1 | #include "EXTERN.h" |
1 | #include "EXTERN.h" |
2 | #include "perl.h" |
2 | #include "perl.h" |
3 | #include "XSUB.h" |
3 | #include "XSUB.h" |
|
|
4 | |
|
|
5 | #define X_STACKSIZE sizeof (void *) * 512 * 1024 // 2-4mb should be enough, really |
|
|
6 | #include "xthread.h" |
|
|
7 | #include "schmorp.h" |
|
|
8 | |
|
|
9 | #ifdef I_DLFCN |
|
|
10 | #include <dlfcn.h> |
|
|
11 | #endif |
|
|
12 | |
|
|
13 | // how stupid is that, the 1.2 header files define CL_VERSION_1_1, |
|
|
14 | // but then fail to define the api functions unless you ALSO define |
|
|
15 | // this. This breaks 100% of the opencl 1.1 apps, for what reason? |
|
|
16 | // after all, the functions are deprecated, not removed. |
|
|
17 | // in addition, you cannot test for this in any future-proof way. |
|
|
18 | // each time a new opencl version comes out, you need to make a new |
|
|
19 | // release. |
|
|
20 | #define CL_USE_DEPRECATED_OPENCL_1_2_APIS /* just guessing, you stupid idiots */ |
|
|
21 | |
|
|
22 | #ifndef PREFER_1_1 |
|
|
23 | #define PREFER_1_1 1 |
|
|
24 | #endif |
|
|
25 | |
|
|
26 | #if PREFER_1_1 |
|
|
27 | #define CL_USE_DEPRECATED_OPENCL_1_1_APIS |
|
|
28 | #endif |
4 | |
29 | |
5 | #ifdef __APPLE__ |
30 | #ifdef __APPLE__ |
6 | #include <OpenCL/opencl.h> |
31 | #include <OpenCL/opencl.h> |
7 | #else |
32 | #else |
8 | #include <CL/opencl.h> |
33 | #include <CL/opencl.h> |
9 | #endif |
34 | #endif |
10 | #if HAVE_OPENGL |
35 | |
11 | #ifdef __APPLE__ |
36 | #ifndef CL_VERSION_1_2 |
12 | #include <OpenGL/gl.h> |
37 | #undef PREFER_1_1 |
13 | #include <OpenCL/cl_gl.h> |
38 | #define PREFER_1_1 1 |
14 | #else |
|
|
15 | #include <GL/gl.h> |
|
|
16 | #include <CL/cl_gl.h> |
|
|
17 | #endif |
|
|
18 | #endif |
39 | #endif |
19 | |
40 | |
20 | typedef cl_platform_id OpenCL__Platform; |
41 | typedef cl_platform_id OpenCL__Platform; |
21 | typedef cl_device_id OpenCL__Device; |
42 | typedef cl_device_id OpenCL__Device; |
22 | typedef cl_context OpenCL__Context; |
43 | typedef cl_context OpenCL__Context; |
… | |
… | |
36 | typedef cl_program OpenCL__Program; |
57 | typedef cl_program OpenCL__Program; |
37 | typedef cl_kernel OpenCL__Kernel; |
58 | typedef cl_kernel OpenCL__Kernel; |
38 | typedef cl_event OpenCL__Event; |
59 | typedef cl_event OpenCL__Event; |
39 | typedef cl_event OpenCL__UserEvent; |
60 | typedef cl_event OpenCL__UserEvent; |
40 | |
61 | |
|
|
62 | typedef struct mapped * OpenCL__Mapped; |
|
|
63 | |
41 | typedef SV *FUTURE; |
64 | typedef SV *FUTURE; |
|
|
65 | |
|
|
66 | static HV |
|
|
67 | *stash_platform, |
|
|
68 | *stash_device, |
|
|
69 | *stash_context, |
|
|
70 | *stash_queue, |
|
|
71 | *stash_program, |
|
|
72 | *stash_kernel, |
|
|
73 | *stash_sampler, |
|
|
74 | *stash_event, |
|
|
75 | *stash_userevent, |
|
|
76 | *stash_memory, |
|
|
77 | *stash_buffer, |
|
|
78 | *stash_bufferobj, |
|
|
79 | *stash_image, |
|
|
80 | *stash_image1d, |
|
|
81 | *stash_image1darray, |
|
|
82 | *stash_image1dbuffer, |
|
|
83 | *stash_image2d, |
|
|
84 | *stash_image2darray, |
|
|
85 | *stash_image3d, |
|
|
86 | *stash_mapped, |
|
|
87 | *stash_mappedbuffer, |
|
|
88 | *stash_mappedimage; |
|
|
89 | |
|
|
90 | /*****************************************************************************/ |
|
|
91 | |
|
|
92 | // name must include a leading underscore |
|
|
93 | // all of this horrors would be unneceesary if somebody wrote a proper OpenGL module |
|
|
94 | // for perl. doh. |
|
|
95 | static void * |
|
|
96 | glsym (const char *name) |
|
|
97 | { |
|
|
98 | void *fun = 0; |
|
|
99 | |
|
|
100 | #if defined I_DLFCN && defined RTLD_DEFAULT |
|
|
101 | fun = dlsym (RTLD_DEFAULT, name + 1); |
|
|
102 | if (!fun) fun = dlsym (RTLD_DEFAULT, name); |
|
|
103 | |
|
|
104 | if (!fun) |
|
|
105 | { |
|
|
106 | static void *libgl; |
|
|
107 | static const char *glso[] = { |
|
|
108 | "libGL.so.1", |
|
|
109 | "libGL.so.3", |
|
|
110 | "libGL.so.4.0", |
|
|
111 | "libGL.so", |
|
|
112 | "/usr/lib/libGL.so", |
|
|
113 | "/usr/X11R6/lib/libGL.1.dylib" |
|
|
114 | }; |
|
|
115 | int i; |
|
|
116 | |
|
|
117 | for (i = 0; !libgl && i < sizeof (glso) / sizeof (glso [0]); ++i) |
|
|
118 | { |
|
|
119 | libgl = dlopen (glso [i], RTLD_LAZY); |
|
|
120 | if (libgl) |
|
|
121 | break; |
|
|
122 | } |
|
|
123 | |
|
|
124 | if (libgl) |
|
|
125 | { |
|
|
126 | fun = dlsym (libgl, name + 1); |
|
|
127 | if (!fun) fun = dlsym (libgl, name); |
|
|
128 | } |
|
|
129 | } |
|
|
130 | #endif |
|
|
131 | |
|
|
132 | return fun; |
|
|
133 | } |
42 | |
134 | |
43 | /*****************************************************************************/ |
135 | /*****************************************************************************/ |
44 | |
136 | |
45 | /* up to two temporary buffers */ |
137 | /* up to two temporary buffers */ |
46 | static void * |
138 | static void * |
… | |
… | |
49 | enum { buffers = 3 }; |
141 | enum { buffers = 3 }; |
50 | static int idx; |
142 | static int idx; |
51 | static void *buf [buffers]; |
143 | static void *buf [buffers]; |
52 | static size_t len [buffers]; |
144 | static size_t len [buffers]; |
53 | |
145 | |
54 | idx = ++idx % buffers; |
146 | idx = (idx + 1) % buffers; |
55 | |
147 | |
56 | if (len [idx] < size) |
148 | if (len [idx] < size) |
57 | { |
149 | { |
58 | free (buf [idx]); |
150 | free (buf [idx]); |
59 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
151 | len [idx] = ((size + 31) & ~4095) + 4096 - 32; |
… | |
… | |
127 | if (res) \ |
219 | if (res) \ |
128 | FAIL (name); |
220 | FAIL (name); |
129 | |
221 | |
130 | /*****************************************************************************/ |
222 | /*****************************************************************************/ |
131 | |
223 | |
132 | #define NEW_MORTAL_OBJ(class,ptr) sv_setref_pv (sv_newmortal (), class, ptr) |
|
|
133 | #define XPUSH_NEW_OBJ(class,ptr) XPUSHs (NEW_MORTAL_OBJ (class, ptr)) |
|
|
134 | |
|
|
135 | static void * |
|
|
136 | SvPTROBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
137 | { |
|
|
138 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
139 | return (void *)SvIV (SvRV (sv)); |
|
|
140 | |
|
|
141 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
142 | } |
|
|
143 | |
|
|
144 | /*****************************************************************************/ |
|
|
145 | |
|
|
146 | static cl_context_properties * |
224 | static cl_context_properties * |
147 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
225 | SvCONTEXTPROPERTIES (const char *func, const char *svname, SV *sv, cl_context_properties *extra, int extracount) |
148 | { |
226 | { |
149 | if (!sv || !SvOK (sv)) |
227 | if (!sv || !SvOK (sv)) |
150 | if (extra) |
228 | if (extra) |
… | |
… | |
153 | return 0; |
231 | return 0; |
154 | |
232 | |
155 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
233 | if (SvROK (sv) && SvTYPE (SvRV (sv)) == SVt_PVAV) |
156 | { |
234 | { |
157 | AV *av = (AV *)SvRV (sv); |
235 | AV *av = (AV *)SvRV (sv); |
158 | int i, len = av_len (av); |
236 | int i, len = av_len (av) + 1; |
159 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
237 | cl_context_properties *p = tmpbuf (sizeof (cl_context_properties) * (len + extracount + 1)); |
160 | cl_context_properties *l = p; |
238 | cl_context_properties *l = p; |
161 | |
239 | |
162 | if (len & 1) |
240 | if (len & 1) |
163 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
241 | croak ("%s: %s is not a property list (must be even number of elements)", func, svname); |
… | |
… | |
165 | while (extracount--) |
243 | while (extracount--) |
166 | *l++ = *extra++; |
244 | *l++ = *extra++; |
167 | |
245 | |
168 | for (i = 0; i < len; i += 2) |
246 | for (i = 0; i < len; i += 2) |
169 | { |
247 | { |
170 | cl_context_properties t = SvIV (*av_fetch (av, i, 0)); |
248 | cl_context_properties t = SvIV (*av_fetch (av, i , 0)); |
171 | cl_context_properties v; |
249 | SV *p_sv = *av_fetch (av, i + 1, 0); |
|
|
250 | cl_context_properties v = SvIV (p_sv); // code below can override |
172 | |
251 | |
173 | switch (t) |
252 | switch (t) |
174 | { |
253 | { |
|
|
254 | case CL_GLX_DISPLAY_KHR: |
|
|
255 | if (!SvOK (p_sv)) |
|
|
256 | { |
|
|
257 | void *func = glsym ("_glXGetCurrentDisplay"); |
|
|
258 | if (func) |
|
|
259 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
260 | } |
|
|
261 | break; |
|
|
262 | |
|
|
263 | case CL_GL_CONTEXT_KHR: |
|
|
264 | if (!SvOK (p_sv)) |
|
|
265 | { |
|
|
266 | void *func = glsym ("_glXGetCurrentContext"); |
|
|
267 | if (func) |
|
|
268 | v = (cl_context_properties)((void *(*)(void))func)(); |
|
|
269 | } |
|
|
270 | break; |
|
|
271 | |
175 | default: |
272 | default: |
176 | /* unknown property, treat as int */ |
273 | /* unknown property, treat as int */ |
177 | v = SvIV (*av_fetch (av, i, 0)); |
|
|
178 | break; |
274 | break; |
179 | } |
275 | } |
180 | |
276 | |
181 | *l++ = t; |
277 | *l++ = t; |
182 | *l++ = v; |
278 | *l++ = v; |
… | |
… | |
190 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
286 | croak ("%s: %s is not a property list (either undef or [type => value, ...])", func, svname); |
191 | } |
287 | } |
192 | |
288 | |
193 | /*****************************************************************************/ |
289 | /*****************************************************************************/ |
194 | |
290 | |
|
|
291 | static SV * |
|
|
292 | new_clobj (HV *stash, IV id) |
|
|
293 | { |
|
|
294 | return sv_2mortal (sv_bless (newRV_noinc (newSViv (id)), stash)); |
|
|
295 | } |
|
|
296 | |
|
|
297 | #define PUSH_CLOBJ(stash,id) PUSHs (new_clobj ((stash), (IV)(id))) |
|
|
298 | #define XPUSH_CLOBJ(stash,id) XPUSHs (new_clobj ((stash), (IV)(id))) |
|
|
299 | |
|
|
300 | /* cl objects are either \$iv, or [$iv, ...] */ |
|
|
301 | /* they can be upgraded at runtime to the array form */ |
|
|
302 | static void * |
|
|
303 | SvCLOBJ (const char *func, const char *svname, SV *sv, const char *pkg) |
|
|
304 | { |
|
|
305 | if (SvROK (sv) && sv_derived_from (sv, pkg)) |
|
|
306 | return (void *)SvIV (SvRV (sv)); |
|
|
307 | |
|
|
308 | croak ("%s: %s is not of type %s", func, svname, pkg); |
|
|
309 | } |
|
|
310 | |
|
|
311 | /*****************************************************************************/ |
|
|
312 | /* callback stuff */ |
|
|
313 | |
|
|
314 | /* default context callback, log to stderr */ |
|
|
315 | static void CL_CALLBACK |
|
|
316 | context_default_notify (const char *msg, const void *info, size_t cb, void *data) |
|
|
317 | { |
|
|
318 | fprintf (stderr, "OpenCL Context Notify: %s\n", msg); |
|
|
319 | } |
|
|
320 | |
|
|
321 | typedef struct |
|
|
322 | { |
|
|
323 | int free_cb; |
|
|
324 | void (*push)(void *data1, void *data2, void *data3); |
|
|
325 | } eq_vtbl; |
|
|
326 | |
|
|
327 | typedef struct eq_item |
|
|
328 | { |
|
|
329 | struct eq_item *next; |
|
|
330 | eq_vtbl *vtbl; |
|
|
331 | SV *cb; |
|
|
332 | void *data1, *data2, *data3; |
|
|
333 | } eq_item; |
|
|
334 | |
|
|
335 | static void (*eq_signal_func)(void *signal_arg, int value); |
|
|
336 | static void *eq_signal_arg; |
|
|
337 | static xmutex_t eq_lock = X_MUTEX_INIT; |
|
|
338 | static eq_item *eq_head, *eq_tail; |
|
|
339 | |
|
|
340 | static void |
|
|
341 | eq_enq (eq_vtbl *vtbl, SV *cb, void *data1, void *data2, void *data3) |
|
|
342 | { |
|
|
343 | eq_item *item = malloc (sizeof (eq_item)); |
|
|
344 | |
|
|
345 | item->next = 0; |
|
|
346 | item->vtbl = vtbl; |
|
|
347 | item->cb = cb; |
|
|
348 | item->data1 = data1; |
|
|
349 | item->data2 = data2; |
|
|
350 | item->data3 = data3; |
|
|
351 | |
|
|
352 | X_LOCK (eq_lock); |
|
|
353 | |
|
|
354 | *(eq_head ? &eq_tail->next : &eq_head) = item; |
|
|
355 | eq_tail = item; |
|
|
356 | |
|
|
357 | X_UNLOCK (eq_lock); |
|
|
358 | |
|
|
359 | eq_signal_func (eq_signal_arg, 0); |
|
|
360 | } |
|
|
361 | |
|
|
362 | static eq_item * |
|
|
363 | eq_dec (void) |
|
|
364 | { |
|
|
365 | eq_item *res; |
|
|
366 | |
|
|
367 | X_LOCK (eq_lock); |
|
|
368 | |
|
|
369 | res = eq_head; |
|
|
370 | if (res) |
|
|
371 | eq_head = res->next; |
|
|
372 | |
|
|
373 | X_UNLOCK (eq_lock); |
|
|
374 | |
|
|
375 | return res; |
|
|
376 | } |
|
|
377 | |
|
|
378 | static void |
|
|
379 | eq_poll (void) |
|
|
380 | { |
|
|
381 | eq_item *item; |
|
|
382 | |
|
|
383 | while ((item = eq_dec ())) |
|
|
384 | { |
|
|
385 | ENTER; |
|
|
386 | SAVETMPS; |
|
|
387 | |
|
|
388 | dSP; |
|
|
389 | PUSHMARK (SP); |
|
|
390 | EXTEND (SP, 2); |
|
|
391 | |
|
|
392 | if (item->vtbl->free_cb) |
|
|
393 | sv_2mortal (item->cb); |
|
|
394 | |
|
|
395 | PUTBACK; |
|
|
396 | item->vtbl->push (item->data1, item->data2, item->data3); |
|
|
397 | |
|
|
398 | SV *cb = item->cb; |
|
|
399 | free (item); |
|
|
400 | |
|
|
401 | call_sv (cb, G_DISCARD | G_VOID); |
|
|
402 | |
|
|
403 | FREETMPS; |
|
|
404 | LEAVE; |
|
|
405 | } |
|
|
406 | } |
|
|
407 | |
|
|
408 | static void |
|
|
409 | eq_poll_interrupt (pTHX_ void *c_arg, int value) |
|
|
410 | { |
|
|
411 | eq_poll (); |
|
|
412 | } |
|
|
413 | |
|
|
414 | /*****************************************************************************/ |
|
|
415 | /* context notify */ |
|
|
416 | |
|
|
417 | static void |
|
|
418 | eq_context_push (void *data1, void *data2, void *data3) |
|
|
419 | { |
|
|
420 | dSP; |
|
|
421 | PUSHs (sv_2mortal (newSVpv (data1, 0))); |
|
|
422 | PUSHs (sv_2mortal (newSVpvn (data2, (STRLEN)data3))); |
|
|
423 | PUTBACK; |
|
|
424 | |
|
|
425 | free (data1); |
|
|
426 | free (data2); |
|
|
427 | } |
|
|
428 | |
|
|
429 | static eq_vtbl eq_context_vtbl = { 0, eq_context_push }; |
|
|
430 | |
|
|
431 | static void CL_CALLBACK |
|
|
432 | eq_context_notify (const char *msg, const void *pvt, size_t cb, void *user_data) |
|
|
433 | { |
|
|
434 | void *pvt_copy = malloc (cb); |
|
|
435 | memcpy (pvt_copy, pvt, cb); |
|
|
436 | eq_enq (&eq_context_vtbl, user_data, strdup (msg), pvt_copy, (void *)cb); |
|
|
437 | } |
|
|
438 | |
|
|
439 | #define CONTEXT_NOTIFY_CALLBACK \ |
|
|
440 | void (CL_CALLBACK *pfn_notify)(const char *, const void *, size_t, void *) = context_default_notify; \ |
|
|
441 | void *user_data = 0; \ |
|
|
442 | \ |
|
|
443 | if (SvOK (notify)) \ |
|
|
444 | { \ |
|
|
445 | pfn_notify = eq_context_notify; \ |
|
|
446 | user_data = s_get_cv (notify); \ |
|
|
447 | } |
|
|
448 | |
|
|
449 | static SV * |
|
|
450 | new_clobj_context (cl_context ctx, void *user_data) |
|
|
451 | { |
|
|
452 | SV *sv = new_clobj (stash_context, (IV)ctx); |
|
|
453 | |
|
|
454 | if (user_data) |
|
|
455 | sv_magicext (SvRV (sv), user_data, PERL_MAGIC_ext, 0, 0, 0); |
|
|
456 | |
|
|
457 | return sv; |
|
|
458 | } |
|
|
459 | |
|
|
460 | #define XPUSH_CLOBJ_CONTEXT XPUSHs (new_clobj_context (ctx, user_data)); |
|
|
461 | |
|
|
462 | /*****************************************************************************/ |
|
|
463 | /* build/compile/link notify */ |
|
|
464 | |
|
|
465 | static void |
|
|
466 | eq_program_push (void *data1, void *data2, void *data3) |
|
|
467 | { |
|
|
468 | dSP; |
|
|
469 | PUSH_CLOBJ (stash_program, data1); |
|
|
470 | PUTBACK; |
|
|
471 | } |
|
|
472 | |
|
|
473 | static eq_vtbl eq_program_vtbl = { 1, eq_program_push }; |
|
|
474 | |
|
|
475 | static void CL_CALLBACK |
|
|
476 | eq_program_notify (cl_program program, void *user_data) |
|
|
477 | { |
|
|
478 | eq_enq (&eq_program_vtbl, user_data, (void *)program, 0, 0); |
|
|
479 | } |
|
|
480 | |
|
|
481 | struct build_args |
|
|
482 | { |
|
|
483 | cl_program program; |
|
|
484 | char *options; |
|
|
485 | void *user_data; |
|
|
486 | cl_uint num_devices; |
|
|
487 | }; |
|
|
488 | |
|
|
489 | X_THREAD_PROC (build_program_thread) |
|
|
490 | { |
|
|
491 | struct build_args *arg = thr_arg; |
|
|
492 | |
|
|
493 | clBuildProgram (arg->program, arg->num_devices, arg->num_devices ? (void *)(arg + 1) : 0, arg->options, 0, 0); |
|
|
494 | |
|
|
495 | if (arg->user_data) |
|
|
496 | eq_program_notify (arg->program, arg->user_data); |
|
|
497 | else |
|
|
498 | clReleaseProgram (arg->program); |
|
|
499 | |
|
|
500 | free (arg->options); |
|
|
501 | free (arg); |
|
|
502 | } |
|
|
503 | |
|
|
504 | static void |
|
|
505 | build_program_async (cl_program program, cl_uint num_devices, const cl_device_id *device_list, const char *options, void *user_data) |
|
|
506 | { |
|
|
507 | struct build_args *arg = malloc (sizeof (struct build_args) + sizeof (*device_list) * num_devices); |
|
|
508 | |
|
|
509 | arg->program = program; |
|
|
510 | arg->options = strdup (options); |
|
|
511 | arg->user_data = user_data; |
|
|
512 | arg->num_devices = num_devices; |
|
|
513 | memcpy (arg + 1, device_list, sizeof (*device_list) * num_devices); |
|
|
514 | |
|
|
515 | xthread_t id; |
|
|
516 | thread_create (&id, build_program_thread, arg); |
|
|
517 | } |
|
|
518 | |
|
|
519 | /*****************************************************************************/ |
|
|
520 | /* event objects */ |
|
|
521 | |
|
|
522 | static void |
|
|
523 | eq_event_push (void *data1, void *data2, void *data3) |
|
|
524 | { |
|
|
525 | dSP; |
|
|
526 | PUSH_CLOBJ (stash_event, data1); |
|
|
527 | PUSHs (sv_2mortal (newSViv ((IV)data2))); |
|
|
528 | PUTBACK; |
|
|
529 | } |
|
|
530 | |
|
|
531 | static eq_vtbl eq_event_vtbl = { 1, eq_event_push }; |
|
|
532 | |
|
|
533 | static void CL_CALLBACK |
|
|
534 | eq_event_notify (cl_event event, cl_int event_command_exec_status, void *user_data) |
|
|
535 | { |
|
|
536 | clRetainEvent (event); |
|
|
537 | eq_enq (&eq_event_vtbl, user_data, (void *)event, (void *)(IV)event_command_exec_status, 0); |
|
|
538 | } |
|
|
539 | |
|
|
540 | /*****************************************************************************/ |
|
|
541 | /* utilities for XS code */ |
|
|
542 | |
195 | static size_t |
543 | static size_t |
196 | img_row_pitch (cl_mem img) |
544 | img_row_pitch (cl_mem img) |
197 | { |
545 | { |
198 | size_t res; |
546 | size_t res; |
199 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
547 | clGetImageInfo (img, CL_IMAGE_ROW_PITCH, sizeof (res), &res, 0); |
200 | return res; |
548 | return res; |
201 | } |
549 | } |
202 | |
550 | |
203 | static cl_event * |
551 | static cl_event * |
204 | event_list (SV **items, int count) |
552 | event_list (SV **items, cl_uint *rcount, cl_event extra) |
205 | { |
553 | { |
206 | if (!count) |
554 | cl_uint count = *rcount; |
|
|
555 | |
|
|
556 | if (count > 0x7fffffff) // yeha, it's a hack - the caller might have underflowed |
|
|
557 | count = 0; |
|
|
558 | |
|
|
559 | if (!count && !extra) |
207 | return 0; |
560 | return 0; |
208 | |
561 | |
209 | cl_event *list = tmpbuf (sizeof (cl_event) * count); |
562 | cl_event *list = tmpbuf (sizeof (cl_event) * (count + 1)); |
|
|
563 | int i = 0; |
210 | |
564 | |
211 | while (count--) |
565 | while (count--) |
|
|
566 | if (SvOK (items [count])) |
212 | list [count] = SvPTROBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
567 | list [i++] = SvCLOBJ ("clEnqueue", "wait_events", items [count], "OpenCL::Event"); |
213 | |
568 | |
|
|
569 | if (extra) |
|
|
570 | list [i++] = extra; |
|
|
571 | |
|
|
572 | *rcount = i; |
|
|
573 | |
214 | return list; |
574 | return i ? list : 0; |
215 | } |
575 | } |
216 | |
576 | |
217 | #define EVENT_LIST(items,count) \ |
577 | #define EVENT_LIST(skip) \ |
218 | cl_uint event_list_count = (count); \ |
578 | cl_uint event_list_count = items - (skip); \ |
219 | cl_event *event_list_ptr = event_list (&ST (items), event_list_count) |
579 | cl_event *event_list_ptr = event_list (&ST (skip), &event_list_count, 0) |
220 | |
580 | |
221 | #define INFO(class) \ |
581 | #define INFO(class) \ |
222 | { \ |
582 | { \ |
223 | size_t size; \ |
583 | size_t size; \ |
224 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
584 | NEED_SUCCESS (Get ## class ## Info, (self, name, 0, 0, &size)); \ |
… | |
… | |
228 | SvCUR_set (sv, size); \ |
588 | SvCUR_set (sv, size); \ |
229 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
589 | NEED_SUCCESS (Get ## class ## Info, (self, name, size, SvPVX (sv), 0)); \ |
230 | XPUSHs (sv); \ |
590 | XPUSHs (sv); \ |
231 | } |
591 | } |
232 | |
592 | |
|
|
593 | /*****************************************************************************/ |
|
|
594 | /* mapped_xxx */ |
|
|
595 | |
|
|
596 | static OpenCL__Mapped |
|
|
597 | SvMAPPED (SV *sv) |
|
|
598 | { |
|
|
599 | // no typechecking atm., keep your fingers crossed |
|
|
600 | return (OpenCL__Mapped)SvMAGIC (SvRV (sv))->mg_ptr; |
|
|
601 | } |
|
|
602 | |
|
|
603 | struct mapped |
|
|
604 | { |
|
|
605 | cl_command_queue queue; |
|
|
606 | cl_mem memobj; |
|
|
607 | void *ptr; |
|
|
608 | size_t cb; |
|
|
609 | cl_event event; |
|
|
610 | size_t row_pitch; |
|
|
611 | size_t slice_pitch; |
|
|
612 | }; |
|
|
613 | |
|
|
614 | static SV * |
|
|
615 | 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) |
|
|
616 | { |
|
|
617 | SV *data = newSV (0); |
|
|
618 | SvUPGRADE (data, SVt_PVMG); |
|
|
619 | |
|
|
620 | OpenCL__Mapped mapped; |
|
|
621 | New (0, mapped, 1, struct mapped); |
|
|
622 | |
|
|
623 | clRetainCommandQueue (queue); |
|
|
624 | |
|
|
625 | mapped->queue = queue; |
|
|
626 | mapped->memobj = memobj; |
|
|
627 | mapped->ptr = ptr; |
|
|
628 | mapped->cb = cb; |
|
|
629 | mapped->event = ev; |
|
|
630 | mapped->row_pitch = row_pitch; |
|
|
631 | mapped->slice_pitch = slice_pitch; |
|
|
632 | |
|
|
633 | sv_magicext (data, 0, PERL_MAGIC_ext, 0, (char *)mapped, 0); |
|
|
634 | |
|
|
635 | if (SvLEN (data)) |
|
|
636 | Safefree (data); |
|
|
637 | |
|
|
638 | SvPVX (data) = (char *)ptr; |
|
|
639 | SvCUR_set (data, cb); |
|
|
640 | SvLEN_set (data, 0); |
|
|
641 | SvPOK_only (data); |
|
|
642 | |
|
|
643 | SV *obj = sv_2mortal (sv_bless (newRV_noinc (data), stash)); |
|
|
644 | |
|
|
645 | if (!(flags & CL_MAP_WRITE)) |
|
|
646 | SvREADONLY_on (data); |
|
|
647 | |
|
|
648 | return obj; |
|
|
649 | } |
|
|
650 | |
|
|
651 | static void |
|
|
652 | mapped_detach (SV *sv, OpenCL__Mapped mapped) |
|
|
653 | { |
|
|
654 | SV *data = SvRV (sv); |
|
|
655 | |
|
|
656 | // the next check checks both before AND after detach, where SvPVX should STILL be 0 |
|
|
657 | if (SvPVX (data) != (char *)mapped->ptr) |
|
|
658 | warn ("FATAL: OpenCL memory mapped scalar changed location, detected"); |
|
|
659 | else |
|
|
660 | { |
|
|
661 | SvREADONLY_off (data); |
|
|
662 | SvCUR_set (data, 0); |
|
|
663 | SvPVX (data) = 0; |
|
|
664 | SvOK_off (data); |
|
|
665 | } |
|
|
666 | |
|
|
667 | mapped->ptr = 0; |
|
|
668 | } |
|
|
669 | |
|
|
670 | static void |
|
|
671 | mapped_unmap (SV *self, OpenCL__Mapped mapped, cl_command_queue queue, SV **wait_list, cl_uint event_list_count) |
|
|
672 | { |
|
|
673 | cl_event *event_list_ptr = event_list (wait_list, &event_list_count, mapped->event); |
|
|
674 | cl_event ev; |
|
|
675 | |
|
|
676 | NEED_SUCCESS (EnqueueUnmapMemObject, (queue, mapped->memobj, mapped->ptr, event_list_count, event_list_ptr, &ev)); |
|
|
677 | |
|
|
678 | clReleaseEvent (mapped->event); |
|
|
679 | mapped->event = ev; |
|
|
680 | |
|
|
681 | mapped_detach (self, mapped); |
|
|
682 | } |
|
|
683 | |
|
|
684 | /*****************************************************************************/ |
|
|
685 | |
233 | MODULE = OpenCL PACKAGE = OpenCL |
686 | MODULE = OpenCL PACKAGE = OpenCL |
234 | |
687 | |
235 | PROTOTYPES: ENABLE |
688 | PROTOTYPES: ENABLE |
|
|
689 | |
|
|
690 | void |
|
|
691 | poll () |
|
|
692 | CODE: |
|
|
693 | eq_poll (); |
|
|
694 | |
|
|
695 | void |
|
|
696 | _eq_initialise (IV func, IV arg) |
|
|
697 | CODE: |
|
|
698 | eq_signal_func = (void (*)(void *, int))func; |
|
|
699 | eq_signal_arg = (void*)arg; |
236 | |
700 | |
237 | BOOT: |
701 | BOOT: |
238 | { |
702 | { |
239 | HV *stash = gv_stashpv ("OpenCL", 1); |
703 | HV *stash = gv_stashpv ("OpenCL", 1); |
240 | static const ivstr *civ, const_iv[] = { |
704 | static const ivstr *civ, const_iv[] = { |
… | |
… | |
247 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
711 | { sizeof (cl_long ), "SIZEOF_LONG" }, |
248 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
712 | { sizeof (cl_ulong ), "SIZEOF_ULONG" }, |
249 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
713 | { sizeof (cl_half ), "SIZEOF_HALF" }, |
250 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
714 | { sizeof (cl_float ), "SIZEOF_FLOAT" }, |
251 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
715 | { sizeof (cl_double), "SIZEOF_DOUBLE" }, |
252 | { HAVE_OPENGL , "HAVE_OPENGL" }, |
|
|
253 | #include "constiv.h" |
716 | #include "constiv.h" |
254 | }; |
717 | }; |
|
|
718 | |
255 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
719 | for (civ = const_iv + sizeof (const_iv) / sizeof (const_iv [0]); civ > const_iv; civ--) |
256 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
720 | newCONSTSUB (stash, (char *)civ[-1].name, newSViv (civ[-1].iv)); |
|
|
721 | |
|
|
722 | stash_platform = gv_stashpv ("OpenCL::Platform", GV_ADD); |
|
|
723 | stash_device = gv_stashpv ("OpenCL::Device", GV_ADD); |
|
|
724 | stash_context = gv_stashpv ("OpenCL::Context", GV_ADD); |
|
|
725 | stash_queue = gv_stashpv ("OpenCL::Queue", GV_ADD); |
|
|
726 | stash_program = gv_stashpv ("OpenCL::Program", GV_ADD); |
|
|
727 | stash_kernel = gv_stashpv ("OpenCL::Kernel", GV_ADD); |
|
|
728 | stash_sampler = gv_stashpv ("OpenCL::Sampler", GV_ADD); |
|
|
729 | stash_event = gv_stashpv ("OpenCL::Event", GV_ADD); |
|
|
730 | stash_userevent = gv_stashpv ("OpenCL::UserEvent", GV_ADD); |
|
|
731 | stash_memory = gv_stashpv ("OpenCL::Memory", GV_ADD); |
|
|
732 | stash_buffer = gv_stashpv ("OpenCL::Buffer", GV_ADD); |
|
|
733 | stash_bufferobj = gv_stashpv ("OpenCL::BufferObj", GV_ADD); |
|
|
734 | stash_image = gv_stashpv ("OpenCL::Image", GV_ADD); |
|
|
735 | stash_image1d = gv_stashpv ("OpenCL::Image1D", GV_ADD); |
|
|
736 | stash_image1darray = gv_stashpv ("OpenCL::Image1DArray", GV_ADD); |
|
|
737 | stash_image1dbuffer = gv_stashpv ("OpenCL::Image1DBuffer", GV_ADD); |
|
|
738 | stash_image2d = gv_stashpv ("OpenCL::Image2D", GV_ADD); |
|
|
739 | stash_image2darray = gv_stashpv ("OpenCL::Image2DArray", GV_ADD); |
|
|
740 | stash_image3d = gv_stashpv ("OpenCL::Image3D", GV_ADD); |
|
|
741 | stash_mapped = gv_stashpv ("OpenCL::Mapped", GV_ADD); |
|
|
742 | stash_mappedbuffer = gv_stashpv ("OpenCL::MappedBuffer", GV_ADD); |
|
|
743 | stash_mappedimage = gv_stashpv ("OpenCL::MappedImage", GV_ADD); |
|
|
744 | |
|
|
745 | sv_setiv (perl_get_sv ("OpenCL::POLL_FUNC", TRUE), (IV)eq_poll_interrupt); |
257 | } |
746 | } |
258 | |
747 | |
259 | cl_int |
748 | cl_int |
260 | errno () |
749 | errno () |
261 | CODE: |
750 | CODE: |
262 | errno = res; |
751 | RETVAL = res; |
|
|
752 | OUTPUT: |
|
|
753 | RETVAL |
263 | |
754 | |
264 | const char * |
755 | const char * |
265 | err2str (cl_int err) |
756 | err2str (cl_int err = res) |
266 | |
757 | |
267 | const char * |
758 | const char * |
268 | enum2str (cl_uint value) |
759 | enum2str (cl_uint value) |
269 | |
760 | |
270 | void |
761 | void |
… | |
… | |
278 | list = tmpbuf (sizeof (*list) * count); |
769 | list = tmpbuf (sizeof (*list) * count); |
279 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
770 | NEED_SUCCESS (GetPlatformIDs, (count, list, 0)); |
280 | |
771 | |
281 | EXTEND (SP, count); |
772 | EXTEND (SP, count); |
282 | for (i = 0; i < count; ++i) |
773 | for (i = 0; i < count; ++i) |
283 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", list [i])); |
774 | PUSH_CLOBJ (stash_platform, (IV)list [i]); |
284 | |
775 | |
285 | void |
776 | void |
286 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
777 | context_from_type (cl_context_properties *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
287 | PPCODE: |
778 | PPCODE: |
|
|
779 | CONTEXT_NOTIFY_CALLBACK; |
288 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
780 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (properties, type, 0, 0, &res)); |
289 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
781 | XPUSH_CLOBJ_CONTEXT; |
290 | |
782 | |
291 | void |
783 | void |
292 | context (cl_context_properties *properties = 0, FUTURE devices, FUTURE notify = 0) |
784 | context (FUTURE properties, FUTURE devices, FUTURE notify) |
293 | PPCODE: |
785 | PPCODE: |
294 | /* der Gipfel der Kunst */ |
786 | /* der Gipfel der Kunst */ |
295 | |
787 | |
296 | void |
788 | void |
297 | wait_for_events (...) |
789 | wait_for_events (...) |
298 | CODE: |
790 | CODE: |
299 | EVENT_LIST (0, items); |
791 | EVENT_LIST (0); |
300 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
792 | NEED_SUCCESS (WaitForEvents, (event_list_count, event_list_ptr)); |
301 | |
793 | |
302 | PROTOTYPES: DISABLE |
794 | PROTOTYPES: DISABLE |
303 | |
795 | |
304 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
796 | MODULE = OpenCL PACKAGE = OpenCL::Platform |
305 | |
797 | |
306 | void |
798 | void |
307 | info (OpenCL::Platform self, cl_platform_info name) |
799 | info (OpenCL::Platform self, cl_platform_info name) |
308 | PPCODE: |
800 | PPCODE: |
309 | INFO (Platform) |
801 | INFO (Platform) |
|
|
802 | |
|
|
803 | void |
|
|
804 | unload_compiler (OpenCL::Platform self) |
|
|
805 | CODE: |
|
|
806 | #if CL_VERSION_1_2 |
|
|
807 | clUnloadPlatformCompiler (self); |
|
|
808 | #endif |
310 | |
809 | |
311 | #BEGIN:platform |
810 | #BEGIN:platform |
312 | |
811 | |
313 | void |
812 | void |
314 | profile (OpenCL::Platform self) |
813 | profile (OpenCL::Platform self) |
… | |
… | |
340 | list = tmpbuf (sizeof (*list) * count); |
839 | list = tmpbuf (sizeof (*list) * count); |
341 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
840 | NEED_SUCCESS (GetDeviceIDs, (self, type, count, list, 0)); |
342 | |
841 | |
343 | EXTEND (SP, count); |
842 | EXTEND (SP, count); |
344 | for (i = 0; i < count; ++i) |
843 | for (i = 0; i < count; ++i) |
345 | PUSHs (sv_setref_pv (sv_newmortal (), "OpenCL::Device", list [i])); |
844 | PUSH_CLOBJ (stash_device, list [i]); |
346 | |
845 | |
347 | void |
846 | void |
348 | context (OpenCL::Platform self, cl_context_properties *properties = 0, SV *devices, FUTURE notify = 0) |
847 | context (OpenCL::Platform self, cl_context_properties *properties, SV *devices, SV *notify = &PL_sv_undef) |
349 | PPCODE: |
848 | PPCODE: |
350 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
849 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
351 | croak ("OpenCL::Platform argument 'device' must be an arrayref with device objects, in call"); |
850 | croak ("OpenCL::Platform::context argument 'device' must be an arrayref with device objects, in call"); |
352 | |
851 | |
353 | AV *av = (AV *)SvRV (devices); |
852 | AV *av = (AV *)SvRV (devices); |
354 | cl_uint num_devices = av_len (av) + 1; |
853 | cl_uint num_devices = av_len (av) + 1; |
355 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
854 | cl_device_id *device_list = tmpbuf (sizeof (cl_device_id) * num_devices); |
|
|
855 | |
356 | int i; |
856 | int i; |
357 | |
|
|
358 | for (i = num_devices; i--; ) |
857 | for (i = num_devices; i--; ) |
359 | device_list [i] = SvPTROBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
858 | device_list [i] = SvCLOBJ ("clCreateContext", "devices", *av_fetch (av, i, 0), "OpenCL::Device"); |
360 | |
859 | |
|
|
860 | CONTEXT_NOTIFY_CALLBACK; |
361 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, 0, 0, &res)); |
861 | NEED_SUCCESS_ARG (cl_context ctx, CreateContext, (properties, num_devices, device_list, pfn_notify, user_data, &res)); |
362 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
862 | XPUSH_CLOBJ_CONTEXT; |
363 | |
863 | |
364 | void |
864 | void |
365 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, FUTURE notify = 0) |
865 | context_from_type (OpenCL::Platform self, SV *properties = 0, cl_device_type type = CL_DEVICE_TYPE_DEFAULT, SV *notify = &PL_sv_undef) |
366 | PPCODE: |
866 | PPCODE: |
367 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
867 | cl_context_properties extra[] = { CL_CONTEXT_PLATFORM, (cl_context_properties)self }; |
368 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
868 | cl_context_properties *props = SvCONTEXTPROPERTIES ("OpenCL::Platform::context_from_type", "properties", properties, extra, 2); |
|
|
869 | |
|
|
870 | CONTEXT_NOTIFY_CALLBACK; |
369 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
871 | NEED_SUCCESS_ARG (cl_context ctx, CreateContextFromType, (props, type, 0, 0, &res)); |
370 | XPUSH_NEW_OBJ ("OpenCL::Context", ctx); |
872 | XPUSH_CLOBJ_CONTEXT; |
371 | |
873 | |
372 | MODULE = OpenCL PACKAGE = OpenCL::Device |
874 | MODULE = OpenCL PACKAGE = OpenCL::Device |
373 | |
875 | |
374 | void |
876 | void |
375 | info (OpenCL::Device self, cl_device_info name) |
877 | info (OpenCL::Device self, cl_device_info name) |
… | |
… | |
383 | PPCODE: |
885 | PPCODE: |
384 | cl_device_type value [1]; |
886 | cl_device_type value [1]; |
385 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
887 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_TYPE, sizeof (value), value, 0)); |
386 | EXTEND (SP, 1); |
888 | EXTEND (SP, 1); |
387 | const int i = 0; |
889 | const int i = 0; |
388 | PUSHs (sv_2mortal (newSViv (value [i]))); |
890 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
389 | |
891 | |
390 | void |
892 | void |
391 | vendor_id (OpenCL::Device self) |
893 | vendor_id (OpenCL::Device self) |
392 | ALIAS: |
894 | ALIAS: |
393 | vendor_id = CL_DEVICE_VENDOR_ID |
895 | vendor_id = CL_DEVICE_VENDOR_ID |
… | |
… | |
414 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
916 | native_vector_width_int = CL_DEVICE_NATIVE_VECTOR_WIDTH_INT |
415 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
917 | native_vector_width_long = CL_DEVICE_NATIVE_VECTOR_WIDTH_LONG |
416 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
918 | native_vector_width_float = CL_DEVICE_NATIVE_VECTOR_WIDTH_FLOAT |
417 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
919 | native_vector_width_double = CL_DEVICE_NATIVE_VECTOR_WIDTH_DOUBLE |
418 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
920 | native_vector_width_half = CL_DEVICE_NATIVE_VECTOR_WIDTH_HALF |
419 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
921 | reference_count_ext = CL_DEVICE_REFERENCE_COUNT_EXT |
420 | PPCODE: |
922 | PPCODE: |
421 | cl_uint value [1]; |
923 | cl_uint value [1]; |
422 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
924 | NEED_SUCCESS (GetDeviceInfo, (self, ix, sizeof (value), value, 0)); |
423 | EXTEND (SP, 1); |
925 | EXTEND (SP, 1); |
424 | const int i = 0; |
926 | const int i = 0; |
… | |
… | |
538 | PPCODE: |
1040 | PPCODE: |
539 | cl_command_queue_properties value [1]; |
1041 | cl_command_queue_properties value [1]; |
540 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1042 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
541 | EXTEND (SP, 1); |
1043 | EXTEND (SP, 1); |
542 | const int i = 0; |
1044 | const int i = 0; |
543 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1045 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
544 | |
1046 | |
545 | void |
1047 | void |
546 | platform (OpenCL::Device self) |
1048 | platform (OpenCL::Device self) |
547 | PPCODE: |
1049 | PPCODE: |
548 | cl_platform_id value [1]; |
1050 | cl_platform_id value [1]; |
549 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
1051 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PLATFORM, sizeof (value), value, 0)); |
550 | EXTEND (SP, 1); |
1052 | EXTEND (SP, 1); |
551 | const int i = 0; |
1053 | const int i = 0; |
552 | { |
1054 | { |
553 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Platform", value [i])); |
1055 | PUSH_CLOBJ (stash_platform, value [i]); |
554 | } |
1056 | } |
555 | |
1057 | |
556 | void |
1058 | void |
557 | name (OpenCL::Device self) |
1059 | name (OpenCL::Device self) |
558 | ALIAS: |
1060 | ALIAS: |
… | |
… | |
577 | cl_device_id value [1]; |
1079 | cl_device_id value [1]; |
578 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
1080 | NEED_SUCCESS (GetDeviceInfo, (self, CL_DEVICE_PARENT_DEVICE_EXT, sizeof (value), value, 0)); |
579 | EXTEND (SP, 1); |
1081 | EXTEND (SP, 1); |
580 | const int i = 0; |
1082 | const int i = 0; |
581 | { |
1083 | { |
582 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1084 | PUSH_CLOBJ (stash_device, value [i]); |
583 | } |
1085 | } |
584 | |
1086 | |
585 | void |
1087 | void |
586 | partition_types_ext (OpenCL::Device self) |
1088 | partition_types_ext (OpenCL::Device self) |
587 | ALIAS: |
1089 | ALIAS: |
… | |
… | |
614 | |
1116 | |
615 | void |
1117 | void |
616 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
1118 | queue (OpenCL::Context self, OpenCL::Device device, cl_command_queue_properties properties = 0) |
617 | PPCODE: |
1119 | PPCODE: |
618 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
1120 | NEED_SUCCESS_ARG (cl_command_queue queue, CreateCommandQueue, (self, device, properties, &res)); |
619 | XPUSH_NEW_OBJ ("OpenCL::Queue", queue); |
1121 | XPUSH_CLOBJ (stash_queue, queue); |
620 | |
1122 | |
621 | void |
1123 | void |
622 | user_event (OpenCL::Context self) |
1124 | user_event (OpenCL::Context self) |
623 | PPCODE: |
1125 | PPCODE: |
624 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
1126 | NEED_SUCCESS_ARG (cl_event ev, CreateUserEvent, (self, &res)); |
625 | XPUSH_NEW_OBJ ("OpenCL::UserEvent", ev); |
1127 | XPUSH_CLOBJ (stash_userevent, ev); |
626 | |
1128 | |
627 | void |
1129 | void |
628 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
1130 | buffer (OpenCL::Context self, cl_mem_flags flags, size_t len) |
629 | PPCODE: |
1131 | PPCODE: |
630 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
1132 | if (flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR)) |
631 | croak ("clCreateBuffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
1133 | croak ("OpenCL::Context::buffer: cannot use/copy host ptr when no data is given, use $context->buffer_sv instead?"); |
632 | |
1134 | |
633 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
1135 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, 0, &res)); |
634 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1136 | XPUSH_CLOBJ (stash_bufferobj, mem); |
635 | |
1137 | |
636 | void |
1138 | void |
637 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
1139 | buffer_sv (OpenCL::Context self, cl_mem_flags flags, SV *data) |
638 | PPCODE: |
1140 | PPCODE: |
639 | STRLEN len; |
1141 | STRLEN len; |
640 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1142 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
641 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
1143 | if (!(flags & (CL_MEM_USE_HOST_PTR | CL_MEM_COPY_HOST_PTR))) |
642 | croak ("clCreateBuffer: have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
1144 | croak ("OpenCL::Context::buffer_sv: you have to specify use or copy host ptr when buffer data is given, use $context->buffer instead?"); |
643 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
1145 | NEED_SUCCESS_ARG (cl_mem mem, CreateBuffer, (self, flags, len, ptr, &res)); |
644 | XPUSH_NEW_OBJ ("OpenCL::BufferObj", mem); |
1146 | XPUSH_CLOBJ (stash_bufferobj, mem); |
645 | |
1147 | |
|
|
1148 | #if CL_VERSION_1_2 |
|
|
1149 | |
646 | void |
1150 | void |
647 | 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) |
1151 | 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) |
648 | PPCODE: |
1152 | PPCODE: |
649 | STRLEN len; |
1153 | STRLEN len; |
650 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1154 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
651 | const cl_image_format format = { channel_order, channel_type }; |
1155 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1156 | const cl_image_desc desc = { |
|
|
1157 | type, |
|
|
1158 | width, height, depth, |
|
|
1159 | array_size, row_pitch, slice_pitch, |
|
|
1160 | num_mip_level, num_samples, |
|
|
1161 | type == CL_MEM_OBJECT_IMAGE1D_BUFFER ? (cl_mem)SvCLOBJ ("OpenCL::Context::Image", "data", data, "OpenCL::Buffer") : 0 |
|
|
1162 | }; |
652 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
1163 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
653 | XPUSH_NEW_OBJ ("OpenCL::Image2D", mem); |
1164 | HV *stash = stash_image; |
|
|
1165 | switch (type) |
|
|
1166 | { |
|
|
1167 | case CL_MEM_OBJECT_IMAGE1D_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1168 | case CL_MEM_OBJECT_IMAGE1D: stash = stash_image1d; break; |
|
|
1169 | case CL_MEM_OBJECT_IMAGE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1170 | case CL_MEM_OBJECT_IMAGE2D: stash = stash_image2d; break; |
|
|
1171 | case CL_MEM_OBJECT_IMAGE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1172 | case CL_MEM_OBJECT_IMAGE3D: stash = stash_image3d; break; |
|
|
1173 | } |
|
|
1174 | XPUSH_CLOBJ (stash, mem); |
654 | |
1175 | |
|
|
1176 | #endif |
|
|
1177 | |
655 | void |
1178 | void |
656 | 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) |
1179 | 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) |
657 | PPCODE: |
1180 | PPCODE: |
658 | STRLEN len; |
1181 | STRLEN len; |
659 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
1182 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
660 | const cl_image_format format = { channel_order, channel_type }; |
1183 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1184 | #if PREFER_1_1 |
|
|
1185 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage2D, (self, flags, &format, width, height, row_pitch, ptr, &res)); |
|
|
1186 | #else |
|
|
1187 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE2D, width, height, 0, 0, row_pitch, 0, 0, 0, 0 }; |
|
|
1188 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1189 | #endif |
|
|
1190 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1191 | |
|
|
1192 | void |
|
|
1193 | 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) |
|
|
1194 | PPCODE: |
|
|
1195 | STRLEN len; |
|
|
1196 | char *ptr = SvOK (data) ? SvPVbyte (data, len) : 0; |
|
|
1197 | const cl_image_format format = { channel_order, channel_type }; |
|
|
1198 | #if PREFER_1_1 |
661 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
1199 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage3D, (self, flags, &format, width, height, depth, row_pitch, slice_pitch, ptr, &res)); |
662 | XPUSH_NEW_OBJ ("OpenCL::Image3D", mem); |
1200 | #else |
|
|
1201 | const cl_image_desc desc = { CL_MEM_OBJECT_IMAGE3D, width, height, depth, 0, row_pitch, slice_pitch, 0, 0, 0 }; |
|
|
1202 | NEED_SUCCESS_ARG (cl_mem mem, CreateImage, (self, flags, &format, &desc, ptr, &res)); |
|
|
1203 | #endif |
|
|
1204 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1205 | |
|
|
1206 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1207 | |
|
|
1208 | void |
|
|
1209 | gl_buffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint bufobj) |
|
|
1210 | PPCODE: |
|
|
1211 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLBuffer, (self, flags, bufobj, &res)); |
|
|
1212 | XPUSH_CLOBJ (stash_bufferobj, mem); |
|
|
1213 | |
|
|
1214 | void |
|
|
1215 | gl_renderbuffer (OpenCL::Context self, cl_mem_flags flags, cl_GLuint renderbuffer) |
|
|
1216 | PPCODE: |
|
|
1217 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLRenderbuffer, (self, flags, renderbuffer, &res)); |
|
|
1218 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1219 | |
|
|
1220 | #if CL_VERSION_1_2 |
|
|
1221 | |
|
|
1222 | void |
|
|
1223 | gl_texture (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1224 | ALIAS: |
|
|
1225 | PPCODE: |
|
|
1226 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture, (self, flags, target, miplevel, texture, &res)); |
|
|
1227 | cl_gl_object_type type; |
|
|
1228 | NEED_SUCCESS (GetGLObjectInfo, (mem, &type, 0)); // TODO: use target instead? |
|
|
1229 | HV *stash = stash_memory; |
|
|
1230 | switch (type) |
|
|
1231 | { |
|
|
1232 | case CL_GL_OBJECT_TEXTURE_BUFFER: stash = stash_image1dbuffer; break; |
|
|
1233 | case CL_GL_OBJECT_TEXTURE1D: stash = stash_image1d; break; |
|
|
1234 | case CL_GL_OBJECT_TEXTURE1D_ARRAY: stash = stash_image2darray; break; |
|
|
1235 | case CL_GL_OBJECT_TEXTURE2D: stash = stash_image2d; break; |
|
|
1236 | case CL_GL_OBJECT_TEXTURE2D_ARRAY: stash = stash_image2darray; break; |
|
|
1237 | case CL_GL_OBJECT_TEXTURE3D: stash = stash_image3d; break; |
|
|
1238 | } |
|
|
1239 | XPUSH_CLOBJ (stash, mem); |
|
|
1240 | |
|
|
1241 | #endif |
|
|
1242 | |
|
|
1243 | void |
|
|
1244 | gl_texture2d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1245 | PPCODE: |
|
|
1246 | #if PREFER_1_1 |
|
|
1247 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture2D, (self, flags, target, miplevel, texture, &res)); |
|
|
1248 | #else |
|
|
1249 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1250 | #endif |
|
|
1251 | XPUSH_CLOBJ (stash_image2d, mem); |
|
|
1252 | |
|
|
1253 | void |
|
|
1254 | gl_texture3d (OpenCL::Context self, cl_mem_flags flags, cl_GLenum target, cl_GLint miplevel, cl_GLuint texture) |
|
|
1255 | PPCODE: |
|
|
1256 | #if PREFER_1_1 |
|
|
1257 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture3D, (self, flags, target, miplevel, texture, &res)); |
|
|
1258 | #else |
|
|
1259 | NEED_SUCCESS_ARG (cl_mem mem, CreateFromGLTexture , (self, flags, target, miplevel, texture, &res)); |
|
|
1260 | #endif |
|
|
1261 | XPUSH_CLOBJ (stash_image3d, mem); |
|
|
1262 | |
|
|
1263 | #endif |
663 | |
1264 | |
664 | void |
1265 | void |
665 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
1266 | supported_image_formats (OpenCL::Context self, cl_mem_flags flags, cl_mem_object_type image_type) |
666 | PPCODE: |
1267 | PPCODE: |
667 | { |
1268 | { |
… | |
… | |
685 | |
1286 | |
686 | void |
1287 | void |
687 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
1288 | sampler (OpenCL::Context self, cl_bool normalized_coords, cl_addressing_mode addressing_mode, cl_filter_mode filter_mode) |
688 | PPCODE: |
1289 | PPCODE: |
689 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
1290 | NEED_SUCCESS_ARG (cl_sampler sampler, CreateSampler, (self, normalized_coords, addressing_mode, filter_mode, &res)); |
690 | XPUSH_NEW_OBJ ("OpenCL::Sampler", sampler); |
1291 | XPUSH_CLOBJ (stash_sampler, sampler); |
691 | |
1292 | |
692 | void |
1293 | void |
693 | program_with_source (OpenCL::Context self, SV *program) |
1294 | program_with_source (OpenCL::Context self, SV *program) |
694 | PPCODE: |
1295 | PPCODE: |
695 | STRLEN len; |
1296 | STRLEN len; |
696 | size_t len2; |
1297 | size_t len2; |
697 | const char *ptr = SvPVbyte (program, len); |
1298 | const char *ptr = SvPVbyte (program, len); |
698 | |
1299 | |
699 | len2 = len; |
1300 | len2 = len; |
700 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
1301 | NEED_SUCCESS_ARG (cl_program prog, CreateProgramWithSource, (self, 1, &ptr, &len2, &res)); |
701 | XPUSH_NEW_OBJ ("OpenCL::Program", prog); |
1302 | XPUSH_CLOBJ (stash_program, prog); |
702 | |
1303 | |
703 | #BEGIN:context |
1304 | #BEGIN:context |
704 | |
1305 | |
705 | void |
1306 | void |
706 | reference_count (OpenCL::Context self) |
1307 | reference_count (OpenCL::Context self) |
… | |
… | |
723 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
1324 | NEED_SUCCESS (GetContextInfo, (self, CL_CONTEXT_DEVICES, size, value, 0)); |
724 | int i, n = size / sizeof (*value); |
1325 | int i, n = size / sizeof (*value); |
725 | EXTEND (SP, n); |
1326 | EXTEND (SP, n); |
726 | for (i = 0; i < n; ++i) |
1327 | for (i = 0; i < n; ++i) |
727 | { |
1328 | { |
728 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1329 | PUSH_CLOBJ (stash_device, value [i]); |
729 | } |
1330 | } |
730 | |
1331 | |
731 | void |
1332 | void |
732 | properties (OpenCL::Context self) |
1333 | properties (OpenCL::Context self) |
733 | PPCODE: |
1334 | PPCODE: |
… | |
… | |
748 | DESTROY (OpenCL::Queue self) |
1349 | DESTROY (OpenCL::Queue self) |
749 | CODE: |
1350 | CODE: |
750 | clReleaseCommandQueue (self); |
1351 | clReleaseCommandQueue (self); |
751 | |
1352 | |
752 | void |
1353 | void |
753 | enqueue_read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
1354 | read_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, size_t len, SV *data, ...) |
|
|
1355 | ALIAS: |
|
|
1356 | enqueue_read_buffer = 0 |
754 | PPCODE: |
1357 | PPCODE: |
755 | cl_event ev = 0; |
1358 | cl_event ev = 0; |
756 | EVENT_LIST (6, items - 6); |
1359 | EVENT_LIST (6); |
757 | |
1360 | |
758 | SvUPGRADE (data, SVt_PV); |
1361 | SvUPGRADE (data, SVt_PV); |
759 | SvGROW (data, len); |
1362 | SvGROW (data, len); |
760 | SvPOK_only (data); |
1363 | SvPOK_only (data); |
761 | SvCUR_set (data, len); |
1364 | SvCUR_set (data, len); |
762 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1365 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
763 | |
1366 | |
764 | if (ev) |
1367 | if (ev) |
765 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1368 | XPUSH_CLOBJ (stash_event, ev); |
766 | |
1369 | |
767 | void |
1370 | void |
768 | enqueue_write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
1371 | write_buffer (OpenCL::Queue self, OpenCL::Buffer mem, cl_bool blocking, size_t offset, SV *data, ...) |
|
|
1372 | ALIAS: |
|
|
1373 | enqueue_write_buffer = 0 |
769 | PPCODE: |
1374 | PPCODE: |
770 | cl_event ev = 0; |
1375 | cl_event ev = 0; |
771 | STRLEN len; |
1376 | STRLEN len; |
772 | char *ptr = SvPVbyte (data, len); |
1377 | char *ptr = SvPVbyte (data, len); |
773 | EVENT_LIST (5, items - 5); |
1378 | EVENT_LIST (5); |
774 | |
1379 | |
775 | NEED_SUCCESS (EnqueueReadBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1380 | NEED_SUCCESS (EnqueueWriteBuffer, (self, mem, blocking, offset, len, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
776 | |
1381 | |
777 | if (ev) |
1382 | if (ev) |
778 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1383 | XPUSH_CLOBJ (stash_event, ev); |
779 | |
1384 | |
|
|
1385 | #if CL_VERSION_1_2 |
|
|
1386 | |
780 | void |
1387 | void |
781 | enqueue_copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
1388 | fill_buffer (OpenCL::Queue self, OpenCL::Buffer mem, SV *data, size_t offset, size_t size, ...) |
|
|
1389 | ALIAS: |
|
|
1390 | enqueue_fill_buffer = 0 |
782 | PPCODE: |
1391 | PPCODE: |
783 | cl_event ev = 0; |
1392 | cl_event ev = 0; |
|
|
1393 | STRLEN len; |
|
|
1394 | char *ptr = SvPVbyte (data, len); |
|
|
1395 | EVENT_LIST (5); |
|
|
1396 | |
|
|
1397 | NEED_SUCCESS (EnqueueFillBuffer, (self, mem, ptr, len, offset, size, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1398 | |
|
|
1399 | if (ev) |
|
|
1400 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1401 | |
|
|
1402 | void |
|
|
1403 | fill_image (OpenCL::Queue self, OpenCL::Image img, NV r, NV g, NV b, NV a, size_t x, size_t y, size_t z, size_t width, size_t height, size_t depth, ...) |
|
|
1404 | ALIAS: |
|
|
1405 | enqueue_fill_image = 0 |
|
|
1406 | PPCODE: |
|
|
1407 | cl_event ev = 0; |
|
|
1408 | STRLEN len; |
|
|
1409 | const size_t origin [3] = { x, y, z }; |
|
|
1410 | const size_t region [3] = { width, height, depth }; |
|
|
1411 | EVENT_LIST (12); |
|
|
1412 | |
|
|
1413 | const cl_float c_f [4] = { r, g, b, a }; |
|
|
1414 | const cl_uint c_u [4] = { r, g, b, a }; |
|
|
1415 | const cl_int c_s [4] = { r, g, b, a }; |
|
|
1416 | const void *c_fus [3] = { &c_f, &c_u, &c_s }; |
|
|
1417 | static const char fus [] = { 0, 0, 0, 0, 0, 0, 0, 2, 2, 2, 1, 1, 1, 0, 0 }; |
|
|
1418 | cl_image_format format; |
|
|
1419 | NEED_SUCCESS (GetImageInfo, (img, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
1420 | assert (sizeof (fus) == CL_FLOAT + 1 - CL_SNORM_INT8); |
|
|
1421 | if (format.image_channel_data_type < CL_SNORM_INT8 || CL_FLOAT < format.image_channel_data_type) |
|
|
1422 | croak ("enqueue_fill_image: image has unsupported channel type, only opencl 1.2 channel types supported."); |
|
|
1423 | |
|
|
1424 | NEED_SUCCESS (EnqueueFillImage, (self, img, c_fus [fus [format.image_channel_data_type - CL_SNORM_INT8]], |
|
|
1425 | origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1426 | |
|
|
1427 | if (ev) |
|
|
1428 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1429 | |
|
|
1430 | #endif |
|
|
1431 | |
|
|
1432 | void |
|
|
1433 | copy_buffer (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_offset, size_t dst_offset, size_t len, ...) |
|
|
1434 | ALIAS: |
|
|
1435 | enqueue_copy_buffer = 0 |
|
|
1436 | PPCODE: |
|
|
1437 | cl_event ev = 0; |
784 | EVENT_LIST (6, items - 6); |
1438 | EVENT_LIST (6); |
785 | |
1439 | |
786 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1440 | NEED_SUCCESS (EnqueueCopyBuffer, (self, src, dst, src_offset, dst_offset, len, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
787 | |
1441 | |
788 | if (ev) |
1442 | if (ev) |
789 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1443 | XPUSH_CLOBJ (stash_event, ev); |
790 | |
1444 | |
791 | void |
1445 | void |
792 | 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, ...) |
1446 | read_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
|
|
1447 | ALIAS: |
|
|
1448 | enqueue_read_buffer_rect = 0 |
793 | PPCODE: |
1449 | PPCODE: |
794 | cl_event ev = 0; |
1450 | cl_event ev = 0; |
795 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1451 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
796 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1452 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
797 | const size_t region[3] = { width, height, depth }; |
1453 | const size_t region[3] = { width, height, depth }; |
798 | EVENT_LIST (17, items - 17); |
1454 | EVENT_LIST (17); |
799 | |
1455 | |
800 | if (!buf_row_pitch) |
1456 | if (!buf_row_pitch) |
801 | buf_row_pitch = region [0]; |
1457 | buf_row_pitch = region [0]; |
802 | |
1458 | |
803 | if (!buf_slice_pitch) |
1459 | if (!buf_slice_pitch) |
… | |
… | |
816 | SvPOK_only (data); |
1472 | SvPOK_only (data); |
817 | SvCUR_set (data, len); |
1473 | SvCUR_set (data, len); |
818 | 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)); |
1474 | NEED_SUCCESS (EnqueueReadBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
819 | |
1475 | |
820 | if (ev) |
1476 | if (ev) |
821 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1477 | XPUSH_CLOBJ (stash_event, ev); |
822 | |
1478 | |
823 | void |
1479 | void |
824 | 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, ...) |
1480 | write_buffer_rect (OpenCL::Queue self, OpenCL::Memory buf, cl_bool blocking, size_t buf_x, size_t buf_y, size_t buf_z, size_t host_x, size_t host_y, size_t host_z, size_t width, size_t height, size_t depth, size_t buf_row_pitch, size_t buf_slice_pitch, size_t host_row_pitch, size_t host_slice_pitch, SV *data, ...) |
|
|
1481 | ALIAS: |
|
|
1482 | enqueue_write_buffer_rect = 0 |
825 | PPCODE: |
1483 | PPCODE: |
826 | cl_event ev = 0; |
1484 | cl_event ev = 0; |
827 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
1485 | const size_t buf_origin [3] = { buf_x , buf_y , buf_z }; |
828 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
1486 | const size_t host_origin[3] = { host_x, host_y, host_z }; |
829 | const size_t region[3] = { width, height, depth }; |
1487 | const size_t region[3] = { width, height, depth }; |
830 | STRLEN len; |
1488 | STRLEN len; |
831 | char *ptr = SvPVbyte (data, len); |
1489 | char *ptr = SvPVbyte (data, len); |
832 | EVENT_LIST (17, items - 17); |
1490 | EVENT_LIST (17); |
833 | |
1491 | |
834 | if (!buf_row_pitch) |
1492 | if (!buf_row_pitch) |
835 | buf_row_pitch = region [0]; |
1493 | buf_row_pitch = region [0]; |
836 | |
1494 | |
837 | if (!buf_slice_pitch) |
1495 | if (!buf_slice_pitch) |
… | |
… | |
846 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
1504 | size_t min_len = host_row_pitch * host_slice_pitch * region [2]; |
847 | |
1505 | |
848 | if (len < min_len) |
1506 | if (len < min_len) |
849 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1507 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
850 | |
1508 | |
851 | 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)); |
1509 | NEED_SUCCESS (EnqueueWriteBufferRect, (self, buf, blocking, buf_origin, host_origin, region, buf_row_pitch, buf_slice_pitch, host_row_pitch, host_slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
852 | |
1510 | |
853 | if (ev) |
1511 | if (ev) |
854 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1512 | XPUSH_CLOBJ (stash_event, ev); |
855 | |
1513 | |
856 | void |
1514 | void |
857 | 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, ...) |
1515 | copy_buffer_rect (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t src_row_pitch, size_t src_slice_pitch, size_t dst_row_pitch, size_t dst_slice_pitch, ...) |
|
|
1516 | ALIAS: |
|
|
1517 | enqueue_copy_buffer_rect = 0 |
858 | PPCODE: |
1518 | PPCODE: |
859 | cl_event ev = 0; |
1519 | cl_event ev = 0; |
860 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1520 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
861 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1521 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
862 | const size_t region[3] = { width, height, depth }; |
1522 | const size_t region[3] = { width, height, depth }; |
863 | EVENT_LIST (16, items - 16); |
1523 | EVENT_LIST (16); |
864 | |
1524 | |
865 | 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)); |
1525 | NEED_SUCCESS (EnqueueCopyBufferRect, (self, src, dst, src_origin, dst_origin, region, src_row_pitch, src_slice_pitch, dst_row_pitch, dst_slice_pitch, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
866 | |
1526 | |
867 | if (ev) |
1527 | if (ev) |
868 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1528 | XPUSH_CLOBJ (stash_event, ev); |
869 | |
1529 | |
870 | void |
1530 | void |
871 | 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, ...) |
1531 | read_image (OpenCL::Queue self, OpenCL::Image src, cl_bool blocking, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
|
|
1532 | ALIAS: |
|
|
1533 | enqueue_read_image = 0 |
872 | PPCODE: |
1534 | PPCODE: |
873 | cl_event ev = 0; |
1535 | cl_event ev = 0; |
874 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1536 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
875 | const size_t region[3] = { width, height, depth }; |
1537 | const size_t region[3] = { width, height, depth }; |
876 | EVENT_LIST (12, items - 12); |
1538 | EVENT_LIST (12); |
877 | |
1539 | |
878 | if (!row_pitch) |
1540 | if (!row_pitch) |
879 | row_pitch = img_row_pitch (src); |
1541 | row_pitch = img_row_pitch (src); |
880 | |
1542 | |
881 | if (depth > 1 && !slice_pitch) |
1543 | if (depth > 1 && !slice_pitch) |
… | |
… | |
888 | SvPOK_only (data); |
1550 | SvPOK_only (data); |
889 | SvCUR_set (data, len); |
1551 | SvCUR_set (data, len); |
890 | 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)); |
1552 | NEED_SUCCESS (EnqueueReadImage, (self, src, blocking, src_origin, region, row_pitch, slice_pitch, SvPVX (data), event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
891 | |
1553 | |
892 | if (ev) |
1554 | if (ev) |
893 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1555 | XPUSH_CLOBJ (stash_event, ev); |
894 | |
1556 | |
895 | void |
1557 | void |
896 | 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, ...) |
1558 | write_image (OpenCL::Queue self, OpenCL::Image dst, cl_bool blocking, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, size_t row_pitch, size_t slice_pitch, SV *data, ...) |
|
|
1559 | ALIAS: |
|
|
1560 | enqueue_write_image = 0 |
897 | PPCODE: |
1561 | PPCODE: |
898 | cl_event ev = 0; |
1562 | cl_event ev = 0; |
899 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1563 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
900 | const size_t region[3] = { width, height, depth }; |
1564 | const size_t region[3] = { width, height, depth }; |
901 | STRLEN len; |
1565 | STRLEN len; |
902 | char *ptr = SvPVbyte (data, len); |
1566 | char *ptr = SvPVbyte (data, len); |
903 | EVENT_LIST (12, items - 12); |
1567 | EVENT_LIST (12); |
904 | |
1568 | |
905 | if (!row_pitch) |
1569 | if (!row_pitch) |
906 | row_pitch = img_row_pitch (dst); |
1570 | row_pitch = img_row_pitch (dst); |
907 | |
1571 | |
908 | if (depth > 1 && !slice_pitch) |
1572 | if (depth > 1 && !slice_pitch) |
… | |
… | |
911 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
1575 | size_t min_len = slice_pitch ? slice_pitch * depth : row_pitch * height; |
912 | |
1576 | |
913 | if (len < min_len) |
1577 | if (len < min_len) |
914 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
1578 | croak ("clEnqueueWriteImage: data string is shorter than what would be transferred"); |
915 | |
1579 | |
916 | 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)); |
1580 | NEED_SUCCESS (EnqueueWriteImage, (self, dst, blocking, dst_origin, region, row_pitch, slice_pitch, ptr, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
917 | |
1581 | |
918 | if (ev) |
1582 | if (ev) |
919 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1583 | XPUSH_CLOBJ (stash_event, ev); |
920 | |
1584 | |
921 | void |
1585 | void |
922 | 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, ...) |
1586 | copy_image (OpenCL::Queue self, OpenCL::Image src, OpenCL::Image dst, size_t src_x, size_t src_y, size_t src_z, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
1587 | ALIAS: |
|
|
1588 | enqueue_copy_image = 0 |
923 | PPCODE: |
1589 | PPCODE: |
924 | cl_event ev = 0; |
1590 | cl_event ev = 0; |
925 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1591 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
926 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
1592 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
927 | const size_t region[3] = { width, height, depth }; |
1593 | const size_t region[3] = { width, height, depth }; |
928 | EVENT_LIST (12, items - 12); |
1594 | EVENT_LIST (12); |
929 | |
1595 | |
930 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1596 | NEED_SUCCESS (EnqueueCopyImage, (self, src, dst, src_origin, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
931 | |
1597 | |
932 | if (ev) |
1598 | if (ev) |
933 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1599 | XPUSH_CLOBJ (stash_event, ev); |
934 | |
1600 | |
935 | void |
1601 | void |
936 | 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, ...) |
1602 | copy_image_to_buffer (OpenCL::Queue self, OpenCL::Image src, OpenCL::Buffer dst, size_t src_x, size_t src_y, size_t src_z, size_t width, size_t height, size_t depth, size_t dst_offset, ...) |
|
|
1603 | ALIAS: |
|
|
1604 | enqueue_copy_image_to_buffer = 0 |
937 | PPCODE: |
1605 | PPCODE: |
938 | cl_event ev = 0; |
1606 | cl_event ev = 0; |
939 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
1607 | const size_t src_origin[3] = { src_x, src_y, src_z }; |
|
|
1608 | const size_t region [3] = { width, height, depth }; |
|
|
1609 | EVENT_LIST (10); |
|
|
1610 | |
|
|
1611 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1612 | |
|
|
1613 | if (ev) |
|
|
1614 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1615 | |
|
|
1616 | void |
|
|
1617 | copy_buffer_to_image (OpenCL::Queue self, OpenCL::Buffer src, OpenCL::Image dst, size_t src_offset, size_t dst_x, size_t dst_y, size_t dst_z, size_t width, size_t height, size_t depth, ...) |
|
|
1618 | ALIAS: |
|
|
1619 | enqueue_copy_buffer_to_image = 0 |
|
|
1620 | PPCODE: |
|
|
1621 | cl_event ev = 0; |
|
|
1622 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
1623 | const size_t region [3] = { width, height, depth }; |
|
|
1624 | EVENT_LIST (10); |
|
|
1625 | |
|
|
1626 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1627 | |
|
|
1628 | if (ev) |
|
|
1629 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1630 | |
|
|
1631 | void |
|
|
1632 | map_buffer (OpenCL::Queue self, OpenCL::Buffer buf, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t offset = 0, size_t cb = 0, ...) |
|
|
1633 | ALIAS: |
|
|
1634 | enqueue_map_buffer = 0 |
|
|
1635 | PPCODE: |
|
|
1636 | cl_event ev; |
|
|
1637 | EVENT_LIST (6); |
|
|
1638 | |
|
|
1639 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapBuffer, (self, buf, blocking, map_flags, offset, cb, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1640 | XPUSHs (mapped_new (stash_mappedbuffer, self, buf, map_flags, ptr, cb, ev, 0, 0)); |
|
|
1641 | |
|
|
1642 | void |
|
|
1643 | map_image (OpenCL::Queue self, OpenCL::Image img, cl_bool blocking = 1, cl_map_flags map_flags = CL_MAP_READ | CL_MAP_WRITE, size_t x = 0, size_t y = 0, size_t z = 0, size_t width = 0, size_t height = 0, size_t depth = 0, ...) |
|
|
1644 | ALIAS: |
|
|
1645 | enqueue_map_image = 0 |
|
|
1646 | PPCODE: |
|
|
1647 | cl_event ev; |
|
|
1648 | const size_t origin[3] = { x, y, z }; |
940 | const size_t region[3] = { width, height, depth }; |
1649 | const size_t region[3] = { width, height, depth }; |
|
|
1650 | size_t row_pitch, slice_pitch; |
941 | EVENT_LIST (10, items - 10); |
1651 | EVENT_LIST (10); |
942 | |
1652 | |
|
|
1653 | NEED_SUCCESS_ARG (void *ptr, EnqueueMapImage, (self, img, blocking, map_flags, origin, region, &row_pitch, &slice_pitch, event_list_count, event_list_ptr, &ev, &res)); |
|
|
1654 | |
|
|
1655 | size_t cb = slice_pitch ? slice_pitch * region [2] |
|
|
1656 | : row_pitch ? row_pitch * region [1] |
|
|
1657 | : region [0]; |
|
|
1658 | |
|
|
1659 | XPUSHs (mapped_new (stash_mappedimage, self, img, map_flags, ptr, cb, ev, row_pitch, slice_pitch)); |
|
|
1660 | |
|
|
1661 | void |
|
|
1662 | unmap (OpenCL::Queue self, OpenCL::Mapped mapped, ...) |
|
|
1663 | PPCODE: |
|
|
1664 | mapped_unmap (ST (1), mapped, self, &ST (2), items - 2); |
|
|
1665 | if (GIMME_V != G_VOID) |
|
|
1666 | { |
|
|
1667 | clRetainEvent (mapped->event); |
|
|
1668 | XPUSH_CLOBJ (stash_event, mapped->event); |
|
|
1669 | } |
|
|
1670 | |
|
|
1671 | void |
|
|
1672 | task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
1673 | ALIAS: |
|
|
1674 | enqueue_task = 0 |
|
|
1675 | PPCODE: |
|
|
1676 | cl_event ev = 0; |
|
|
1677 | EVENT_LIST (2); |
|
|
1678 | |
943 | NEED_SUCCESS (EnqueueCopyImageToBuffer, (self, src, dst, src_origin, region, dst_offset, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1679 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
944 | |
1680 | |
945 | if (ev) |
1681 | if (ev) |
946 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1682 | XPUSH_CLOBJ (stash_event, ev); |
947 | |
1683 | |
948 | void |
1684 | void |
949 | 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, ...) |
|
|
950 | PPCODE: |
|
|
951 | cl_event ev = 0; |
|
|
952 | const size_t dst_origin[3] = { dst_x, dst_y, dst_z }; |
|
|
953 | const size_t region[3] = { width, height, depth }; |
|
|
954 | EVENT_LIST (10, items - 10); |
|
|
955 | |
|
|
956 | NEED_SUCCESS (EnqueueCopyBufferToImage, (self, src, dst, src_offset, dst_origin, region, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
957 | |
|
|
958 | if (ev) |
|
|
959 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
960 | |
|
|
961 | void |
|
|
962 | enqueue_task (OpenCL::Queue self, OpenCL::Kernel kernel, ...) |
|
|
963 | PPCODE: |
|
|
964 | cl_event ev = 0; |
|
|
965 | EVENT_LIST (2, items - 2); |
|
|
966 | |
|
|
967 | NEED_SUCCESS (EnqueueTask, (self, kernel, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
968 | |
|
|
969 | if (ev) |
|
|
970 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
|
|
971 | |
|
|
972 | void |
|
|
973 | 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, ...) |
1685 | nd_range_kernel (OpenCL::Queue self, OpenCL::Kernel kernel, SV *global_work_offset, SV *global_work_size, SV *local_work_size = &PL_sv_undef, ...) |
|
|
1686 | ALIAS: |
|
|
1687 | enqueue_nd_range_kernel = 0 |
974 | PPCODE: |
1688 | PPCODE: |
975 | cl_event ev = 0; |
1689 | cl_event ev = 0; |
976 | size_t *gwo = 0, *gws, *lws = 0; |
1690 | size_t *gwo = 0, *gws, *lws = 0; |
977 | int gws_len; |
1691 | int gws_len; |
978 | size_t *lists; |
1692 | size_t *lists; |
979 | int i; |
1693 | int i; |
980 | EVENT_LIST (5, items - 5); |
1694 | EVENT_LIST (5); |
981 | |
1695 | |
982 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
1696 | if (!SvROK (global_work_size) || SvTYPE (SvRV (global_work_size)) != SVt_PVAV) |
983 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
1697 | croak ("clEnqueueNDRangeKernel: global_work_size must be an array reference"); |
984 | |
1698 | |
985 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
1699 | gws_len = AvFILLp (SvRV (global_work_size)) + 1; |
986 | |
1700 | |
987 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
1701 | lists = tmpbuf (sizeof (size_t) * 3 * gws_len); |
988 | |
1702 | |
989 | gws = lists + gws_len * 0; |
1703 | gws = lists + gws_len * 0; |
990 | for (i = 0; i < gws_len; ++i) |
1704 | for (i = 0; i < gws_len; ++i) |
|
|
1705 | { |
991 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
1706 | gws [i] = SvIV (AvARRAY (SvRV (global_work_size))[i]); |
|
|
1707 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1708 | if (!gws [i]) |
|
|
1709 | croak ("clEnqueueNDRangeKernel: global_work_size[%d] is zero, must be non-zero", i); |
|
|
1710 | } |
992 | |
1711 | |
993 | if (SvOK (global_work_offset)) |
1712 | if (SvOK (global_work_offset)) |
994 | { |
1713 | { |
995 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
1714 | if (!SvROK (global_work_offset) || SvTYPE (SvRV (global_work_offset)) != SVt_PVAV) |
996 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
1715 | croak ("clEnqueueNDRangeKernel: global_work_offset must be undef or an array reference"); |
… | |
… | |
1003 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1722 | gwo [i] = SvIV (AvARRAY (SvRV (global_work_offset))[i]); |
1004 | } |
1723 | } |
1005 | |
1724 | |
1006 | if (SvOK (local_work_size)) |
1725 | if (SvOK (local_work_size)) |
1007 | { |
1726 | { |
1008 | if (SvOK (local_work_size) && !SvROK (local_work_size) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1727 | if ((SvOK (local_work_size) && !SvROK (local_work_size)) || SvTYPE (SvRV (local_work_size)) != SVt_PVAV) |
1009 | croak ("clEnqueueNDRangeKernel: global_work_size must be undef or an array reference"); |
1728 | croak ("clEnqueueNDRangeKernel: local_work_size must be undef or an array reference"); |
1010 | |
1729 | |
1011 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1730 | if (AvFILLp (SvRV (local_work_size)) + 1 != gws_len) |
1012 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1731 | croak ("clEnqueueNDRangeKernel: local_work_local must be undef or an array of same size as global_work_size"); |
1013 | |
1732 | |
1014 | lws = lists + gws_len * 2; |
1733 | lws = lists + gws_len * 2; |
1015 | for (i = 0; i < gws_len; ++i) |
1734 | for (i = 0; i < gws_len; ++i) |
|
|
1735 | { |
1016 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
1736 | lws [i] = SvIV (AvARRAY (SvRV (local_work_size))[i]); |
|
|
1737 | // at least nvidia crashes for 0-sized work group sizes, work around |
|
|
1738 | if (!lws [i]) |
|
|
1739 | croak ("clEnqueueNDRangeKernel: local_work_size[%d] is zero, must be non-zero", i); |
|
|
1740 | } |
1017 | } |
1741 | } |
1018 | |
1742 | |
1019 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1743 | NEED_SUCCESS (EnqueueNDRangeKernel, (self, kernel, gws_len, gwo, gws, lws, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
1020 | |
1744 | |
1021 | if (ev) |
1745 | if (ev) |
1022 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1746 | XPUSH_CLOBJ (stash_event, ev); |
1023 | |
1747 | |
|
|
1748 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1749 | |
1024 | void |
1750 | void |
1025 | enqueue_marker (OpenCL::Queue self) |
1751 | acquire_gl_objects (OpenCL::Queue self, SV *objects, ...) |
|
|
1752 | ALIAS: |
|
|
1753 | release_gl_objects = 1 |
|
|
1754 | enqueue_acquire_gl_objects = 0 |
|
|
1755 | enqueue_release_gl_objects = 1 |
1026 | PPCODE: |
1756 | PPCODE: |
|
|
1757 | if (!SvROK (objects) || SvTYPE (SvRV (objects)) != SVt_PVAV) |
|
|
1758 | croak ("OpenCL::Queue::enqueue_acquire/release_gl_objects argument 'objects' must be an arrayref with memory objects, in call"); |
|
|
1759 | |
1027 | cl_event ev; |
1760 | cl_event ev = 0; |
1028 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
1761 | EVENT_LIST (2); |
1029 | XPUSH_NEW_OBJ ("OpenCL::Event", ev); |
1762 | AV *av = (AV *)SvRV (objects); |
|
|
1763 | cl_uint num_objects = av_len (av) + 1; |
|
|
1764 | cl_mem *object_list = tmpbuf (sizeof (cl_mem) * num_objects); |
|
|
1765 | int i; |
1030 | |
1766 | |
|
|
1767 | for (i = num_objects; i--; ) |
|
|
1768 | object_list [i] = SvCLOBJ ("OpenCL::Queue::enqueue_acquire/release_gl_objects", "objects", *av_fetch (av, i, 0), "OpenCL::Memory"); |
|
|
1769 | |
|
|
1770 | if (ix) |
|
|
1771 | NEED_SUCCESS (EnqueueReleaseGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1772 | else |
|
|
1773 | NEED_SUCCESS (EnqueueAcquireGLObjects, (self, num_objects, object_list, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1774 | |
|
|
1775 | if (ev) |
|
|
1776 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1777 | |
|
|
1778 | #endif |
|
|
1779 | |
1031 | void |
1780 | void |
1032 | enqueue_wait_for_events (OpenCL::Queue self, ...) |
1781 | wait_for_events (OpenCL::Queue self, ...) |
|
|
1782 | ALIAS: |
|
|
1783 | enqueue_wait_for_events = 0 |
1033 | CODE: |
1784 | CODE: |
1034 | EVENT_LIST (1, items - 1); |
1785 | EVENT_LIST (1); |
|
|
1786 | #if PREFER_1_1 |
1035 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
1787 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1788 | #else |
|
|
1789 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, 0)); |
|
|
1790 | #endif |
1036 | |
1791 | |
1037 | void |
1792 | void |
1038 | enqueue_barrier (OpenCL::Queue self) |
1793 | marker (OpenCL::Queue self, ...) |
|
|
1794 | ALIAS: |
|
|
1795 | enqueue_marker = 0 |
1039 | CODE: |
1796 | PPCODE: |
|
|
1797 | cl_event ev = 0; |
|
|
1798 | EVENT_LIST (1); |
|
|
1799 | #if PREFER_1_1 |
|
|
1800 | if (!event_list_count) |
|
|
1801 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1802 | else |
|
|
1803 | #if CL_VERSION_1_2 |
|
|
1804 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1805 | #else |
|
|
1806 | { |
|
|
1807 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); // also a barrier |
|
|
1808 | NEED_SUCCESS (EnqueueMarker, (self, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1809 | } |
|
|
1810 | #endif |
|
|
1811 | #else |
|
|
1812 | NEED_SUCCESS (EnqueueMarkerWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1813 | #endif |
|
|
1814 | if (ev) |
|
|
1815 | XPUSH_CLOBJ (stash_event, ev); |
|
|
1816 | |
|
|
1817 | void |
|
|
1818 | barrier (OpenCL::Queue self, ...) |
|
|
1819 | ALIAS: |
|
|
1820 | enqueue_barrier = 0 |
|
|
1821 | PPCODE: |
|
|
1822 | cl_event ev = 0; |
|
|
1823 | EVENT_LIST (1); |
|
|
1824 | #if PREFER_1_1 |
|
|
1825 | if (!event_list_count && GIMME_V == G_VOID) |
1040 | NEED_SUCCESS (EnqueueBarrier, (self)); |
1826 | NEED_SUCCESS (EnqueueBarrier, (self)); |
|
|
1827 | else |
|
|
1828 | #if CL_VERSION_1_2 |
|
|
1829 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1830 | #else |
|
|
1831 | { |
|
|
1832 | if (event_list_count) |
|
|
1833 | NEED_SUCCESS (EnqueueWaitForEvents, (self, event_list_count, event_list_ptr)); |
|
|
1834 | |
|
|
1835 | if (GIMME_V != G_VOID) |
|
|
1836 | NEED_SUCCESS (EnqueueMarker, (self, &ev)); |
|
|
1837 | } |
|
|
1838 | #endif |
|
|
1839 | #else |
|
|
1840 | NEED_SUCCESS (EnqueueBarrierWithWaitList, (self, event_list_count, event_list_ptr, GIMME_V != G_VOID ? &ev : 0)); |
|
|
1841 | #endif |
|
|
1842 | if (ev) |
|
|
1843 | XPUSH_CLOBJ (stash_event, ev); |
1041 | |
1844 | |
1042 | void |
1845 | void |
1043 | flush (OpenCL::Queue self) |
1846 | flush (OpenCL::Queue self) |
1044 | CODE: |
1847 | CODE: |
1045 | NEED_SUCCESS (Flush, (self)); |
1848 | NEED_SUCCESS (Flush, (self)); |
… | |
… | |
1063 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1866 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_CONTEXT, sizeof (value), value, 0)); |
1064 | EXTEND (SP, 1); |
1867 | EXTEND (SP, 1); |
1065 | const int i = 0; |
1868 | const int i = 0; |
1066 | { |
1869 | { |
1067 | NEED_SUCCESS (RetainContext, (value [i])); |
1870 | NEED_SUCCESS (RetainContext, (value [i])); |
1068 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
1871 | PUSH_CLOBJ (stash_context, value [i]); |
1069 | } |
1872 | } |
1070 | |
1873 | |
1071 | void |
1874 | void |
1072 | device (OpenCL::Queue self) |
1875 | device (OpenCL::Queue self) |
1073 | PPCODE: |
1876 | PPCODE: |
1074 | cl_device_id value [1]; |
1877 | cl_device_id value [1]; |
1075 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1878 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_DEVICE, sizeof (value), value, 0)); |
1076 | EXTEND (SP, 1); |
1879 | EXTEND (SP, 1); |
1077 | const int i = 0; |
1880 | const int i = 0; |
1078 | { |
1881 | { |
1079 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
1882 | PUSH_CLOBJ (stash_device, value [i]); |
1080 | } |
1883 | } |
1081 | |
1884 | |
1082 | void |
1885 | void |
1083 | reference_count (OpenCL::Queue self) |
1886 | reference_count (OpenCL::Queue self) |
1084 | PPCODE: |
1887 | PPCODE: |
… | |
… | |
1093 | PPCODE: |
1896 | PPCODE: |
1094 | cl_command_queue_properties value [1]; |
1897 | cl_command_queue_properties value [1]; |
1095 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1898 | NEED_SUCCESS (GetCommandQueueInfo, (self, CL_QUEUE_PROPERTIES, sizeof (value), value, 0)); |
1096 | EXTEND (SP, 1); |
1899 | EXTEND (SP, 1); |
1097 | const int i = 0; |
1900 | const int i = 0; |
1098 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1901 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1099 | |
1902 | |
1100 | #END:command_queue |
1903 | #END:command_queue |
1101 | |
1904 | |
1102 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1905 | MODULE = OpenCL PACKAGE = OpenCL::Memory |
1103 | |
1906 | |
… | |
… | |
1118 | PPCODE: |
1921 | PPCODE: |
1119 | cl_mem_object_type value [1]; |
1922 | cl_mem_object_type value [1]; |
1120 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1923 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_TYPE, sizeof (value), value, 0)); |
1121 | EXTEND (SP, 1); |
1924 | EXTEND (SP, 1); |
1122 | const int i = 0; |
1925 | const int i = 0; |
1123 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1926 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1124 | |
1927 | |
1125 | void |
1928 | void |
1126 | flags (OpenCL::Memory self) |
1929 | flags (OpenCL::Memory self) |
1127 | PPCODE: |
1930 | PPCODE: |
1128 | cl_mem_flags value [1]; |
1931 | cl_mem_flags value [1]; |
1129 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1932 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_FLAGS, sizeof (value), value, 0)); |
1130 | EXTEND (SP, 1); |
1933 | EXTEND (SP, 1); |
1131 | const int i = 0; |
1934 | const int i = 0; |
1132 | PUSHs (sv_2mortal (newSViv (value [i]))); |
1935 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1133 | |
1936 | |
1134 | void |
1937 | void |
1135 | size (OpenCL::Memory self) |
1938 | size (OpenCL::Memory self) |
1136 | ALIAS: |
1939 | ALIAS: |
1137 | size = CL_MEM_SIZE |
1940 | size = CL_MEM_SIZE |
… | |
… | |
1171 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1974 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_CONTEXT, sizeof (value), value, 0)); |
1172 | EXTEND (SP, 1); |
1975 | EXTEND (SP, 1); |
1173 | const int i = 0; |
1976 | const int i = 0; |
1174 | { |
1977 | { |
1175 | NEED_SUCCESS (RetainContext, (value [i])); |
1978 | NEED_SUCCESS (RetainContext, (value [i])); |
1176 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
1979 | PUSH_CLOBJ (stash_context, value [i]); |
1177 | } |
1980 | } |
1178 | |
1981 | |
1179 | void |
1982 | void |
1180 | associated_memobject (OpenCL::Memory self) |
1983 | associated_memobject (OpenCL::Memory self) |
1181 | PPCODE: |
1984 | PPCODE: |
… | |
… | |
1183 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1986 | NEED_SUCCESS (GetMemObjectInfo, (self, CL_MEM_ASSOCIATED_MEMOBJECT, sizeof (value), value, 0)); |
1184 | EXTEND (SP, 1); |
1987 | EXTEND (SP, 1); |
1185 | const int i = 0; |
1988 | const int i = 0; |
1186 | { |
1989 | { |
1187 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1990 | NEED_SUCCESS (RetainMemObject, (value [i])); |
1188 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Memory", value [i])); |
1991 | PUSH_CLOBJ (stash_memory, value [i]); |
1189 | } |
1992 | } |
1190 | |
1993 | |
1191 | #END:mem |
1994 | #END:mem |
|
|
1995 | |
|
|
1996 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
1997 | |
|
|
1998 | void |
|
|
1999 | gl_object_info (OpenCL::Memory self) |
|
|
2000 | PPCODE: |
|
|
2001 | cl_gl_object_type type; |
|
|
2002 | cl_GLuint name; |
|
|
2003 | NEED_SUCCESS (GetGLObjectInfo, (self, &type, &name)); |
|
|
2004 | EXTEND (SP, 2); |
|
|
2005 | PUSHs (sv_2mortal (newSVuv (type))); |
|
|
2006 | PUSHs (sv_2mortal (newSVuv (name))); |
|
|
2007 | |
|
|
2008 | #endif |
1192 | |
2009 | |
1193 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
2010 | MODULE = OpenCL PACKAGE = OpenCL::BufferObj |
1194 | |
2011 | |
1195 | void |
2012 | void |
1196 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
2013 | sub_buffer_region (OpenCL::BufferObj self, cl_mem_flags flags, size_t origin, size_t size) |
… | |
… | |
1199 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
2016 | croak ("clCreateSubBuffer: cannot use/copy/alloc host ptr, doesn't make sense, check your flags!"); |
1200 | |
2017 | |
1201 | cl_buffer_region crdata = { origin, size }; |
2018 | cl_buffer_region crdata = { origin, size }; |
1202 | |
2019 | |
1203 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
2020 | NEED_SUCCESS_ARG (cl_mem mem, CreateSubBuffer, (self, flags, CL_BUFFER_CREATE_TYPE_REGION, &crdata, &res)); |
1204 | XPUSH_NEW_OBJ ("OpenCL::Buffer", mem); |
2021 | XPUSH_CLOBJ (stash_buffer, mem); |
1205 | |
2022 | |
1206 | MODULE = OpenCL PACKAGE = OpenCL::Image |
2023 | MODULE = OpenCL PACKAGE = OpenCL::Image |
1207 | |
2024 | |
1208 | void |
2025 | void |
1209 | image_info (OpenCL::Image self, cl_image_info name) |
2026 | image_info (OpenCL::Image self, cl_image_info name) |
1210 | PPCODE: |
2027 | PPCODE: |
1211 | INFO (Image) |
2028 | INFO (Image) |
|
|
2029 | |
|
|
2030 | void |
|
|
2031 | format (OpenCL::Image self) |
|
|
2032 | PPCODE: |
|
|
2033 | cl_image_format format; |
|
|
2034 | NEED_SUCCESS (GetImageInfo, (self, CL_IMAGE_FORMAT, sizeof (format), &format, 0)); |
|
|
2035 | EXTEND (SP, 2); |
|
|
2036 | PUSHs (sv_2mortal (newSVuv (format.image_channel_order))); |
|
|
2037 | PUSHs (sv_2mortal (newSVuv (format.image_channel_data_type))); |
1212 | |
2038 | |
1213 | #BEGIN:image |
2039 | #BEGIN:image |
1214 | |
2040 | |
1215 | void |
2041 | void |
1216 | element_size (OpenCL::Image self) |
2042 | element_size (OpenCL::Image self) |
… | |
… | |
1228 | const int i = 0; |
2054 | const int i = 0; |
1229 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
2055 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1230 | |
2056 | |
1231 | #END:image |
2057 | #END:image |
1232 | |
2058 | |
|
|
2059 | #if cl_apple_gl_sharing || cl_khr_gl_sharing |
|
|
2060 | |
|
|
2061 | #BEGIN:gl_texture |
|
|
2062 | |
|
|
2063 | void |
|
|
2064 | target (OpenCL::Image self) |
|
|
2065 | PPCODE: |
|
|
2066 | cl_GLenum value [1]; |
|
|
2067 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_TEXTURE_TARGET, sizeof (value), value, 0)); |
|
|
2068 | EXTEND (SP, 1); |
|
|
2069 | const int i = 0; |
|
|
2070 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
|
|
2071 | |
|
|
2072 | void |
|
|
2073 | gl_mipmap_level (OpenCL::Image self) |
|
|
2074 | PPCODE: |
|
|
2075 | cl_GLint value [1]; |
|
|
2076 | NEED_SUCCESS (GetGLTextureInfo, (self, CL_GL_MIPMAP_LEVEL, sizeof (value), value, 0)); |
|
|
2077 | EXTEND (SP, 1); |
|
|
2078 | const int i = 0; |
|
|
2079 | PUSHs (sv_2mortal (newSViv (value [i]))); |
|
|
2080 | |
|
|
2081 | #END:gl_texture |
|
|
2082 | |
|
|
2083 | #endif |
|
|
2084 | |
1233 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
2085 | MODULE = OpenCL PACKAGE = OpenCL::Sampler |
1234 | |
2086 | |
1235 | void |
2087 | void |
1236 | DESTROY (OpenCL::Sampler self) |
2088 | DESTROY (OpenCL::Sampler self) |
1237 | CODE: |
2089 | CODE: |
… | |
… | |
1260 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
2112 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_CONTEXT, sizeof (value), value, 0)); |
1261 | EXTEND (SP, 1); |
2113 | EXTEND (SP, 1); |
1262 | const int i = 0; |
2114 | const int i = 0; |
1263 | { |
2115 | { |
1264 | NEED_SUCCESS (RetainContext, (value [i])); |
2116 | NEED_SUCCESS (RetainContext, (value [i])); |
1265 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2117 | PUSH_CLOBJ (stash_context, value [i]); |
1266 | } |
2118 | } |
1267 | |
2119 | |
1268 | void |
2120 | void |
1269 | normalized_coords (OpenCL::Sampler self) |
2121 | normalized_coords (OpenCL::Sampler self) |
1270 | PPCODE: |
2122 | PPCODE: |
1271 | cl_addressing_mode value [1]; |
2123 | cl_addressing_mode value [1]; |
1272 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
2124 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_NORMALIZED_COORDS, sizeof (value), value, 0)); |
1273 | EXTEND (SP, 1); |
2125 | EXTEND (SP, 1); |
1274 | const int i = 0; |
2126 | const int i = 0; |
1275 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2127 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1276 | |
2128 | |
1277 | void |
2129 | void |
1278 | addressing_mode (OpenCL::Sampler self) |
2130 | addressing_mode (OpenCL::Sampler self) |
1279 | PPCODE: |
2131 | PPCODE: |
1280 | cl_filter_mode value [1]; |
2132 | cl_filter_mode value [1]; |
1281 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
2133 | NEED_SUCCESS (GetSamplerInfo, (self, CL_SAMPLER_ADDRESSING_MODE, sizeof (value), value, 0)); |
1282 | EXTEND (SP, 1); |
2134 | EXTEND (SP, 1); |
1283 | const int i = 0; |
2135 | const int i = 0; |
1284 | PUSHs (sv_2mortal (newSViv (value [i]))); |
2136 | PUSHs (sv_2mortal (newSVuv (value [i]))); |
1285 | |
2137 | |
1286 | void |
2138 | void |
1287 | filter_mode (OpenCL::Sampler self) |
2139 | filter_mode (OpenCL::Sampler self) |
1288 | PPCODE: |
2140 | PPCODE: |
1289 | cl_bool value [1]; |
2141 | cl_bool value [1]; |
… | |
… | |
1300 | DESTROY (OpenCL::Program self) |
2152 | DESTROY (OpenCL::Program self) |
1301 | CODE: |
2153 | CODE: |
1302 | clReleaseProgram (self); |
2154 | clReleaseProgram (self); |
1303 | |
2155 | |
1304 | void |
2156 | void |
1305 | build (OpenCL::Program self, OpenCL::Device device, SV *options = &PL_sv_undef) |
2157 | build (OpenCL::Program self, SV *devices = &PL_sv_undef, SV *options = &PL_sv_undef, SV *notify = &PL_sv_undef) |
|
|
2158 | ALIAS: |
|
|
2159 | build_async = 1 |
1306 | CODE: |
2160 | CODE: |
|
|
2161 | void (CL_CALLBACK *pfn_notify)(cl_program program, void *user_data) = 0; |
|
|
2162 | void *user_data = 0; |
|
|
2163 | cl_uint num_devices = 0; |
|
|
2164 | cl_device_id *device_list = 0; |
|
|
2165 | |
|
|
2166 | if (SvOK (devices)) |
|
|
2167 | { |
|
|
2168 | if (!SvROK (devices) || SvTYPE (SvRV (devices)) != SVt_PVAV) |
|
|
2169 | croak ("clProgramBuild: devices must be undef or an array of OpenCL::Device objects."); |
|
|
2170 | |
|
|
2171 | AV *av = (AV *)SvRV (devices); |
|
|
2172 | num_devices = av_len (av) + 1; |
|
|
2173 | |
|
|
2174 | if (num_devices) |
|
|
2175 | { |
|
|
2176 | device_list = tmpbuf (sizeof (*device_list) * num_devices); |
|
|
2177 | int count; |
|
|
2178 | for (count = 0; count < num_devices; ++count) |
|
|
2179 | device_list [count] = SvCLOBJ ("clBuildProgram", "devices", *av_fetch (av, count, 1), "OpenCL::Device"); |
|
|
2180 | } |
|
|
2181 | } |
|
|
2182 | |
|
|
2183 | if (SvOK (notify)) |
|
|
2184 | { |
|
|
2185 | NEED_SUCCESS (RetainProgram, (self)); |
|
|
2186 | pfn_notify = eq_program_notify; |
|
|
2187 | user_data = SvREFCNT_inc (s_get_cv (notify)); |
|
|
2188 | } |
|
|
2189 | |
|
|
2190 | if (ix) |
|
|
2191 | build_program_async (self, num_devices, device_list, SvPVbyte_nolen (options), user_data); |
|
|
2192 | else |
1307 | NEED_SUCCESS (BuildProgram, (self, 1, &device, SvPVbyte_nolen (options), 0, 0)); |
2193 | NEED_SUCCESS (BuildProgram, (self, num_devices, device_list, SvPVbyte_nolen (options), pfn_notify, user_data)); |
1308 | |
2194 | |
1309 | void |
2195 | void |
1310 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
2196 | build_info (OpenCL::Program self, OpenCL::Device device, cl_program_build_info name) |
1311 | PPCODE: |
2197 | PPCODE: |
1312 | size_t size; |
2198 | size_t size; |
… | |
… | |
1347 | |
2233 | |
1348 | void |
2234 | void |
1349 | kernel (OpenCL::Program program, SV *function) |
2235 | kernel (OpenCL::Program program, SV *function) |
1350 | PPCODE: |
2236 | PPCODE: |
1351 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
2237 | NEED_SUCCESS_ARG (cl_kernel kernel, CreateKernel, (program, SvPVbyte_nolen (function), &res)); |
1352 | XPUSH_NEW_OBJ ("OpenCL::Kernel", kernel); |
2238 | XPUSH_CLOBJ (stash_kernel, kernel); |
|
|
2239 | |
|
|
2240 | void |
|
|
2241 | kernels_in_program (OpenCL::Program program) |
|
|
2242 | PPCODE: |
|
|
2243 | cl_uint num_kernels; |
|
|
2244 | NEED_SUCCESS (CreateKernelsInProgram, (program, 0, 0, &num_kernels)); |
|
|
2245 | cl_kernel *kernels = tmpbuf (sizeof (cl_kernel) * num_kernels); |
|
|
2246 | NEED_SUCCESS (CreateKernelsInProgram, (program, num_kernels, kernels, 0)); |
|
|
2247 | |
|
|
2248 | int i; |
|
|
2249 | EXTEND (SP, num_kernels); |
|
|
2250 | for (i = 0; i < num_kernels; ++i) |
|
|
2251 | PUSH_CLOBJ (stash_kernel, kernels [i]); |
1353 | |
2252 | |
1354 | void |
2253 | void |
1355 | info (OpenCL::Program self, cl_program_info name) |
2254 | info (OpenCL::Program self, cl_program_info name) |
1356 | PPCODE: |
2255 | PPCODE: |
1357 | INFO (Program) |
2256 | INFO (Program) |
… | |
… | |
1375 | { |
2274 | { |
1376 | SV *sv = sv_2mortal (newSV (sizes [i])); |
2275 | SV *sv = sv_2mortal (newSV (sizes [i])); |
1377 | SvUPGRADE (sv, SVt_PV); |
2276 | SvUPGRADE (sv, SVt_PV); |
1378 | SvPOK_only (sv); |
2277 | SvPOK_only (sv); |
1379 | SvCUR_set (sv, sizes [i]); |
2278 | SvCUR_set (sv, sizes [i]); |
1380 | ptrs [i] = SvPVX (sv); |
2279 | ptrs [i] = (void *)SvPVX (sv); |
1381 | PUSHs (sv); |
2280 | PUSHs (sv); |
1382 | } |
2281 | } |
1383 | |
2282 | |
1384 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
2283 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_BINARIES , sizeof (*ptrs ) * n, ptrs , &size)); |
1385 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
2284 | if (size != sizeof (*ptrs) * n) XSRETURN_EMPTY; |
… | |
… | |
1405 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
2304 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_CONTEXT, sizeof (value), value, 0)); |
1406 | EXTEND (SP, 1); |
2305 | EXTEND (SP, 1); |
1407 | const int i = 0; |
2306 | const int i = 0; |
1408 | { |
2307 | { |
1409 | NEED_SUCCESS (RetainContext, (value [i])); |
2308 | NEED_SUCCESS (RetainContext, (value [i])); |
1410 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2309 | PUSH_CLOBJ (stash_context, value [i]); |
1411 | } |
2310 | } |
1412 | |
2311 | |
1413 | void |
2312 | void |
1414 | devices (OpenCL::Program self) |
2313 | devices (OpenCL::Program self) |
1415 | PPCODE: |
2314 | PPCODE: |
… | |
… | |
1419 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
2318 | NEED_SUCCESS (GetProgramInfo, (self, CL_PROGRAM_DEVICES, size, value, 0)); |
1420 | int i, n = size / sizeof (*value); |
2319 | int i, n = size / sizeof (*value); |
1421 | EXTEND (SP, n); |
2320 | EXTEND (SP, n); |
1422 | for (i = 0; i < n; ++i) |
2321 | for (i = 0; i < n; ++i) |
1423 | { |
2322 | { |
1424 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Device", value [i])); |
2323 | PUSH_CLOBJ (stash_device, value [i]); |
1425 | } |
2324 | } |
1426 | |
2325 | |
1427 | void |
2326 | void |
1428 | source (OpenCL::Program self) |
2327 | source (OpenCL::Program self) |
1429 | PPCODE: |
2328 | PPCODE: |
… | |
… | |
1455 | DESTROY (OpenCL::Kernel self) |
2354 | DESTROY (OpenCL::Kernel self) |
1456 | CODE: |
2355 | CODE: |
1457 | clReleaseKernel (self); |
2356 | clReleaseKernel (self); |
1458 | |
2357 | |
1459 | void |
2358 | void |
|
|
2359 | setf (OpenCL::Kernel self, const char *format, ...) |
|
|
2360 | CODE: |
|
|
2361 | int i; |
|
|
2362 | for (i = 2; ; ++i) |
|
|
2363 | { |
|
|
2364 | while (*format == ' ') |
|
|
2365 | ++format; |
|
|
2366 | |
|
|
2367 | char type = *format++; |
|
|
2368 | |
|
|
2369 | if (!type) |
|
|
2370 | break; |
|
|
2371 | |
|
|
2372 | if (i >= items) |
|
|
2373 | croak ("OpenCL::Kernel::setf format string too long (not enough arguments)"); |
|
|
2374 | |
|
|
2375 | SV *sv = ST (i); |
|
|
2376 | |
|
|
2377 | union |
|
|
2378 | { |
|
|
2379 | cl_char cc; cl_uchar cC; cl_short cs; cl_ushort cS; |
|
|
2380 | cl_int ci; cl_uint cI; cl_long cl; cl_ulong cL; |
|
|
2381 | cl_half ch; cl_float cf; cl_double cd; |
|
|
2382 | cl_mem cm; |
|
|
2383 | cl_sampler ca; |
|
|
2384 | size_t cz; |
|
|
2385 | cl_event ce; |
|
|
2386 | } arg; |
|
|
2387 | size_t size; |
|
|
2388 | int nullarg = 0; |
|
|
2389 | |
|
|
2390 | switch (type) |
|
|
2391 | { |
|
|
2392 | case 'c': arg.cc = SvIV (sv); size = sizeof (arg.cc); break; |
|
|
2393 | case 'C': arg.cC = SvUV (sv); size = sizeof (arg.cC); break; |
|
|
2394 | case 's': arg.cs = SvIV (sv); size = sizeof (arg.cs); break; |
|
|
2395 | case 'S': arg.cS = SvUV (sv); size = sizeof (arg.cS); break; |
|
|
2396 | case 'i': arg.ci = SvIV (sv); size = sizeof (arg.ci); break; |
|
|
2397 | case 'I': arg.cI = SvUV (sv); size = sizeof (arg.cI); break; |
|
|
2398 | case 'l': arg.cl = SvIV (sv); size = sizeof (arg.cl); break; |
|
|
2399 | case 'L': arg.cL = SvUV (sv); size = sizeof (arg.cL); break; |
|
|
2400 | |
|
|
2401 | case 'h': arg.ch = SvUV (sv); size = sizeof (arg.ch); break; |
|
|
2402 | case 'f': arg.cf = SvNV (sv); size = sizeof (arg.cf); break; |
|
|
2403 | case 'd': arg.cd = SvNV (sv); size = sizeof (arg.cd); break; |
|
|
2404 | |
|
|
2405 | case 'z': nullarg = 1; size = SvIV (sv); break; |
|
|
2406 | |
|
|
2407 | case 'm': nullarg = !SvOK (sv); arg.cm = SvCLOBJ ("OpenCL::Kernel::setf", "m", sv, "OpenCL::Memory" ); size = sizeof (arg.cm); break; |
|
|
2408 | case 'a': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "a", sv, "OpenCL::Sampler"); size = sizeof (arg.ca); break; |
|
|
2409 | case 'e': nullarg = !SvOK (sv); arg.ca = SvCLOBJ ("OpenCL::Kernel::setf", "e", sv, "OpenCL::Event" ); size = sizeof (arg.ce); break; |
|
|
2410 | |
|
|
2411 | default: |
|
|
2412 | croak ("OpenCL::Kernel::setf format character '%c' not supported", type); |
|
|
2413 | } |
|
|
2414 | |
|
|
2415 | res = clSetKernelArg (self, i - 2, size, nullarg ? 0 : &arg); |
|
|
2416 | if (res) |
|
|
2417 | croak ("OpenCL::Kernel::setf kernel parameter '%c' (#%d): %s", type, i - 2, err2str (res)); |
|
|
2418 | } |
|
|
2419 | |
|
|
2420 | if (i != items) |
|
|
2421 | croak ("OpenCL::Kernel::setf format string too short (too many arguments)"); |
|
|
2422 | |
|
|
2423 | void |
1460 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
2424 | set_char (OpenCL::Kernel self, cl_uint idx, cl_char value) |
1461 | CODE: |
2425 | CODE: |
1462 | clSetKernelArg (self, idx, sizeof (value), &value); |
2426 | clSetKernelArg (self, idx, sizeof (value), &value); |
1463 | |
2427 | |
1464 | void |
2428 | void |
… | |
… | |
1512 | clSetKernelArg (self, idx, sizeof (value), &value); |
2476 | clSetKernelArg (self, idx, sizeof (value), &value); |
1513 | |
2477 | |
1514 | void |
2478 | void |
1515 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
2479 | set_memory (OpenCL::Kernel self, cl_uint idx, OpenCL::Memory_ornull value) |
1516 | CODE: |
2480 | CODE: |
|
|
2481 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2482 | |
|
|
2483 | void |
|
|
2484 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
|
|
2485 | CODE: |
|
|
2486 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2487 | |
|
|
2488 | void |
|
|
2489 | set_image (OpenCL::Kernel self, cl_uint idx, OpenCL::Image_ornull value) |
|
|
2490 | ALIAS: |
|
|
2491 | set_image2d = 0 |
|
|
2492 | set_image3d = 0 |
|
|
2493 | CODE: |
|
|
2494 | clSetKernelArg (self, idx, sizeof (value), value ? &value : 0); |
|
|
2495 | |
|
|
2496 | void |
|
|
2497 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
2498 | CODE: |
1517 | clSetKernelArg (self, idx, sizeof (value), &value); |
2499 | clSetKernelArg (self, idx, sizeof (value), &value); |
1518 | |
2500 | |
1519 | void |
2501 | void |
1520 | set_buffer (OpenCL::Kernel self, cl_uint idx, OpenCL::Buffer_ornull value) |
2502 | set_local (OpenCL::Kernel self, cl_uint idx, size_t size) |
1521 | CODE: |
2503 | CODE: |
1522 | clSetKernelArg (self, idx, sizeof (value), &value); |
2504 | clSetKernelArg (self, idx, size, 0); |
1523 | |
|
|
1524 | void |
|
|
1525 | set_image2d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image2D_ornull value) |
|
|
1526 | CODE: |
|
|
1527 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1528 | |
|
|
1529 | void |
|
|
1530 | set_image3d (OpenCL::Kernel self, cl_uint idx, OpenCL::Image3D_ornull value) |
|
|
1531 | CODE: |
|
|
1532 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1533 | |
|
|
1534 | void |
|
|
1535 | set_sampler (OpenCL::Kernel self, cl_uint idx, OpenCL::Sampler value) |
|
|
1536 | CODE: |
|
|
1537 | clSetKernelArg (self, idx, sizeof (value), &value); |
|
|
1538 | |
2505 | |
1539 | void |
2506 | void |
1540 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
2507 | set_event (OpenCL::Kernel self, cl_uint idx, OpenCL::Event value) |
1541 | CODE: |
2508 | CODE: |
1542 | clSetKernelArg (self, idx, sizeof (value), &value); |
2509 | clSetKernelArg (self, idx, sizeof (value), &value); |
… | |
… | |
1578 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
2545 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_CONTEXT, sizeof (value), value, 0)); |
1579 | EXTEND (SP, 1); |
2546 | EXTEND (SP, 1); |
1580 | const int i = 0; |
2547 | const int i = 0; |
1581 | { |
2548 | { |
1582 | NEED_SUCCESS (RetainContext, (value [i])); |
2549 | NEED_SUCCESS (RetainContext, (value [i])); |
1583 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2550 | PUSH_CLOBJ (stash_context, value [i]); |
1584 | } |
2551 | } |
1585 | |
2552 | |
1586 | void |
2553 | void |
1587 | program (OpenCL::Kernel self) |
2554 | program (OpenCL::Kernel self) |
1588 | PPCODE: |
2555 | PPCODE: |
… | |
… | |
1590 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
2557 | NEED_SUCCESS (GetKernelInfo, (self, CL_KERNEL_PROGRAM, sizeof (value), value, 0)); |
1591 | EXTEND (SP, 1); |
2558 | EXTEND (SP, 1); |
1592 | const int i = 0; |
2559 | const int i = 0; |
1593 | { |
2560 | { |
1594 | NEED_SUCCESS (RetainProgram, (value [i])); |
2561 | NEED_SUCCESS (RetainProgram, (value [i])); |
1595 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Program", value [i])); |
2562 | PUSH_CLOBJ (stash_program, value [i]); |
1596 | } |
2563 | } |
1597 | |
2564 | |
1598 | #END:kernel |
2565 | #END:kernel |
1599 | |
2566 | |
1600 | void |
2567 | void |
… | |
… | |
1660 | wait (OpenCL::Event self) |
2627 | wait (OpenCL::Event self) |
1661 | CODE: |
2628 | CODE: |
1662 | clWaitForEvents (1, &self); |
2629 | clWaitForEvents (1, &self); |
1663 | |
2630 | |
1664 | void |
2631 | void |
|
|
2632 | cb (OpenCL::Event self, cl_int command_exec_callback_type, SV *cb) |
|
|
2633 | CODE: |
|
|
2634 | clSetEventCallback (self, command_exec_callback_type, eq_event_notify, SvREFCNT_inc (s_get_cv (cb))); |
|
|
2635 | |
|
|
2636 | void |
1665 | info (OpenCL::Event self, cl_event_info name) |
2637 | info (OpenCL::Event self, cl_event_info name) |
1666 | PPCODE: |
2638 | PPCODE: |
1667 | INFO (Event) |
2639 | INFO (Event) |
1668 | |
2640 | |
1669 | #BEGIN:event |
2641 | #BEGIN:event |
… | |
… | |
1675 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
2647 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_COMMAND_QUEUE, sizeof (value), value, 0)); |
1676 | EXTEND (SP, 1); |
2648 | EXTEND (SP, 1); |
1677 | const int i = 0; |
2649 | const int i = 0; |
1678 | { |
2650 | { |
1679 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
2651 | NEED_SUCCESS (RetainCommandQueue, (value [i])); |
1680 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Queue", value [i])); |
2652 | PUSH_CLOBJ (stash_queue, value [i]); |
1681 | } |
2653 | } |
1682 | |
2654 | |
1683 | void |
2655 | void |
1684 | command_type (OpenCL::Event self) |
2656 | command_type (OpenCL::Event self) |
1685 | PPCODE: |
2657 | PPCODE: |
… | |
… | |
1708 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
2680 | NEED_SUCCESS (GetEventInfo, (self, CL_EVENT_CONTEXT, sizeof (value), value, 0)); |
1709 | EXTEND (SP, 1); |
2681 | EXTEND (SP, 1); |
1710 | const int i = 0; |
2682 | const int i = 0; |
1711 | { |
2683 | { |
1712 | NEED_SUCCESS (RetainContext, (value [i])); |
2684 | NEED_SUCCESS (RetainContext, (value [i])); |
1713 | PUSHs (NEW_MORTAL_OBJ ("OpenCL::Context", value [i])); |
2685 | PUSH_CLOBJ (stash_context, value [i]); |
1714 | } |
2686 | } |
1715 | |
2687 | |
1716 | #END:event |
2688 | #END:event |
1717 | |
2689 | |
1718 | void |
2690 | void |
… | |
… | |
1743 | void |
2715 | void |
1744 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
2716 | set_status (OpenCL::UserEvent self, cl_int execution_status) |
1745 | CODE: |
2717 | CODE: |
1746 | clSetUserEventStatus (self, execution_status); |
2718 | clSetUserEventStatus (self, execution_status); |
1747 | |
2719 | |
|
|
2720 | MODULE = OpenCL PACKAGE = OpenCL::Mapped |
|
|
2721 | |
|
|
2722 | void |
|
|
2723 | DESTROY (SV *self) |
|
|
2724 | CODE: |
|
|
2725 | OpenCL__Mapped mapped = SvMAPPED (self); |
|
|
2726 | |
|
|
2727 | clEnqueueUnmapMemObject (mapped->queue, mapped->memobj, mapped->ptr, 1, &mapped->event, 0); |
|
|
2728 | mapped_detach (self, mapped); |
|
|
2729 | |
|
|
2730 | clReleaseCommandQueue (mapped->queue); |
|
|
2731 | clReleaseEvent (mapped->event); |
|
|
2732 | Safefree (mapped); |
|
|
2733 | |
|
|
2734 | void |
|
|
2735 | unmap (OpenCL::Mapped self, ...) |
|
|
2736 | CODE: |
|
|
2737 | mapped_unmap (ST (0), self, self->queue, &ST (1), items - 1); |
|
|
2738 | |
|
|
2739 | bool |
|
|
2740 | mapped (OpenCL::Mapped self) |
|
|
2741 | CODE: |
|
|
2742 | RETVAL = !!self->ptr; |
|
|
2743 | OUTPUT: |
|
|
2744 | RETVAL |
|
|
2745 | |
|
|
2746 | void |
|
|
2747 | wait (OpenCL::Mapped self) |
|
|
2748 | PPCODE: |
|
|
2749 | if (self->event) |
|
|
2750 | NEED_SUCCESS (WaitForEvents, (1, &self->event)); |
|
|
2751 | |
|
|
2752 | void |
|
|
2753 | event (OpenCL::Mapped self) |
|
|
2754 | PPCODE: |
|
|
2755 | if (!self->event) |
|
|
2756 | XSRETURN_UNDEF; |
|
|
2757 | |
|
|
2758 | clRetainEvent (self->event); |
|
|
2759 | XPUSH_CLOBJ (stash_event, self->event); |
|
|
2760 | |
|
|
2761 | size_t |
|
|
2762 | size (OpenCL::Mapped self) |
|
|
2763 | CODE: |
|
|
2764 | RETVAL = self->cb; |
|
|
2765 | OUTPUT: |
|
|
2766 | RETVAL |
|
|
2767 | |
|
|
2768 | IV |
|
|
2769 | ptr (OpenCL::Mapped self) |
|
|
2770 | CODE: |
|
|
2771 | RETVAL = PTR2IV (self->ptr); |
|
|
2772 | OUTPUT: |
|
|
2773 | RETVAL |
|
|
2774 | |
|
|
2775 | void |
|
|
2776 | set (OpenCL::Mapped self, size_t offset, SV *data) |
|
|
2777 | CODE: |
|
|
2778 | STRLEN len; |
|
|
2779 | const char *ptr = SvPVbyte (data, len); |
|
|
2780 | |
|
|
2781 | if (offset < 0 || offset + len > self->cb) |
|
|
2782 | croak ("OpenCL::Mapped::set out of bound condition detected"); |
|
|
2783 | |
|
|
2784 | memcpy (offset + (char *)self->ptr, ptr, len); |
|
|
2785 | |
|
|
2786 | MODULE = OpenCL PACKAGE = OpenCL::MappedBuffer |
|
|
2787 | |
|
|
2788 | MODULE = OpenCL PACKAGE = OpenCL::MappedImage |
|
|
2789 | |
|
|
2790 | IV |
|
|
2791 | row_pitch (OpenCL::Mapped self) |
|
|
2792 | ALIAS: |
|
|
2793 | slice_pitch = 1 |
|
|
2794 | CODE: |
|
|
2795 | RETVAL = ix ? self->slice_pitch : self->row_pitch; |
|
|
2796 | OUTPUT: |
|
|
2797 | RETVAL |
|
|
2798 | |
|
|
2799 | |
|
|
2800 | |