 | const size_t region[3] = { width, height, depth }; |
961 | EVENT_LIST (12, items - 12); |
|
|
962 | |
1778 | |
963 | if (!row_pitch) |
1779 | if (!row_pitch) |
964 | row_pitch = img_row_pitch (src); |
1780 | row_pitch = img_row_pitch (src); |
965 | |
1781 | |
966 | if (depth > 1 && !slice_pitch) |
1782 | if (depth > 1 && !slice_pitch) |
… | |
… | |
970 | |
1786 | |
971 | SvUPGRADE (data, SVt_PV); |
1787 | SvUPGRADE (data, SVt_PV); |
972 | SvGROW (data, len); |
1788 | SvGROW (data, len); |
973 | SvPOK_only (data); |
1789 | SvPOK_only (data); |
974 | SvCUR_set (data, len); |
1790 | SvCUR_set (data, len); |
|
|
1791 | |
|
|
1792 | cl_event ev = 0; |
975 | NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1793 | NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
976 | |
1794 | |
977 | if (ev) |
1795 | if (ev) |
978 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1796 | XPUSH_CLOBJ (stash_event, ev); |
979 | |
1797 | |
980 | void |
1798 | void |
981 | enqueue_write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
1799 | 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, ...) |
|
|
1800 | ALIAS: |
|
|
1801 | enqueue_write_image = 0 |
982 | PPCODE: |
1802 | PPCODE: |
983 | cl_event ev = 0; |
1803 | EVENT_LIST (12); |
|
|
1804 | |
984 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1805 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
985 | const size_t region[3] = { width, height, depth }; |
1806 | const size_t region[3] = { width, height, depth }; |
986 | STRLEN len; |
1807 | STRLEN len; |
987 | char *ptr = SvPVbyte (data, len); |
1808 | char *ptr = SvPVbyte (data, len); |
988 | EVENT_LIST (12, items - 12); |
|
|
989 | |
1809 | |
990 | if (!row_pitch) |
1810 | if (!row_pitch) |
991 | row_pitch = img_row_pitch (dst); |
1811 | row_pitch = img_row_pitch (dst); |
992 | |
1812 | |
993 | if (depth > 1 && !slice_pitch) |
1813 | if (depth > 1 && !slice_pitch) |
… | |
… | |
996 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1816 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
997 | |
1817 | |
998 | if (len < min_len) |
1818 | if (len < min_len) |
999 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1819 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1000 | |
1820 | |
|
|
1821 | cl_event ev = 0; |
1001 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1822 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1002 | |
1823 | |
1003 | if (ev) |
1824 | if (ev) |
1004 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1825 | XPUSH_CLOBJ (stash_event, ev); |
1005 | |
1826 | |
1006 | void |
1827 | void |
1007 | enqueue_copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
1828 | 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, ...) |
|
|
1829 | ALIAS: |
|
|
1830 | enqueue_copy_image = 0 |
1008 | PPCODE: |
1831 | PPCODE: |
1009 | cl_event ev = 0; |
1832 | EVENT_LIST (12); |
|
|
1833 | |
1010 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1834 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1011 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1835 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1012 | const size_t region[3] = { width, height, depth }; |
1836 | const size_t region[3] = { width, height, depth }; |
1013 | EVENT_LIST (12, items - 12); |
|
|
1014 | |
1837 | |
|
|
1838 | cl_event ev = 0; |
1015 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1839 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1016 | |
1840 | |
1017 | if (ev) |
1841 | if (ev) |
1018 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1842 | XPUSH_CLOBJ (stash_event, ev); |
1019 | |
1843 | |
1020 | void |
1844 | void |
1021 | enqueue_copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
1845 | 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, ...) |
|
|
1846 | ALIAS: |
|
|
1847 | enqueue_copy_image_to_buffer = 0 |
1022 | PPCODE: |
1848 | PPCODE: |
|
|
1849 | EVENT_LIST (10); |
|
|
1850 | |
|
|
1851 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1852 | const size_t region [3] = { width, height, depth }; |
|
|
1853 | |
1023 | cl_event ev = 0; |
1854 | cl_event ev = 0; |
|
|
1855 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1856 | |
|
|
1857 | if (ev) |
|
|
1858 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1859 | |
|
|
1860 | void |
|
|
1861 | 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, ...) |
|
|
1862 | ALIAS: |
|
|
1863 | enqueue_copy_buffer_to_image = 0 |
|
|
1864 | PPCODE: |
|
|
1865 | EVENT_LIST (10); |
|
|
1866 | |
1024 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1867 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1868 | const size_t region [3] = { width, height, depth }; |
|
|
1869 | |
|
|
1870 | cl_event ev = 0; |
|
|
1871 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1872 | |
|
|
1873 | if (ev) |
|
|
1874 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1875 | |
|
|
1876 | void |
|
|
1877 | 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, ...) |
|
|
1878 | ALIAS: |
|
|
1879 | enqueue_map_buffer = 0 |
|
|
1880 | PPCODE: |
|
|
1881 | EVENT_LIST (6); |
|
|
1882 | |
|
|
1883 | size_t cb = SvIV (cb_); |
|
|
1884 | |
|
|
1885 | if (!SvOK (cb_)) |
|
|
1886 | { |
|
|
1887 | NEED_SUCCESS (GetMemObjectInfo, (buf, CL_MEM_SIZE, sizeof (cb), &cb, 0)); |
|
|
1888 | cb -= offset; |
|
|
1889 | } |
|
|
1890 | |
|
|
1891 | cl_event ev; |
|
|
1892 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1893 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1894 | |
|
|
1895 | void |
|
|
1896 | 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, ...) |
|
|
1897 | ALIAS: |
|
|
1898 | enqueue_map_image = 0 |
|
|
1899 | PPCODE: |
|
|
1900 | size_t width = SvIV (width_); |
|
|
1901 | if (!SvOK (width_)) |
|
|
1902 | { |
|
|
1903 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_WIDTH, sizeof (width), &width, 0)); |
|
|
1904 | width -= x; |
|
|
1905 | } |
|
|
1906 | |
|
|
1907 | size_t height = SvIV (width_); |
|
|
1908 | if (!SvOK (height_)) |
|
|
1909 | { |
|
|
1910 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_HEIGHT, sizeof (height), &height, 0)); |
|
|
1911 | height -= y; |
|
|
1912 | } |
|
|
1913 | |
|
|
1914 | size_t depth = SvIV (width_); |
|
|
1915 | if (!SvOK (depth_)) |
|
|
1916 | { |
|
|
1917 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_DEPTH, sizeof (depth), &depth, 0)); |
|
|
1918 | depth -= z; |
|
|
1919 | |
|
|
1920 | // stupid opencl returns 0 for depth, but requires 1 for 2d images |
|
|
1921 | if (!depth) |
|
|
1922 | depth = 1; |
|
|
1923 | } |
|
|
1924 | |
|
|
1925 | const size_t origin[3] = { x, y, z }; |
1025 | const size_t region[3] = { width, height, depth }; |
1926 | const size_t region[3] = { width, height, depth }; |
|
|
1927 | size_t row_pitch, slice_pitch; |
1026 | EVENT_LIST (10, items - 10); |
1928 | EVENT_LIST (10); |
1027 | |
1929 | |
|
|
1930 | cl_event ev; |
|
|
1931 | 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)); |
|
|
1932 | |
|
|
1933 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1934 | : row_pitch ? row_pitch * region [1] |
|
|
1935 | : region [0]; |
|
|
1936 | |
|
|
1937 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1938 | |
|
|
1939 | void |
|
|
1940 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1941 | PPCODE: |
|
|
1942 | mapped_unmap (cv, ST (1), mapped, self, &ST (2), items - 2); |
|
|
1943 | if (GIMME_V != G_VOID) |
|
|
1944 | { |
|
|
1945 | clRetainEvent (mapped->event); |
|
|
1946 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1947 | } |
|
|
1948 | |
|
|
1949 | void |
|
|
1950 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1951 | ALIAS: |
|
|
1952 | enqueue_task = 0 |
|
|
1953 | PPCODE: |
|
|
1954 | EVENT_LIST (2); |
|
|
1955 | |
|
|
1956 | cl_event ev = 0; |
1028 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1957 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1029 | |
1958 | |
1030 | if (ev) |
1959 | if (ev) |
1031 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1960 | XPUSH_CLOBJ (stash_event, ev); |
1032 | |
1961 | |
1033 | void |
1962 | void |
1034 | enqueue_copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
1035 | PPCODE: |
|
|
1036 | cl_event ev = 0; |
|
|
1037 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1038 | const size_t region[3] = { width, height, depth }; |
|
|
1039 | EVENT_LIST (10, items - 10); |
|
|
1040 | |
|
|
1041 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1042 | |
|
|
1043 | if (ev) |
|
|
1044 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1045 | |
|
|
1046 | void |
|
|
1047 | enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1048 | PPCODE: |
|
|
1049 | cl_event ev = 0; |
|
|
1050 | EVENT_LIST (2, items - 2); |
|
|
1051 | |
|
|
1052 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1053 | |
|
|
1054 | if (ev) |
|
|
1055 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1056 | |
|
|
1057 | void |
|
|
1058 | enqueue_nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
1963 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1964 | ALIAS: |
|
|
1965 | enqueue_nd_range_kernel = 0 |
1059 | PPCODE: |
1966 | PPCODE: |
1060 | cl_event ev = 0; |
1967 | EVENT_LIST (5); |
|
|
1968 | |
1061 | size_t *gwo = 0, *gws, *lws = 0; |
1969 | size_t *gwo = 0, *gws, *lws = 0; |
1062 | int gws_len; |
1970 | int gws_len; |
1063 | size_t *lists; |
1971 | size_t *lists; |
1064 | int i; |
1972 | int i; |
1065 | EVENT_LIST (5, items - 5); |
|
|
1066 | |
1973 | |
1067 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1974 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1068 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1975 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1069 | |
1976 | |
1070 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1977 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1071 | |
1978 | |
1072 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1979 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1073 | |
1980 | |
1074 | gws = lists + gws_len * 0; |
1981 | gws = lists + gws_len * 0; |
1075 | for (i = 0; i < gws_len; ++i) |
1982 | for (i = 0; i < gws_len; ++i) |
|
|
1983 | { |
1076 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
1984 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1985 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1986 | if (!gws [i]) |
|
|
1987 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1988 | } |
1077 | |
1989 | |
1078 | if (SvOK (global_work_offset)) |
1990 | if (SvOK (global_work_offset)) |
1079 | { |
1991 | { |
1080 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1992 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1081 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
1993 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
1088 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
2000 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1089 | } |
2001 | } |
1090 | |
2002 | |
1091 | if (SvOK (local_work_size)) |
2003 | if (SvOK (local_work_size)) |
1092 | { |
2004 | { |
1093 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
2005 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1094 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
2006 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
1095 | |
2007 | |
1096 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
2008 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1097 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
2009 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1098 | |
2010 | |
1099 | lws = lists + gws_len * 2; |
2011 | lws = lists + gws_len * 2; |
1100 | for (i = 0; i < gws_len; ++i) |
2012 | for (i = 0; i < gws_len; ++i) |
|
|
2013 | { |
1101 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
2014 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
2015 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
2016 | if (!lws [i]) |
|
|
2017 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
2018 | } |
1102 | } |
2019 | } |
1103 | |
2020 | |
|
|
2021 | cl_event ev = 0; |
1104 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2022 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1105 | |
2023 | |
1106 | if (ev) |
2024 | if (ev) |
1107 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2025 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2026 | |
|
|
2027 | #if CL_VERSION_1_2 |
|
|
2028 | |
|
|
2029 | void |
|
|
2030 | migrate_mem_objects (OpenCL::Queue self, SV *objects, cl_mem_migration_flags flags, ...) |
|
|
2031 | ALIAS: |
|
|
2032 | enqueue_migrate_mem_objects = 0 |
|
|
2033 | PPCODE: |
|
|
2034 | EVENT_LIST (3); |
|
|
2035 | |
|
|
2036 | cl_uint obj_count; |
|
|
2037 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
|
|
2038 | |
|
|
2039 | cl_event ev = 0; |
|
|
2040 | NEED_SUCCESS (EnqueueMigrateMemObjects, (self, obj_count, obj_list, flags, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2041 | |
|
|
2042 | if (ev) |
|
|
2043 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2044 | |
|
|
2045 | #endif |
1108 | |
2046 | |
1109 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2047 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
1110 | |
2048 | |
1111 | void |
2049 | void |
1112 | enqueue_acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
2050 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
1113 | ALIAS: |
2051 | ALIAS: |
|
|
2052 | release_gl_objects = 1 |
|
|
2053 | enqueue_acquire_gl_objects = 0 |
1114 | enqueue_release_gl_objects = 1 |
2054 | enqueue_release_gl_objects = 1 |
1115 | CODE: |
2055 | PPCODE: |
1116 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
2056 | EVENT_LIST (2); |
1117 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
2057 | |
|
|
2058 | cl_uint obj_count; |
|
|
2059 | cl_mem *obj_list = object_list (cv, 0, "objects", objects, "OpenCL::Memory", &obj_count); |
1118 | |
2060 | |
1119 | cl_event ev = 0; |
2061 | cl_event ev = 0; |
1120 | EVENT_LIST (2, items - 2); |
|
|
1121 | AV *av = (AV *)SvRV (objects); |
|
|
1122 | cl_uint num_objects = av_len (av) + 1; |
|
|
1123 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1124 | int i; |
|
|
1125 | |
|
|
1126 | for (i = num_objects; i--; ) |
|
|
1127 | object_list [i] = SvPTROBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1128 | |
2062 | |
1129 | if (ix) |
2063 | if (ix) |
1130 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2064 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1131 | else |
2065 | else |
1132 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
2066 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, obj_count, obj_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1133 | |
2067 | |
1134 | if (ev) |
2068 | if (ev) |
1135 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
2069 | XPUSH_CLOBJ (stash_event, ev); |
1136 | |
2070 | |
1137 | #endif |
2071 | #endif |
1138 | |
2072 | |
1139 | void |
2073 | void |
1140 | enqueue_marker (OpenCL::Queue self) |
|
|
1141 | PPCODE: |
|
|
1142 | cl_event ev; |
|
|
1143 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1144 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
1145 | |
|
|
1146 | void |
|
|
1147 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
2074 | wait_for_events (OpenCL::Queue self, ...) |
|
|
2075 | ALIAS: |
|
|
2076 | enqueue_wait_for_events = 0 |
1148 | CODE: |
2077 | CODE: |
1149 | EVENT_LIST (1, items - 1); |
2078 | EVENT_LIST (1); |
|
|
2079 | #if PREFER_1_1 |
1150 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
2080 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2081 | #else |
|
|
2082 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
2083 | #endif |
1151 | |
2084 | |
1152 | void |
2085 | void |
1153 | enqueue_barrier (OpenCL::Queue self) |
2086 | marker (OpenCL::Queue self, ...) |
|
|
2087 | ALIAS: |
|
|
2088 | enqueue_marker = 0 |
1154 | CODE: |
2089 | PPCODE: |
|
|
2090 | EVENT_LIST (1); |
|
|
2091 | cl_event ev = 0; |
|
|
2092 | #if PREFER_1_1 |
|
|
2093 | if (!event_list_count) |
|
|
2094 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2095 | else |
|
|
2096 | #if CL_VERSION_1_2 |
|
|
2097 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2098 | #else |
|
|
2099 | { |
|
|
2100 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
2101 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2102 | } |
|
|
2103 | #endif |
|
|
2104 | #else |
|
|
2105 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2106 | #endif |
|
|
2107 | if (ev) |
|
|
2108 | XPUSH_CLOBJ (stash_event, ev); |
|
|
2109 | |
|
|
2110 | void |
|
|
2111 | barrier (OpenCL::Queue self, ...) |
|
|
2112 | ALIAS: |
|
|
2113 | enqueue_barrier = 0 |
|
|
2114 | PPCODE: |
|
|
2115 | EVENT_LIST (1); |
|
|
2116 | cl_event ev = 0; |
|
|
2117 | #if PREFER_1_1 |
|
|
2118 | if (!event_list_count && GIMME_V == G_VOID) |
1155 | NEED_SUCCESS (EnqueueBarrier, (self)); |
2119 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
2120 | else |
|
|
2121 | #if CL_VERSION_1_2 |
|
|
2122 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2123 | #else |
|
|
2124 | { |
|
|
2125 | if (event_list_count) |
|
|
2126 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
2127 | |
|
|
2128 | if (GIMME_V != G_VOID) |
|
|
2129 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
2130 | } |
|
|
2131 | #endif |
|
|
2132 | #else |
|
|
2133 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
2134 | #endif |
|
|
2135 | if (ev) |
|
|
2136 | XPUSH_CLOBJ (stash_event, ev); |
1156 | |
2137 | |
1157 | void |
2138 | void |
1158 | flush (OpenCL::Queue self) |
2139 | flush (OpenCL::Queue self) |
1159 | CODE: |
2140 | CODE: |
1160 | NEED_SUCCESS (Flush, (self)); |
2141 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1178 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
2159 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1179 | EXTEND (SP, 1); |
2160 | EXTEND (SP, 1); |
1180 | const int i = 0; |
2161 | const int i = 0; |
1181 | { |
2162 | { |
1182 | NEED_SUCCESS (RetainContext, (value [i])); |
2163 | NEED_SUCCESS (RetainContext, (value [i])); |
1183 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2164 | PUSH_CLOBJ (stash_context, value [i]); |
1184 | } |
2165 | } |
1185 | |
2166 | |
1186 | void |
2167 | void |
1187 | device (OpenCL::Queue self) |
2168 | device (OpenCL::Queue self) |
1188 | PPCODE: |
2169 | PPCODE: |
1189 | cl_device_id value [1]; |
2170 | cl_device_id value [1]; |
1190 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
2171 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1191 | EXTEND (SP, 1); |
2172 | EXTEND (SP, 1); |
1192 | const int i = 0; |
2173 | const int i = 0; |
1193 | { |
2174 | { |
1194 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2175 | PUSH_CLOBJ (stash_device, value [i]); |
1195 | } |
2176 | } |
1196 | |
2177 | |
1197 | void |
2178 | void |
1198 | reference_count (OpenCL::Queue self) |
2179 | reference_count (OpenCL::Queue self) |
1199 | PPCODE: |
2180 | PPCODE: |
… | |
… | |
1208 | PPCODE: |
2189 | PPCODE: |
1209 | cl_command_queue_properties value [1]; |
2190 | cl_command_queue_properties value [1]; |
1210 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
2191 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1211 | EXTEND (SP, 1); |
2192 | EXTEND (SP, 1); |
1212 | const int i = 0; |
2193 | const int i = 0; |
1213 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2194 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1214 | |
2195 | |
1215 | #END:command_queue |
2196 | #END:command_queue |
1216 | |
2197 | |
1217 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
2198 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1218 | |
2199 | |
… | |
… | |
1223 | |
2204 | |
1224 | void |
2205 | void |
1225 | info (OpenCL::Memory self, cl_mem_info name) |
2206 | info (OpenCL::Memory self, cl_mem_info name) |
1226 | PPCODE: |
2207 | PPCODE: |
1227 | INFO (MemObject) |
2208 | INFO (MemObject) |
|
|
2209 | |
|
|
2210 | void |
|
|
2211 | destructor_callback (OpenCL::Memory self, SV *notify) |
|
|
2212 | PPCODE: |
1228 | |
2213 | |
1229 | #BEGIN:mem |
2214 | #BEGIN:mem |
1230 | |
2215 | |
1231 | void |
2216 | void |
1232 | type (OpenCL::Memory self) |
2217 | type (OpenCL::Memory self) |
1233 | PPCODE: |
2218 | PPCODE: |
1234 | cl_mem_object_type value [1]; |
2219 | cl_mem_object_type value [1]; |
1235 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
2220 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1236 | EXTEND (SP, 1); |
2221 | EXTEND (SP, 1); |
1237 | const int i = 0; |
2222 | const int i = 0; |
1238 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2223 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1239 | |
2224 | |
1240 | void |
2225 | void |
1241 | flags (OpenCL::Memory self) |
2226 | flags (OpenCL::Memory self) |
1242 | PPCODE: |
2227 | PPCODE: |
1243 | cl_mem_flags value [1]; |
2228 | cl_mem_flags value [1]; |
1244 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
2229 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1245 | EXTEND (SP, 1); |
2230 | EXTEND (SP, 1); |
1246 | const int i = 0; |
2231 | const int i = 0; |
1247 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2232 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1248 | |
2233 | |
1249 | void |
2234 | void |
1250 | size (OpenCL::Memory self) |
2235 | size (OpenCL::Memory self) |
1251 | ALIAS: |
2236 | ALIAS: |
1252 | size = CL_MEM_SIZE |
2237 | size = CL_MEM_SIZE |
… | |
… | |
1286 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
2271 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1287 | EXTEND (SP, 1); |
2272 | EXTEND (SP, 1); |
1288 | const int i = 0; |
2273 | const int i = 0; |
1289 | { |
2274 | { |
1290 | NEED_SUCCESS (RetainContext, (value [i])); |
2275 | NEED_SUCCESS (RetainContext, (value [i])); |
1291 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2276 | PUSH_CLOBJ (stash_context, value [i]); |
1292 | } |
2277 | } |
1293 | |
2278 | |
1294 | void |
2279 | void |
1295 | associated_memobject (OpenCL::Memory self) |
2280 | associated_memobject (OpenCL::Memory self) |
1296 | PPCODE: |
2281 | PPCODE: |
… | |
… | |
1298 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
2283 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1299 | EXTEND (SP, 1); |
2284 | EXTEND (SP, 1); |
1300 | const int i = 0; |
2285 | const int i = 0; |
1301 | { |
2286 | { |
1302 | NEED_SUCCESS (RetainMemObject, (value [i])); |
2287 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1303 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
2288 | PUSH_CLOBJ (stash_memory, value [i]); |
1304 | } |
2289 | } |
1305 | |
2290 | |
1306 | #END:mem |
2291 | #END:mem |
1307 | |
2292 | |
1308 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
2293 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
… | |
… | |
1328 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
2313 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1329 | |
2314 | |
1330 | cl_buffer_region crdata = { origin, size }; |
2315 | cl_buffer_region crdata = { origin, size }; |
1331 | |
2316 | |
1332 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
2317 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
1333 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
2318 | XPUSH_CLOBJ (stash_buffer, mem); |
1334 | |
2319 | |
1335 | MODULE = OpenCL PACKAGE = OpenCL::Image |
2320 | MODULE = OpenCL PACKAGE = OpenCL::Image |
1336 | |
2321 | |
1337 | void |
2322 | void |
1338 | image_info (OpenCL::Image self, cl_image_info name) |
2323 | image_info (OpenCL::Image self, cl_image_info name) |
1339 | PPCODE: |
2324 | PPCODE: |
1340 | INFO (Image) |
2325 | INFO (Image) |
|
|
2326 | |
|
|
2327 | void |
|
|
2328 | format (OpenCL::Image self) |
|
|
2329 | PPCODE: |
|
|
2330 | cl_image_format format; |
|
|
2331 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2332 | EXTEND (SP, 2); |
|
|
2333 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2334 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
1341 | |
2335 | |
1342 | #BEGIN:image |
2336 | #BEGIN:image |
1343 | |
2337 | |
1344 | void |
2338 | void |
1345 | element_size (OpenCL::Image self) |
2339 | element_size (OpenCL::Image self) |
… | |
… | |
1415 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2409 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
1416 | EXTEND (SP, 1); |
2410 | EXTEND (SP, 1); |
1417 | const int i = 0; |
2411 | const int i = 0; |
1418 | { |
2412 | { |
1419 | NEED_SUCCESS (RetainContext, (value [i])); |
2413 | NEED_SUCCESS (RetainContext, (value [i])); |
1420 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2414 | PUSH_CLOBJ (stash_context, value [i]); |
1421 | } |
2415 | } |
1422 | |
2416 | |
1423 | void |
2417 | void |
1424 | normalized_coords (OpenCL::Sampler self) |
2418 | normalized_coords (OpenCL::Sampler self) |
1425 | PPCODE: |
2419 | PPCODE: |
1426 | cl_addressing_mode value [1]; |
2420 | cl_addressing_mode value [1]; |
1427 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
2421 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
1428 | EXTEND (SP, 1); |
2422 | EXTEND (SP, 1); |
1429 | const int i = 0; |
2423 | const int i = 0; |
1430 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2424 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1431 | |
2425 | |
1432 | void |
2426 | void |
1433 | addressing_mode (OpenCL::Sampler self) |
2427 | addressing_mode (OpenCL::Sampler self) |
1434 | PPCODE: |
2428 | PPCODE: |
1435 | cl_filter_mode value [1]; |
2429 | cl_filter_mode value [1]; |
1436 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
2430 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
1437 | EXTEND (SP, 1); |
2431 | EXTEND (SP, 1); |
1438 | const int i = 0; |
2432 | const int i = 0; |
1439 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2433 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1440 | |
2434 | |
1441 | void |
2435 | void |
1442 | filter_mode (OpenCL::Sampler self) |
2436 | filter_mode (OpenCL::Sampler self) |
1443 | PPCODE: |
2437 | PPCODE: |
1444 | cl_bool value [1]; |
2438 | cl_bool value [1]; |
… | |
… | |
1455 | DESTROY (OpenCL::Program self) |
2449 | DESTROY (OpenCL::Program self) |
1456 | CODE: |
2450 | CODE: |
1457 | clReleaseProgram (self); |
2451 | clReleaseProgram (self); |
1458 | |
2452 | |
1459 | void |
2453 | void |
1460 | build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) |
2454 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2455 | ALIAS: |
|
|
2456 | build_async = 1 |
1461 | CODE: |
2457 | CODE: |
|
|
2458 | cl_uint device_count = 0; |
|
|
2459 | cl_device_id *device_list = 0; |
|
|
2460 | |
|
|
2461 | if (SvOK (devices)) |
|
|
2462 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2463 | |
|
|
2464 | void *user_data; |
|
|
2465 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2466 | |
|
|
2467 | if (ix) |
|
|
2468 | build_program_async (self, device_count, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2469 | else |
|
|
2470 | NEED_SUCCESS (BuildProgram, (self, device_count, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
|
|
2471 | |
|
|
2472 | #if CL_VERSION_1_2 |
|
|
2473 | |
|
|
2474 | void |
|
|
2475 | compile (OpenCL::Program self, SV *devices, SV *options = &PL_sv_undef, SV *headers = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2476 | CODE: |
|
|
2477 | cl_uint device_count = 0; |
|
|
2478 | cl_device_id *device_list = 0; |
|
|
2479 | |
|
|
2480 | if (SvOK (devices)) |
|
|
2481 | device_list = object_list (cv, 1, "devices", devices, "OpenCL::Device", &device_count); |
|
|
2482 | |
|
|
2483 | cl_uint header_count = 0; |
|
|
2484 | cl_program *header_list = 0; |
|
|
2485 | const char **header_name = 0; |
|
|
2486 | |
|
|
2487 | if (SvOK (headers)) |
|
|
2488 | { |
|
|
2489 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVHV) |
|
|
2490 | croak ("clCompileProgram: headers must be undef or a hashref of name => OpenCL::Program pairs"); |
|
|
2491 | |
|
|
2492 | HV *hv = (HV *)SvRV (devices); |
|
|
2493 | |
|
|
2494 | header_count = hv_iterinit (hv); |
|
|
2495 | header_list = tmpbuf (sizeof (*header_list) * header_count); |
|
|
2496 | header_name = tmpbuf (sizeof (*header_name) * header_count); |
|
|
2497 | |
|
|
2498 | HE *he; |
|
|
2499 | int i = 0; |
|
|
2500 | while (he = hv_iternext (hv)) |
|
|
2501 | { |
|
|
2502 | header_name [i] = SvPVbyte_nolen (HeSVKEY_force (he)); |
|
|
2503 | header_list [i] = SvCLOBJ (cv, "headers", HeVAL (he), "OpenCL::Program"); |
|
|
2504 | ++i; |
|
|
2505 | } |
|
|
2506 | } |
|
|
2507 | |
|
|
2508 | void *user_data; |
|
|
2509 | program_callback pfn_notify = make_program_callback (notify, &user_data); |
|
|
2510 | |
1462 | NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2511 | NEED_SUCCESS (CompileProgram, (self, device_count, device_list, SvPVbyte_nolen (options), |
|
|
2512 | header_count, header_list, header_name, pfn_notify, user_data)); |
|
|
2513 | |
|
|
2514 | #endif |
1463 | |
2515 | |
1464 | void |
2516 | void |
1465 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2517 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
1466 | PPCODE: |
2518 | PPCODE: |
1467 | size_t size; |
2519 | size_t size; |
… | |
… | |
1496 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
2548 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, ix, size, value, 0)); |
1497 | EXTEND (SP, 1); |
2549 | EXTEND (SP, 1); |
1498 | const int i = 0; |
2550 | const int i = 0; |
1499 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
2551 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
1500 | |
2552 | |
|
|
2553 | void |
|
|
2554 | binary_type (OpenCL::Program self, OpenCL::Device device) |
|
|
2555 | PPCODE: |
|
|
2556 | cl_program_binary_type value [1]; |
|
|
2557 | NEED_SUCCESS (GetProgramBuildInfo, (self, device, CL_PROGRAM_BINARY_TYPE, sizeof (value), value, 0)); |
|
|
2558 | EXTEND (SP, 1); |
|
|
2559 | const int i = 0; |
|
|
2560 | PUSHs (sv_2mortal (newSVuv ((UV)value [i]))); |
|
|
2561 | |
1501 | #END:program_build |
2562 | #END:program_build |
1502 | |
2563 | |
1503 | void |
2564 | void |
1504 | kernel (OpenCL::Program program, SV *function) |
2565 | kernel (OpenCL::Program program, SV *function) |
1505 | PPCODE: |
2566 | PPCODE: |
1506 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
2567 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1507 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2568 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2569 | |
|
|
2570 | void |
|
|
2571 | kernels_in_program (OpenCL::Program program) |
|
|
2572 | PPCODE: |
|
|
2573 | cl_uint num_kernels; |
|
|
2574 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2575 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2576 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2577 | |
|
|
2578 | int i; |
|
|
2579 | EXTEND (SP, num_kernels); |
|
|
2580 | for (i = 0; i < num_kernels; ++i) |
|
|
2581 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
1508 | |
2582 | |
1509 | void |
2583 | void |
1510 | info (OpenCL::Program self, cl_program_info name) |
2584 | info (OpenCL::Program self, cl_program_info name) |
1511 | PPCODE: |
2585 | PPCODE: |
1512 | INFO (Program) |
2586 | INFO (Program) |
… | |
… | |
1530 | { |
2604 | { |
1531 | SV *sv = sv_2mortal (newSV (sizes [i])); |
2605 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1532 | SvUPGRADE (sv, SVt_PV); |
2606 | SvUPGRADE (sv, SVt_PV); |
1533 | SvPOK_only (sv); |
2607 | SvPOK_only (sv); |
1534 | SvCUR_set (sv, sizes [i]); |
2608 | SvCUR_set (sv, sizes [i]); |
1535 | ptrs [i] = SvPVX (sv); |
2609 | ptrs [i] = (void *)SvPVX (sv); |
1536 | PUSHs (sv); |
2610 | PUSHs (sv); |
1537 | } |
2611 | } |
1538 | |
2612 | |
1539 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
2613 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1540 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
2614 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
… | |
… | |
1560 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2634 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
1561 | EXTEND (SP, 1); |
2635 | EXTEND (SP, 1); |
1562 | const int i = 0; |
2636 | const int i = 0; |
1563 | { |
2637 | { |
1564 | NEED_SUCCESS (RetainContext, (value [i])); |
2638 | NEED_SUCCESS (RetainContext, (value [i])); |
1565 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2639 | PUSH_CLOBJ (stash_context, value [i]); |
1566 | } |
2640 | } |
1567 | |
2641 | |
1568 | void |
2642 | void |
1569 | devices (OpenCL::Program self) |
2643 | devices (OpenCL::Program self) |
1570 | PPCODE: |
2644 | PPCODE: |
… | |
… | |
1574 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2648 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
1575 | int i, n = size / sizeof (*value); |
2649 | int i, n = size / sizeof (*value); |
1576 | EXTEND (SP, n); |
2650 | EXTEND (SP, n); |
1577 | for (i = 0; i < n; ++i) |
2651 | for (i = 0; i < n; ++i) |
1578 | { |
2652 | { |
1579 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2653 | PUSH_CLOBJ (stash_device, value [i]); |
1580 | } |
2654 | } |
1581 | |
2655 | |
1582 | void |
2656 | void |
1583 | source (OpenCL::Program self) |
2657 | source (OpenCL::Program self) |
1584 | PPCODE: |
2658 | PPCODE: |
… | |
… | |
1610 | DESTROY (OpenCL::Kernel self) |
2684 | DESTROY (OpenCL::Kernel self) |
1611 | CODE: |
2685 | CODE: |
1612 | clReleaseKernel (self); |
2686 | clReleaseKernel (self); |
1613 | |
2687 | |
1614 | void |
2688 | void |
|
|
2689 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2690 | CODE: |
|
|
2691 | int i; |
|
|
2692 | for (i = 2; ; ++i) |
|
|
2693 | { |
|
|
2694 | while (*format == ' ') |
|
|
2695 | ++format; |
|
|
2696 | |
|
|
2697 | char type = *format++; |
|
|
2698 | |
|
|
2699 | if (!type) |
|
|
2700 | break; |
|
|
2701 | |
|
|
2702 | if (i >= items) |
|
|
2703 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2704 | |
|
|
2705 | SV *sv = ST (i); |
|
|
2706 | |
|
|
2707 | union |
|
|
2708 | { |
|
|
2709 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2710 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2711 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2712 | cl_mem cm; |
|
|
2713 | cl_sampler ca; |
|
|
2714 | size_t cz; |
|
|
2715 | cl_event ce; |
|
|
2716 | } arg; |
|
|
2717 | size_t size; |
|
|
2718 | int nullarg = 0; |
|
|
2719 | |
|
|
2720 | switch (type) |
|
|
2721 | { |
|
|
2722 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2723 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2724 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2725 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2726 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2727 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2728 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2729 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2730 | |
|
|
2731 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2732 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2733 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2734 | |
|
|
2735 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2736 | |
|
|
2737 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ (cv, "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2738 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2739 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ (cv, "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2740 | |
|
|
2741 | default: |
|
|
2742 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2743 | } |
|
|
2744 | |
|
|
2745 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2746 | if (res) |
|
|
2747 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2748 | } |
|
|
2749 | |
|
|
2750 | if (i != items) |
|
|
2751 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2752 | |
|
|
2753 | void |
1615 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
2754 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
1616 | CODE: |
2755 | CODE: |
1617 | clSetKernelArg (self, idx, sizeof (value), &value); |
2756 | clSetKernelArg (self, idx, sizeof (value), &value); |
1618 | |
2757 | |
1619 | void |
2758 | void |
… | |
… | |
1667 | clSetKernelArg (self, idx, sizeof (value), &value); |
2806 | clSetKernelArg (self, idx, sizeof (value), &value); |
1668 | |
2807 | |
1669 | void |
2808 | void |
1670 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
2809 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
1671 | CODE: |
2810 | CODE: |
|
|
2811 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2812 | |
|
|
2813 | void |
|
|
2814 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2815 | CODE: |
|
|
2816 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2817 | |
|
|
2818 | void |
|
|
2819 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2820 | CODE: |
|
|
2821 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2822 | |
|
|
2823 | void |
|
|
2824 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2825 | CODE: |
1672 | clSetKernelArg (self, idx, sizeof (value), &value); |
2826 | clSetKernelArg (self, idx, sizeof (value), &value); |
1673 | |
2827 | |
1674 | void |
2828 | void |
1675 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
2829 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
1676 | CODE: |
2830 | CODE: |
1677 | clSetKernelArg (self, idx, sizeof (value), &value); |
2831 | clSetKernelArg (self, idx, size, 0); |
1678 | |
|
|
1679 | void |
|
|
1680 | set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
1681 | CODE: |
|
|
1682 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1683 | |
|
|
1684 | void |
|
|
1685 | set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
1686 | CODE: |
|
|
1687 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1688 | |
|
|
1689 | void |
|
|
1690 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
1691 | CODE: |
|
|
1692 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1693 | |
2832 | |
1694 | void |
2833 | void |
1695 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
2834 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1696 | CODE: |
2835 | CODE: |
1697 | clSetKernelArg (self, idx, sizeof (value), &value); |
2836 | clSetKernelArg (self, idx, sizeof (value), &value); |
… | |
… | |
1733 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2872 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
1734 | EXTEND (SP, 1); |
2873 | EXTEND (SP, 1); |
1735 | const int i = 0; |
2874 | const int i = 0; |
1736 | { |
2875 | { |
1737 | NEED_SUCCESS (RetainContext, (value [i])); |
2876 | NEED_SUCCESS (RetainContext, (value [i])); |
1738 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2877 | PUSH_CLOBJ (stash_context, value [i]); |
1739 | } |
2878 | } |
1740 | |
2879 | |
1741 | void |
2880 | void |
1742 | program (OpenCL::Kernel self) |
2881 | program (OpenCL::Kernel self) |
1743 | PPCODE: |
2882 | PPCODE: |
… | |
… | |
1745 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2884 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
1746 | EXTEND (SP, 1); |
2885 | EXTEND (SP, 1); |
1747 | const int i = 0; |
2886 | const int i = 0; |
1748 | { |
2887 | { |
1749 | NEED_SUCCESS (RetainProgram, (value [i])); |
2888 | NEED_SUCCESS (RetainProgram, (value [i])); |
1750 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); |
2889 | PUSH_CLOBJ (stash_program, value [i]); |
1751 | } |
2890 | } |
1752 | |
2891 | |
1753 | #END:kernel |
2892 | #END:kernel |
1754 | |
2893 | |
1755 | void |
2894 | void |
… | |
… | |
1802 | const int i = 0; |
2941 | const int i = 0; |
1803 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2942 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1804 | |
2943 | |
1805 | #END:kernel_work_group |
2944 | #END:kernel_work_group |
1806 | |
2945 | |
|
|
2946 | #if CL_VERSION_1_2 |
|
|
2947 | |
|
|
2948 | void |
|
|
2949 | arg_info (OpenCL::Kernel self, cl_uint idx, cl_kernel_arg_info name) |
|
|
2950 | PPCODE: |
|
|
2951 | size_t size; |
|
|
2952 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, 0, 0, &size)); |
|
|
2953 | SV *sv = sv_2mortal (newSV (size)); |
|
|
2954 | SvUPGRADE (sv, SVt_PV); |
|
|
2955 | SvPOK_only (sv); |
|
|
2956 | SvCUR_set (sv, size); |
|
|
2957 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, name, size, SvPVX (sv), 0)); |
|
|
2958 | XPUSHs (sv); |
|
|
2959 | |
|
|
2960 | #BEGIN:kernel_arg |
|
|
2961 | |
|
|
2962 | void |
|
|
2963 | arg_address_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2964 | PPCODE: |
|
|
2965 | cl_kernel_arg_address_qualifier value [1]; |
|
|
2966 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ADDRESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2967 | EXTEND (SP, 1); |
|
|
2968 | const int i = 0; |
|
|
2969 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2970 | |
|
|
2971 | void |
|
|
2972 | arg_access_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2973 | PPCODE: |
|
|
2974 | cl_kernel_arg_access_qualifier value [1]; |
|
|
2975 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_ACCESS_QUALIFIER, sizeof (value), value, 0)); |
|
|
2976 | EXTEND (SP, 1); |
|
|
2977 | const int i = 0; |
|
|
2978 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2979 | |
|
|
2980 | void |
|
|
2981 | arg_type_name (OpenCL::Kernel self, cl_uint idx) |
|
|
2982 | ALIAS: |
|
|
2983 | arg_type_name = CL_KERNEL_ARG_TYPE_NAME |
|
|
2984 | arg_name = CL_KERNEL_ARG_NAME |
|
|
2985 | PPCODE: |
|
|
2986 | size_t size; |
|
|
2987 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, 0, 0, &size)); |
|
|
2988 | char *value = tmpbuf (size); |
|
|
2989 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, ix, size, value, 0)); |
|
|
2990 | EXTEND (SP, 1); |
|
|
2991 | const int i = 0; |
|
|
2992 | PUSHs (sv_2mortal (newSVpv (value, 0))); |
|
|
2993 | |
|
|
2994 | void |
|
|
2995 | arg_type_qualifier (OpenCL::Kernel self, cl_uint idx) |
|
|
2996 | PPCODE: |
|
|
2997 | cl_kernel_arg_type_qualifier value [1]; |
|
|
2998 | NEED_SUCCESS (GetKernelArgInfo, (self, idx, CL_KERNEL_ARG_TYPE_QUALIFIER, sizeof (value), value, 0)); |
|
|
2999 | EXTEND (SP, 1); |
|
|
3000 | const int i = 0; |
|
|
3001 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
3002 | |
|
|
3003 | #END:kernel_arg |
|
|
3004 | |
|
|
3005 | #endif |
|
|
3006 | |
1807 | MODULE = OpenCL PACKAGE = OpenCL::Event |
3007 | MODULE = OpenCL PACKAGE = OpenCL::Event |
1808 | |
3008 | |
1809 | void |
3009 | void |
1810 | DESTROY (OpenCL::Event self) |
3010 | DESTROY (OpenCL::Event self) |
1811 | CODE: |
3011 | CODE: |
… | |
… | |
1813 | |
3013 | |
1814 | void |
3014 | void |
1815 | wait (OpenCL::Event self) |
3015 | wait (OpenCL::Event self) |
1816 | CODE: |
3016 | CODE: |
1817 | clWaitForEvents (1, &self); |
3017 | clWaitForEvents (1, &self); |
|
|
3018 | |
|
|
3019 | void |
|
|
3020 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
3021 | CODE: |
|
|
3022 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
1818 | |
3023 | |
1819 | void |
3024 | void |
1820 | info (OpenCL::Event self, cl_event_info name) |
3025 | info (OpenCL::Event self, cl_event_info name) |
1821 | PPCODE: |
3026 | PPCODE: |
1822 | INFO (Event) |
3027 | INFO (Event) |
… | |
… | |
1830 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
3035 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
1831 | EXTEND (SP, 1); |
3036 | EXTEND (SP, 1); |
1832 | const int i = 0; |
3037 | const int i = 0; |
1833 | { |
3038 | { |
1834 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
3039 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
1835 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); |
3040 | PUSH_CLOBJ (stash_queue, value [i]); |
1836 | } |
3041 | } |
1837 | |
3042 | |
1838 | void |
3043 | void |
1839 | command_type (OpenCL::Event self) |
3044 | command_type (OpenCL::Event self) |
1840 | PPCODE: |
3045 | PPCODE: |
… | |
… | |
1863 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
3068 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
1864 | EXTEND (SP, 1); |
3069 | EXTEND (SP, 1); |
1865 | const int i = 0; |
3070 | const int i = 0; |
1866 | { |
3071 | { |
1867 | NEED_SUCCESS (RetainContext, (value [i])); |
3072 | NEED_SUCCESS (RetainContext, (value [i])); |
1868 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
3073 | PUSH_CLOBJ (stash_context, value [i]); |
1869 | } |
3074 | } |
1870 | |
3075 | |
1871 | #END:event |
3076 | #END:event |
1872 | |
3077 | |
1873 | void |
3078 | void |
… | |
… | |
1898 | void |
3103 | void |
1899 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
3104 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
1900 | CODE: |
3105 | CODE: |
1901 | clSetUserEventStatus (self, execution_status); |
3106 | clSetUserEventStatus (self, execution_status); |
1902 | |
3107 | |
|
|
3108 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
3109 | |
|
|
3110 | void |
|
|
3111 | DESTROY (SV *self) |
|
|
3112 | CODE: |
|
|
3113 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
3114 | |
|
|
3115 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
3116 | mapped_detach (self, mapped); |
|
|
3117 | |
|
|
3118 | clReleaseCommandQueue (mapped->queue); |
|
|
3119 | clReleaseEvent (mapped->event); |
|
|
3120 | Safefree (mapped); |
|
|
3121 | |
|
|
3122 | void |
|
|
3123 | unmap (OpenCL::Mapped self, ...) |
|
|
3124 | CODE: |
|
|
3125 | mapped_unmap (cv, ST (0), self, self->queue, &ST (1), items - 1); |
|
|
3126 | |
|
|
3127 | bool |
|
|
3128 | mapped (OpenCL::Mapped self) |
|
|
3129 | CODE: |
|
|
3130 | RETVAL = !!self->ptr; |
|
|
3131 | OUTPUT: |
|
|
3132 | RETVAL |
|
|
3133 | |
|
|
3134 | void |
|
|
3135 | wait (OpenCL::Mapped self) |
|
|
3136 | PPCODE: |
|
|
3137 | if (self->event) |
|
|
3138 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
3139 | |
|
|
3140 | void |
|
|
3141 | event (OpenCL::Mapped self) |
|
|
3142 | PPCODE: |
|
|
3143 | if (!self->event) |
|
|
3144 | XSRETURN_UNDEF; |
|
|
3145 | |
|
|
3146 | clRetainEvent (self->event); |
|
|
3147 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
3148 | |
|
|
3149 | size_t |
|
|
3150 | size (OpenCL::Mapped self) |
|
|
3151 | CODE: |
|
|
3152 | RETVAL = self->cb; |
|
|
3153 | OUTPUT: |
|
|
3154 | RETVAL |
|
|
3155 | |
|
|
3156 | IV |
|
|
3157 | ptr (OpenCL::Mapped self) |
|
|
3158 | CODE: |
|
|
3159 | RETVAL = PTR2IV (self->ptr); |
|
|
3160 | OUTPUT: |
|
|
3161 | RETVAL |
|
|
3162 | |
|
|
3163 | void |
|
|
3164 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
3165 | CODE: |
|
|
3166 | STRLEN len; |
|
|
3167 | const char *ptr = SvPVbyte (data, len); |
|
|
3168 | |
|
|
3169 | if (offset + len > self->cb) |
|
|
3170 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
3171 | |
|
|
3172 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
3173 | |
|
|
3174 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
3175 | |
|
|
3176 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
3177 | |
|
|
3178 | IV |
|
|
3179 | row_pitch (OpenCL::Mapped self) |
|
|
3180 | ALIAS: |
|
|
3181 | slice_pitch = 1 |
|
|
3182 | CODE: |
|
|
3183 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
3184 | OUTPUT: |
|
|
3185 | RETVAL |
|
|
3186 | |
|
|
3187 | |
|
|
3188 | |